Add char_masks preprocessing script for SCAIL-2 training (task 3.7)

scripts/process_char_masks.py turns per-sample character label-map videos/images
(integer pixel labels: 0 = environment, 1..K = characters -> binding slots) into
the pixel-space semantic-mask tensors the SCAIL training/inference path consumes:
{"mask": [K+1, F_pix, H_pix, W_pix]} (ch0 = environment switch, ch1..K = slots).

- Aligns to the target video's latent grid read from the saved latent metadata
  (F_pix=(F-1)*8+1, H*32, W*32), so char_masks/ lines up file-for-file with
  latents/ / driving_latents/ for PrecomputedDataset.
- Nearest-neighbour resize so integer labels are never blended; labels > K are
  dropped with a warning; ch0 filled uniformly with --environment-switch.
- Reuses process_videos.py helpers (naming, atomic save, VAE factors) and matches
  its typer CLI conventions.

Verified on CPU: a synthetic 2-character label map (plus an out-of-range id)
produces mask (7,17,128,128) with ch0 uniform, slots placed correctly, id>K
dropped, and feeds encode_mask_channels to the 8*(K+1)=56 channels. README +
docs/tasks.md 3.7 updated (upstream label-map generation via SAM/tracking is
dataset-specific and still out of scope).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

.claude/skills/train-model/SKILL.md

@@ -0,0 +1,277 @@
+---
+name: train-model
+description: End-to-end agent for training LTX-2 models. Probes filesystem and GPU, picks the right conditioning mode from the user's intent, prepares the dataset (scenes, captions, references), preprocesses, autotunes, launches, and monitors training. Use when the user wants to train, fine-tune, LoRA, or otherwise produce a custom LTX-2 model.
+argument-hint: [optional source path or run name]
+user-invocable: true
+allowed-tools: Bash, Read, Grep, Glob, Edit, Write, Agent, AskUserQuestion, TodoWrite
+---
+
+# Train Model — Orchestrator
+
+Take the user from "I want to train something" to a running, monitored training job — automating the mechanical glue (dataset layout, captioning, preprocessing, config patching, launch, monitoring) without making silent decisions on their behalf.
+
+> **Source of truth for the trainer:** [`packages/ltx-trainer/docs/`](../../../packages/ltx-trainer/docs/). This skill orchestrates those scripts; it does not duplicate their reference content.
+
+## Hard Invariants
+
+These are non-negotiable. Re-read them before every action that touches the filesystem or starts a process.
+
+1. **No file mutation outside the run workspace without explicit user approval.** The workspace is `./projects/<run-name>/`. Never overwrite, move, delete, or modify any user file or directory outside it without surfacing an explicit ask. Hours of dataset work must never be silently destroyed.
+2. **No heavy work before plan approval.** Captioning, preprocessing, training, and autotune do **not** start until the user approves `plan.md` (Phase 4). Probing the filesystem and running `nvidia-smi` is fine; encoding videos or downloading models is not.
+3. **No silent assumptions.** Every non-trivial default appears under "Assumptions" in `plan.md`.
+4. **No code changes to the trainer package without explicit consent.** If the user's intent doesn't map to a supported configuration, follow the **Escape Hatch** section below — do not unilaterally edit `packages/ltx-trainer/`.
+5. **No fabricated claims about training outcomes or data sufficiency.** Do not assert how well something *will* train, whether a dataset is "too small," how many samples/seconds of audio are "enough," which modality will "learn better," expected quality, or any similar prediction — you have no grounded basis for these, they're frequently wrong, and they mislead users. Stick to facts you can substantiate: what the trainer/docs actually say, observed numbers (loss, step time, VRAM), counts, and the user's own stated goals. If the user asks for a recommendation that depends on such judgment, you may share it **only** as an explicitly-flagged uncertainty ("I'm not sure — you'd have to try it"), never as authoritative fact. When in doubt, say less.
+
+## Keep the User Informed
+
+Most users don't know how this skill works under the hood — they don't know what "preprocessing," "a one-sample sanity check," or "autotune" mean or why they're happening. Narrate the run in plain language so it never feels like a black box:
+
+- **Entering a phase:** one or two sentences on *what you're about to do and why* — in user terms, not jargon.
+- **Leaving a phase:** one line on *what came out of it* (e.g. "captioned 9 clips," "found the fastest stable config: batch 1, ~3s/step").
+- **Explain the non-obvious phases explicitly** — these are the ones that confuse people:
+  - *Sanity check (Phase 6):* "Before the full run, I do a quick dry run on a single clip at your target resolution. It catches out-of-memory or config problems in a couple of minutes instead of failing hours into training."
+  - *Autotune (Phase 6):* "Then I try a few configuration variants on that one clip to pick the fastest one that still fits your GPU — so the full run is as fast as it can be."
+  - *Preprocess (Phase 7):* "I'm encoding your videos into the compressed latents the trainer reads. One-time step; the trained model never sees the raw videos directly."
+- Keep it concise — a sentence or two per transition, not walls of text or raw logs. This is running commentary, not a replacement for the upfront plan (Phase 4) or the status reports (Phase 8).
+- Long-running steps (preprocess, training): say roughly how long it'll take and that you'll report back, so silence doesn't read as "stuck."
+- **Describe what you're doing — don't editorialize about how it'll turn out.** Narration covers *what's happening*; it must not drift into unfounded predictions about training quality or data sufficiency (e.g. "26s of audio is too little," "voice won't learn well"). Those are fabricated claims — see Hard Invariant #5. State facts and the user's choices; leave the "will it be good?" judgment to the user watching the results.
+
+## Phase 0 — Set Up
+
+Create the workspace and todos.
+
+1. Pick a workspace root in this order (use first writable):
+   - `$LTX_TRAININGS_DIR`
+   - `/data/ltx-trainings/`
+   - `/workspace/ltx-trainings/`
+   - `./projects/` (repo-relative — preferred default in this repo)
+2. Derive a tentative `<run-name>` from the user's words; finalise after Phase 1 once the mode is known. Format: `<mode>-<dataset-name>-<YYYYMMDD-HHMM>`.
+3. Create `<workspace>/<run-name>/` and seed empty subdirs: `dataset/`, `outputs/`, `overfit/`.
+4. Create a todo list covering Phases 1–9 so the user can see progress.
+
+## Phase 1 — Intent
+
+Ask one question, framed in user terms (not jargon):
+
+> What do you want the model to learn? Examples: "generate videos from text," "make a LoRA of a specific style," "extend a video forward in time," "add sound effects to a silent video," "fill in masked regions of a video."
+
+Map the answer to one or more conditioning modes via `references/mode-selector.md`. If the requested capability has no mapping, go to **Escape Hatch**.
+
+### Plain concept/style LoRA → ask how it'll be used, default to I2V
+
+A "train a LoRA on X" request (a character/style/concept LoRA, no specific conditioning task) maps to either T2V or I2V. These aren't locked to inference: LoRA weights are pipeline-agnostic (the same checkpoint loads in both T2V and I2V inference), and the `i2v_lora` config applies first-frame conditioning with **`probability: 0.5`** — so it learns **both** conditioned (I2V) and unconditioned (T2V) generation in one run, and the first frame is taken automatically from each training clip (no extra data prep). I2V is therefore a versatile superset.
+
+Ask how they intend to use the result:
+
+> Will you generate videos from **text alone** (T2V), from a **starting image** (I2V), or **both / not sure**?
+
+- **Both / not sure (default):** use `i2v_lora` (probabilistic first-frame) — works for both at inference.
+- **I2V:** `i2v_lora`.
+- **Text only:** `t2v_lora`.
+
+(This only applies to plain concept/style LoRAs. A specific task — extension, inpainting, foley, IC-LoRA, etc. — maps directly to its mode via `mode-selector.md`; no usage question needed.)
+
+### Confirm the mode before proceeding
+
+Once the mode is determined, **state it plainly and confirm it** before doing any probing or work — a wrong inference is cheap to fix here and expensive later:
+
+> "Got it — I'll train an **I2V LoRA** (usable for both image-to-video and text-to-video at inference). Sound right?"
+
+The mode also appears in the plan (Phase 4), but confirm it here so the rest of the flow isn't built on a wrong guess.
+
+## Phase 2 — Probe
+
+No questions in this phase. Inspect what's already there. Use `references/onboarding.md` as the source of truth for the prerequisite checklist and what to do when something is missing.
+
+### Filesystem probe
+- If the user pointed at a path, classify: directory of raw videos, single long video, directory with a metadata file (CSV/JSON/JSONL), existing `.precomputed/`, partial outputs from a prior run.
+- For metadata files, identify columns: `video`/`media_path`, `caption`, `audio`, `reference_video`, `video_mask`, etc. (see `packages/ltx-trainer/docs/dataset-preparation.md`).
+- **Clip lengths (small datasets only):** for datasets up to a few hundred clips, `ffprobe` each clip's frame count and note the **minimum**. Clips shorter than the target frame bucket are silently skipped by `process_dataset.py`, so the shortest clip caps the achievable frame count — feed this into the Phase 3 resolution/frame-count choice (pick a bucket the clips support, or plan multi-bucket). Skip this per-clip probe for large datasets (too slow); rely instead on the post-preprocess reconciliation in Phase 7, which flags any dropped clips regardless of dataset size.
+- Check for an existing `<workspace>/<run-name>/` and whether `outputs/checkpoints/` contains a prior checkpoint (`lora_weights_step_*.safetensors` or `model_weights_step_*.safetensors`, plus a matching `training_state_step_*.pt` when resume state is enabled). This is a **resume candidate** — but note the trainer does *not* auto-resume from the output dir; resuming requires explicitly setting `model.load_checkpoint` in `config.yaml` to that checkpoint path. See `phases/launch-and-monitor.md` for the resume flow.
+- Check disk space at the workspace root. Preprocessed latents, checkpoints, and validation samples add up across a run; surface the available space alongside a rough sense of what one run consumes (one preprocessed bucket scales with sample count and resolution; each checkpoint is several GB), and warn the user if free space looks tight given their dataset size.
+
+### Hardware probe
+- `nvidia-smi --query-gpu=name,memory.total,driver_version --format=csv,noheader` → GPU model, count, VRAM. Stop the run if no CUDA GPU.
+- W&B login state — use wandb's **own** credential resolution (source-agnostic: covers env var, netrc, and the wandb settings file), not a hand-rolled netrc grep and **not** `wandb status` (which misleadingly reports `api_key: null` even when logged in):
+
+  ```bash
+  uv run python -c "import wandb; print(bool(wandb.Api().api_key))"   # True => logged in
+  ```
+
+  `True` → W&B is available, enable it. `False` → genuinely not logged in. If the check errors or is ambiguous, **ask the user** rather than silently disabling — a wrong "W&B off" assumption can incorrectly disable expected tracking.
+- Apply `references/hardware-profiles.md` to derive defaults (32GB / 40–60GB / 80GB+ VRAM tier).
+
+### Prerequisite probe (first-run sanity)
+- `uv` installed (`command -v uv`).
+- Workspace synced (lockfile present + `ltx-trainer` import works).
+- LTX-2 `.safetensors` and Gemma text encoder dir present in `/models/`, `~/models/`, or `$LTX_MODELS_DIR`.
+- Captioner availability: Gemini auth (`GEMINI_API_KEY`/`GOOGLE_API_KEY` or gcloud/Vertex) OR a ≥40 GiB GPU to host the Qwen3-Omni-30B vLLM server (FP8; bf16 needs ≥66 GiB). On typical consumer GPUs (24/32 GB), Gemini is effectively the only local-free option — see `references/onboarding.md`.
+
+For any missing prerequisite, **do not silently fail in a later phase**. Present the finding in chat with the specific next step from `references/onboarding.md`. The skill may offer to auto-install / auto-download missing pieces — but only ever after asking the user explicitly, one item at a time (model downloads are tens of GB each). Never auto-modify shell rc files or system config without consent.
+
+When everything (or what the user agreed to set up) is in place, fold the resolved paths into the plan's Assumptions section.
+
+### Pre-existing artifacts in the run dir
+
+If the run dir (or a user-supplied path) already contains artifacts from a prior session, classify each into one of three buckets — only the third prompts the user:
+
+1. **Deterministically verifiable → verify, then reuse silently (or stop).** `.precomputed/` latents: load a sample, check tensor shapes + modality coverage against the target (Phase 7). Match → reuse, no question. Mismatch/incomplete → stop and ask (reuse-at-old-spec / re-preprocess to a new dir / abort).
+2. **Cheap, fully-derived intermediates → regenerate silently.** `overfit/`, eval renders, sanity/temp configs, one-sample metadata. Delete and redo; don't ask.
+3. **Expensive AND not deterministically verifiable → ask.** Captions (`dataset.json`) and trained checkpoints/outputs. We can't programmatically decide whether existing captions or a half-finished run are what the user wants now, so surface what was found (counts, and how/when produced if knowable) and ask: reuse vs regenerate (for checkpoints: resume vs fresh).
+
+Principle: only ask when reuse-vs-regenerate is a genuine judgment call with cost either way. Never silently delete user-supplied data (hard invariant #1).
+
+## Phase 3 — Ask (minimum viable set)
+
+Ask only what cannot be inferred. Use `AskUserQuestion` with multiple-choice when possible. Typical questions:
+
+- **Target resolution / frame count** — propose a default per mode (e.g., `768x512x49` for T2V LoRA on consumer GPUs); offer overrides.
+- **Training steps** — if dataset size doesn't pin it, propose a default (e.g., 2000 for small LoRA datasets).
+- **LoRA trigger word / concept name** — only for style/concept LoRAs. Ask **only** for the word itself (or whether they want one). **Never** ask or mention *how* it's injected — it's always the `--lora-trigger` flag (passed to `process_dataset.py`, which forwards it to `process_captions.py` where the prepend happens); this is a fixed implementation detail. Presenting caption-injection as an option creates unnecessary confusion.
+- **Captioner backend** — only if more than one path is viable (e.g. a ≥40 GiB GPU can host the Qwen3-Omni-30B server *and* Gemini auth is available). On typical consumer GPUs, default to `gemini_flash` and surface that Gemini auth is required rather than asking.
+- **Model paths** — only if not found in the probe.
+
+Never ask anything answerable by `ls`, `nvidia-smi`, or the W&B credential check above.
+
+## Phase 4 — Plan
+
+Write the plan to `<workspace>/<run-name>/plan.md` using `references/plan-template.md`. Present it to the user in chat (don't just dump the file path). Wait for explicit approval before proceeding.
+
+If the user requests changes, edit the plan and re-present. Do not start Phase 5 until approval.
+
+## Phase 5 — Prepare Dataset
+
+If captioned metadata already exists with all required columns for the chosen mode, skip this phase. Otherwise follow `phases/prepare-dataset.md` — re-read it before acting.
+
+**Captioning gate:** caption a 3-sample spot-check first, show the captions in full, and **STOP for explicit user approval** before captioning the full set. The user must approve or give tuning instructions — never auto-proceed to the full pass. (Details in `phases/prepare-dataset.md`.)
+
+**Conditioning-inputs gate:** modes that need a reference (V2V/A2A/AV2AV IC-LoRA) or a mask (video/audio inpainting) require a per-sample input that encodes the user's specific idea. **Ask the user to provide it** — never invent the method (no defaulting to Canny/depth/generic masks). Only help generate it if the user explicitly asks, following *their* approach. Don't enter preprocessing for these modes without the input present. (Details in `phases/prepare-dataset.md` Step 4.)
+
+## Phase 6 — Sanity Check + Autotune (always run)
+
+**Tell the user what this phase is before starting it** — it's the most opaque to someone who doesn't know the design (see "Keep the User Informed"). In plain terms: a quick single-clip dry run at the target resolution to catch OOM/config errors cheaply, followed by trying a few config variants to pick the fastest stable one.
+
+Run **at the full target resolution** on **one sample** before the full preprocess. Purpose:
+1. Catch OOM / config errors before paying the full preprocessing cost.
+2. Empirically pick the fastest stable config via a small sweep.
+
+Steps:
+1. Pick one sample from the dataset metadata. Preprocess just that sample to `<workspace>/<run-name>/overfit/.precomputed/` (see `phases/preprocess-dataset.md` — use it in "one-sample" mode).
+2. Generate a temp config matching the planned full-run config but with `data.preprocessed_data_root: overfit/.precomputed`, `optimization.steps: 50`, `validation.interval: 50`, `checkpoints.interval: null`.
+3. Run the **baseline trial**: the conservative config from the matched VRAM tier (32GB tier = `t2v_lora_low_vram.yaml` defaults; 80GB+ tier = `t2v_lora.yaml` defaults — see `references/hardware-profiles.md`).
+4. **Success criteria** (all required):
+   - No OOM, no NaN loss, no crash.
+   - All 50 training steps complete.
+   - Validation sample at step 50 generates successfully (validation pass is a real OOM risk — do not skip).
+   - **For audio runs:** the one-sample `audio_latents/` is non-empty (the trainer log should report audio enabled). A joint/audio run that silently produced no audio latents is a failure even if steps complete — see the audio gate in `phases/preprocess-dataset.md`.
+   - **Loss is NOT a success criterion.** Loss can be non-monotonic even when training is healthy.
+5. **Autotune sweep** — incremental, capped at 5 trials total. Each trial = current best + one change. Stop on OOM (revert), no step-time improvement, or 5 trials:
+   - Trial 2: `quantization: null` (disable transformer quantization) if VRAM headroom allows.
+   - Trial 3: `optimizer_type: adamw` (disable 8-bit optimizer) if headroom allows.
+   - Trial 4: `batch_size` up (1 → 2 → 4). Adjust `gradient_accumulation_steps` proportionally to keep effective batch constant. **Note:** batch size can't be meaningfully tested on the one-sample set — defer it (test on the full set later, or just keep `batch_size: 1`, which is preferable for small concept-LoRA datasets anyway).
+   - **Do not sweep:** resolution (user decision), `acceleration.load_text_encoder_in_8bit` (one-time, no step-time impact), `enable_gradient_checkpointing` (the trainer's example configs ship with it on; on the 80GB+ tier you *may* try it off, but for the 22B model it usually OOMs even with tens of GB of apparent headroom — don't expect a win; never turn it off on the 32GB tier).
+6. Collect per trial: step time and peak VRAM. **Prefer the trainer's own end-of-run stats** (it prints total time / step time and peak GPU memory) — no external timing tool is needed (`/usr/bin/time` is often not installed). Append results to `<workspace>/<run-name>/autotune.log`.
+7. Winning trial's deltas are patched into the main `config.yaml`. Summarise the sweep to the user (one line per trial + winner).
+
+If the baseline trial fails, consult `references/troubleshooting.md`, propose a fix, re-run. Never push forward to the full preprocess after a failed sanity check.
+
+## Phase 7 — Full Preprocess
+
+Follow `phases/preprocess-dataset.md`. Re-read it before acting.
+
+**If `.precomputed/` already exists at the target path** (user-supplied or prior run), the phase verifies shapes and modality coverage before reuse. On mismatch it stops and asks — never silently overwrites.
+
+## Phase 8 — Launch & Monitor
+
+Follow `phases/launch-and-monitor.md`. Re-read it before acting. Surface the W&B URL (if enabled) and produce periodic status reports. At completion, the phase writes `<workspace>/<run-name>/outputs/run-summary.md`.
+
+## Phase 9 — Post-Train Validate
+
+After training finishes, follow `phases/post-train-validate.md`. Re-read it before acting. The phase renders the final LoRA against in-distribution, out-of-distribution, and held-out prompts; surfaces the MP4 paths to the user and exits.
+
+**Important constraint:** the post-train validate phase does not solicit a verdict and does not coach on causes for "soft" failures. Soft training quality has no reliable if/then rule book — the user inspects the renders and decides for themselves whether to ship, iterate, or change course. The orchestrator's job ends after Phase 9; iteration is a new invocation with a new run-name.
+
+## Monitor-Only Entry
+
+If invoked against an existing `<workspace>/<run-name>/` that already has a training process running or completed, **skip to Phase 8 in monitor-only mode** instead of restarting anything: report current step, recent loss, ETA, checkpoint list, W&B URL, and the resume command. Distinguish "live process" vs "stopped run with checkpoints" and offer the appropriate next action.
+
+## Escape Hatch: Unsupported Modes
+
+If Phase 1 intent doesn't map to any combination of supported `flexible`-strategy conditions:
+
+1. Stop. Do not edit `packages/ltx-trainer/` on your own.
+2. Explain what's missing in concrete terms: "you want X. The trainer supports A, B, C via conditions D, E, F. X requires a new condition / strategy."
+3. Identify the code change needed (typically a new `Condition` subclass in `ltx_trainer/training_strategies/flexible.py` plus schema wiring in `config.py`).
+4. Ask the user explicitly: "Proceed with the code change, do it yourself, or abort?"
+5. Only on explicit consent, drop out of this skill's orchestrator and edit code as a normal agent task. After the change lands and is tested, return here.
+
+## Ask-vs-Assume Cheat Sheet
+
+| Decision | How |
+|----------|-----|
+| Precision, quantization, optimizer, grad checkpointing | Assume from matched VRAM tier. List in plan's "Assumptions". |
+| Checkpoint/validation interval, seed, W&B project name, output dir | Assume sensible defaults. List in "Assumptions". |
+| Target resolution / frame count | Ask if not given; propose mode-appropriate default. |
+| Step count | Ask if dataset size doesn't pin it. |
+| LoRA trigger word / concept name | Ask for the word only (style/concept LoRAs). Never ask *how* it's injected — always `--lora-trigger`. |
+| Captioner backend | Ask only if multiple backends are viable. |
+| Anything answerable by `ls`, `nvidia-smi`, or the W&B credential check | Never ask. Probe. |
+
+## The Two `load_text_encoder_in_8bit` Flags
+
+Same name, different layers — do not conflate:
+
+| Flag | Layer | Effect |
+|------|-------|--------|
+| `process_dataset.py --load-text-encoder-in-8bit` | Preprocessing CLI | Memory during caption-embedding precompute (Phase 7). One-time per dataset. |
+| `acceleration.load_text_encoder_in_8bit` (trainer YAML) | Trainer config | Memory during validation-sample prompt-embedding caching at training start (Phase 8). One-time per run. |
+
+Both are one-time costs and neither affects per-step training speed. Default per the trainer's shipped configs: **ON** on the 32GB tier (matches `t2v_lora_low_vram.yaml`), **OFF** on the 80GB+ tier (matches `t2v_lora.yaml`). No measured guidance for the 40–60GB tier beyond starting from the 32GB tier and autotuning.
+
+## Workspace Layout
+
+```
+<workspace>/<run-name>/
+  plan.md                  # the approved plan
+  config.yaml              # generated training config (NOT in packages/ltx-trainer/configs/)
+  autotune.log             # per-trial sweep results
+  dataset/
+    dataset.json           # captions + media paths (training split)
+    holdout.jsonl          # held-out split (if reserved)
+    videos/                # source media copies (NO derived files here)
+    .precomputed/          # latents/ audio_latents/ conditions/ (+ references/masks per mode)
+  outputs/
+    checkpoints/           # training checkpoints + states
+    samples/               # in-training validation samples (step_*)
+    eval/                  # Phase 9: in-distribution/ out-of-distribution/ held-out/ + prompts.json
+    run-summary.md         # written at completion
+  logs/                    # all run logs
+  overfit/                 # Phase 6 scratch (one-sample preprocess + sanity/autotune runs)
+```
+
+`<run-name>` default: `<mode>-<dataset-name>-<YYYYMMDD-HHMM>`. Surface in the plan; user may rename.
+
+### Workspace hygiene (keep it clean)
+
+- **Don't create undocumented directories** (e.g. an ad-hoc `scratch/`). Intermediates belong under `overfit/` (sanity/autotune scratch) or a `/tmp` tempdir — not loose in the run dir or the repo root.
+- **Never write derived files into `dataset/videos/`** (the source media dir). Latents go under `.precomputed/`; one-sample/eval metadata goes under `overfit/`, not `dataset/`.
+- **One canonical manifest per artifact** — don't leave duplicate `*_prompts.json` / metadata copies.
+- **Clean up phase byproducts:** the Phase 9 eval must delete its validate-only trainer cruft (see `phases/post-train-validate.md`); `overfit/` is scratch and may be removed after a successful run. The final tree should look like the layout above — no stray `.pt`/`.wav` files, no `eval/checkpoints/`, no duplicate manifests.
+
+## References
+
+- `references/mode-selector.md` — user intent → conditioning mode mapping, LoRA rank guidance (read in Phase 1).
+- `references/onboarding.md` — first-run prerequisite checklist, model download paths, captioner graceful degradation (read in Phase 2).
+- `references/hardware-profiles.md` — GPU VRAM → tier + config defaults (read in Phase 2).
+- `references/config-patching.md` — safe YAML edits + schema constraints (read whenever editing `config.yaml`).
+- `references/troubleshooting.md` — OOM, NaN, validation failures, resume (read on any failure).
+- `references/plan-template.md` — exact plan.md format (read in Phase 4).
+
+## Phase Procedures
+
+These are procedure documents the orchestrator reads when entering each phase. They are not standalone skills — they're never invoked by Claude's skill-discovery system. The orchestrator opens them via the `Read` tool and follows the instructions inline.
+
+- `phases/prepare-dataset.md` — Phase 5: scenes, captioner iteration, IC-LoRA references, metadata, holdout split.
+- `phases/preprocess-dataset.md` — Phases 6 (one-sample) & 7 (full): `process_dataset.py` orchestration, existing-data verification.
+- `phases/launch-and-monitor.md` — Phase 8: launch command, accelerate, W&B, status reports, run-summary writing.
+- `phases/post-train-validate.md` — Phase 9: render the final LoRA against three prompt categories; surface paths only.

.claude/skills/train-model/phases/launch-and-monitor.md

@@ -0,0 +1,235 @@
+# Phase 8 — Launch & Monitor
+
+Procedure document for the `train-model` orchestrator (Phase 8 + monitor-only re-entry). Read this file in full before acting on launch/monitor.
+
+Goal: start the training job, surface the W&B URL, produce periodic status reports, write `run-summary.md` at completion.
+
+The orchestrator's hard invariants apply (see `../SKILL.md`).
+
+## Launch — Single GPU
+
+```bash
+cd packages/ltx-trainer
+uv run python scripts/train.py "<workspace>/<run-name>/config.yaml"
+```
+
+## Launch — Multi-GPU
+
+Use Accelerate. For LoRA, DDP (default) is fine. For full FT, use FSDP.
+
+```bash
+cd packages/ltx-trainer
+
+# DDP (LoRA, multi-GPU)
+uv run accelerate launch scripts/train.py "<workspace>/<run-name>/config.yaml"
+
+# FSDP (full FT, multi-GPU)
+uv run accelerate launch \
+  --config_file configs/accelerate/fsdp.yaml \
+  scripts/train.py "<workspace>/<run-name>/config.yaml"
+```
+
+**Pass `--disable-progress-bars` to `train.py` whenever stdout is redirected to a log file** (every background run) or running multi-GPU. The Rich progress bar rewrites a single line with carriage returns and does **not** flush parseable newlines to a redirected log, so without this flag the log shows no step/loss lines and you're forced to poll `nvidia-smi`. With it, step/loss lines are written normally and the log is greppable.
+
+## Pre-Launch Checks (every launch)
+
+1. Run the self-check from `references/config-patching.md` (paths exist, frame/resolution constraints, generated modalities have matching latents dirs, references/masks dirs present for conditional modes).
+2. Confirm `nvidia-smi` shows expected GPUs available (not occupied by another process).
+3. If `wandb.enabled: true`, confirm credentials still resolve: `uv run python -c "import wandb; print(bool(wandb.Api().api_key))"` → `True`. (Don't use `wandb status`.) If `False`, surface to the user before launching — they may want to `wandb login` or run without tracking.
+
+## Run In Background, Monitor Foreground
+
+Long training runs should not block the agent's response loop, and they must survive past the launching turn.
+
+**Prefer the agent's managed/native background-shell mechanism** (the harness facility for long-running background commands — output streaming + PID/exit tracking, survives across turns). It's the reliable way to launch training: it stays alive, streams to a log the agent can poll, and reports completion. Launch the training command through that mechanism, writing to `<workspace>/<run-name>/logs/train.log` with `--disable-progress-bars`, using **absolute paths** for the config and log (`uv run --directory packages/ltx-trainer` changes the cwd, so a relative config path won't resolve).
+
+**`nohup ... &` is a last-resort fallback only.** Detached jobs can be harder to track and may not survive environment/session cleanup, so only use it if no managed background mechanism is available, and verify the PID is still alive afterward:
+
+```bash
+# Fallback ONLY — prefer the managed background shell above.
+mkdir -p "<workspace>/<run-name>/logs"
+nohup uv run --directory packages/ltx-trainer python scripts/train.py \
+  "<ABSOLUTE path>/<run-name>/config.yaml" --disable-progress-bars \
+  > "<ABSOLUTE path>/<run-name>/logs/train.log" 2>&1 &
+echo $! > "<workspace>/<run-name>/logs/train.pid"
+```
+
+## Status Report
+
+Produce on user request (or every <interval> automatically). Pull from:
+
+- **W&B run URL:** First lines of `train.log` after init, or `wandb.run.url` from a `wandb` Python snippet. Surface as a clickable URL.
+- **Latest step:** `tail -n 200 "<workspace>/<run-name>/logs/train.log" | grep -oE "step [0-9]+" | tail -1`.
+- **Recent loss:** `tail -n 200 "<workspace>/<run-name>/logs/train.log" | grep -oE "loss[: ]+[0-9.]+" | tail -5`.
+- **Checkpoints saved:** `ls -1t "<workspace>/<run-name>/outputs/checkpoints/" 2>/dev/null`.
+- **Validation samples:** `ls -1t "<workspace>/<run-name>/outputs/samples/" 2>/dev/null`.
+- **GPU utilization snapshot:** `nvidia-smi --query-gpu=name,utilization.gpu,memory.used,memory.total --format=csv,noheader`.
+- **ETA:** only report one **grounded in real numbers the trainer has actually emitted** — do not invent or "educated-guess" an ETA before the trainer has produced per-step timings. Compute as `(total_steps - current_step) * recent_avg_step_time`, where `recent_avg_step_time` is measured from **steady-state training steps** (the trainer's reported per-step time or log timestamps), **excluding** one-time setup that doesn't repeat per step: model loading, the step-0 validation pass, and periodic validation passes. The trainer's own early ETA projection is skewed high by the slow step-0 validation and settles after a few steps — wait for it to settle rather than quoting the inflated early figure. Until real step timings exist, say "measuring step time…" rather than guessing a duration.
+
+Format the report tightly — one block, no fluff:
+
+```
+Step 1240 / 2000 (62%) — loss ~0.072 (last 5: 0.071, 0.073, 0.069, 0.075, 0.072)
+ETA: ~1h 24m  |  GPU: 91% util, 39.2 / 48.0 GB
+Latest checkpoint: lora_weights_step_01000.safetensors
+Latest validation: samples/step_01200_*.mp4
+W&B: https://wandb.ai/<entity>/<project>/runs/<id>
+```
+
+## Monitor-Only Mode
+
+Invoked when the orchestrator detects an existing `<workspace>/<run-name>/` with checkpoints or a running process.
+
+1. Check if the training process is live: `[ -f .../logs/train.pid ] && kill -0 $(cat .../logs/train.pid) 2>/dev/null && echo LIVE || echo STOPPED`.
+2. Produce the same status report as above.
+3. If STOPPED:
+   - Compute step from last checkpoint.
+   - Find the latest checkpoint pair (`lora_weights_step_*.safetensors` or `model_weights_step_*.safetensors`, plus a matching `training_state_step_*.pt` when resume state is enabled) under `<workspace>/<run-name>/outputs/checkpoints/`.
+   - Patch `model.load_checkpoint` in `config.yaml` to point at that checkpoint file (this is the only way the trainer knows to resume — there's no auto-detection from `output_dir`).
+   - Surface the resume command: `uv run python scripts/train.py "<workspace>/<run-name>/config.yaml"`.
+   - Ask the user to confirm both the patch and the launch before applying.
+
+## Resume
+
+The trainer **does not auto-resume from `output_dir`**. To resume an interrupted run:
+
+1. Set `model.load_checkpoint` in `config.yaml` to the latest checkpoint file (e.g. `<workspace>/<run-name>/outputs/checkpoints/lora_weights_step_02000.safetensors`).
+2. Launch normally. The trainer loads those weights, then looks for a matching `training_state_step_*.pt` **next to the loaded checkpoint** and restores optimizer/scheduler/step state from it. If the state file is missing, weights load but training starts from step 0.
+3. To load the weights but skip the state restore (e.g. for branching off into a new run from a known-good checkpoint), set `checkpoints.no_resume: true`.
+
+Always ask the user before patching `model.load_checkpoint` or setting `no_resume` — checkpoints are precious.
+
+## Failure During Training
+
+If the training process exits non-zero:
+
+1. Tail the log and identify the error type.
+2. Cross-reference `references/troubleshooting.md`.
+3. Propose a config fix (with the exact diff to `config.yaml`).
+4. Ask the user before applying. Then resume with the patched config.
+
+Never silently restart a failed training run without acknowledging the failure to the user.
+
+## After Training Completes
+
+1. Write a **run summary** to `<workspace>/<run-name>/outputs/run-summary.md` (see below) so the user can find their bearings months later without rereading the trainer docs.
+2. Show final checkpoint path and step count.
+3. Show W&B URL if enabled.
+4. Return control to the orchestrator (Phase 9 — post-train validate runs next).
+
+### Writing `run-summary.md`
+
+The summary is the **landing page** for this run. Anyone (including the user months from now) should be able to read it and understand: what was trained, on what data, with what config, where everything lives, how to use the result, and how to continue. The trainer doesn't produce this — the skill does.
+
+Template (fill from `plan.md`, `config.yaml`, `autotune.log`, dataset metadata, training log):
+
+```markdown
+# <run-name>
+
+**Trained:** <YYYY-MM-DD HH:MM> on <GPU(s)>
+**Final checkpoint:** `outputs/checkpoints/<filename>.safetensors` (step <N>)
+
+## What this LoRA does
+
+<One paragraph from the plan's Goal section — restating the user's intent.>
+
+## Trigger word
+
+`<trigger>` — include in prompts at inference time. (Omit this section if no trigger word.)
+
+## Mode
+
+<Mode name> (<lora|full>). Conditioning: <list of conditions, or "none">.
+
+## Dataset
+
+- Source: `<absolute path>`
+- Captioning: <`qwen_omni` (Qwen3-Omni-30B via vLLM) | `gemini_flash` | "user-supplied">
+  - Captioner instruction used: <verbatim string, or "default">
+- Samples: <N training> + <K held-out> at <W>x<H>x<F>
+- Preprocessed to: `dataset/.precomputed/`
+
+## Training config
+
+Final values after autotune (deltas from baseline noted in `autotune.log`):
+
+| Field | Value |
+|-------|-------|
+| Optimizer | <...> |
+| Mixed precision | <bf16/fp16> |
+| Quantization | <...> |
+| Gradient checkpointing | <on/off> |
+| Batch size × grad accum | <B> × <A> (effective <BxA>) |
+| LoRA rank / alpha | <R> / <A> (or "full FT") |
+| LoRA target modules | <list> (or "n/a") |
+| Steps | <N> |
+| Learning rate | <value> |
+| Step time (final) | ~<T>s |
+| Peak VRAM | ~<V> GB |
+
+Full config: `<workspace>/<run-name>/config.yaml`
+
+## Outputs
+
+- Checkpoints: `outputs/checkpoints/`
+- Validation samples (during training): `outputs/samples/`
+- Post-train eval renders (if Phase 9 ran): `outputs/eval/`
+- W&B run: <url, or "(W&B not enabled)">
+
+## How to use this checkpoint
+
+For inference, point `packages/ltx-pipelines/` at the final checkpoint. Example invocation:
+
+\`\`\`bash
+# (Minimal sketch — adapt to the pipeline you're using.)
+# load base LTX-2 model + apply this LoRA from outputs/checkpoints/<filename>.safetensors
+\`\`\`
+
+## How to continue training
+
+To resume from the final checkpoint (e.g. more steps, different LR), edit `config.yaml`:
+
+\`\`\`yaml
+model:
+  load_checkpoint: "<absolute path to outputs/checkpoints/<filename>.safetensors>"
+optimization:
+  steps: <new total>          # trainer resumes optimizer/scheduler/step from the training_state_step_*.pt sitting next to the checkpoint above
+\`\`\`
+
+Then re-launch with the same command in the "Launched with" section below.
+
+## How this was launched
+
+\`\`\`
+<exact command used, with the workspace's absolute config path>
+\`\`\`
+
+## Reproducibility
+
+- Seed: <value>
+- Workspace: `<absolute path>`
+- Repo commit at launch: `<git rev-parse HEAD output>`
+- LTX-2 model: `<model.model_path from config>`
+- Text encoder: `<model.text_encoder_path from config>`
+```
+
+Write the file using the `Write` tool. Don't embed it in a heredoc — the markdown nested in this skill is illustrative; fill the template with real values from the run's artifacts.
+
+### Next steps (surface to user)
+
+After writing the summary, point the user at:
+- The summary file path.
+- The final checkpoint path.
+- The W&B URL if enabled.
+- The upcoming Phase 9 (post-train validate) — the orchestrator handles the transition.
+
+Suggest, but don't run:
+- Test inference with `packages/ltx-pipelines/`.
+- Push to HF Hub via the trainer's `hub.push_to_hub` config (a separate, lightweight re-launch).
+- Continue training from the final checkpoint (see summary's "How to continue training" section).
+
+## Do Not
+
+- Do not modify `output_dir` contents after a run completes — checkpoints belong to the user now.
+- Do not start a second training run into the same `output_dir` without explicit user approval. Resume requires patching `model.load_checkpoint` (the trainer does not auto-detect prior checkpoints); a true fresh-start from the same dir additionally needs `checkpoints.no_resume: true`.
+- Do not auto-restart a failed run without diagnosis and user approval.

.claude/skills/train-model/phases/post-train-validate.md

@@ -0,0 +1,171 @@
+# Phase 9 — Post-Train Validate
+
+Procedure document for the `train-model` orchestrator (Phase 9). Read this file in full before acting on post-train validation.
+
+Goal: render the final checkpoint against three prompt categories so the user can inspect the result and form their own judgement. Save outputs in an organized layout. **Do not** prompt for pass/fail verdicts, do not infer causes for failures, do not suggest fixes — soft training failures don't have a clean if/then rule book, and pretending otherwise wastes the user's time.
+
+The orchestrator's hard invariants apply (see `../SKILL.md`).
+
+## What this phase does
+
+1. Collect prompts for three categories.
+2. Render the final LoRA against all collected prompts.
+3. Save outputs under `<workspace>/<run-name>/outputs/eval/<category>/`.
+4. Print the paths and exit.
+
+## Categories
+
+### 1 — In-distribution
+
+A few captions from the training set itself. Tests whether the model learned what it was shown.
+
+- Default: 3 random captions from the dataset metadata (seed 42 for reproducibility).
+- Source: `<workspace>/<run-name>/dataset/dataset.json` (the captions used for training, after the held-out split).
+
+### 2 — Out-of-distribution
+
+Prompts the model has never seen, but in the same domain. Tests whether the model generalizes the concept beyond memorized phrasings.
+
+Ask the user once:
+
+> "For out-of-distribution validation, paste 2–3 prompts you'd realistically want to generate at inference time. (If you don't have any specific ones in mind, reply 'default' — I'll use a few generic prompts that include the trigger word.)"
+
+- On `default`: synthesize 3 short prompts using the LoRA's trigger word and a generic scene context (e.g., "<trigger> walking in a forest at dawn"). Note these are generic — they're better than nothing, but real user-style prompts make a stronger test.
+- Otherwise: use the user's prompts verbatim.
+- **For a run with a generated audio modality, the synthesized prompts must describe the audio** — matching how the training captions describe it (inspect a few from `dataset.json` first). A prompt with no audio direction leaves the audio branch unguided and the generated audio comes out poor. E.g. for the talking-head case include spoken-voice/room-tone direction; for music/ambience/foley describe the sound character. (Categories 1 and 3 reuse the real captions verbatim, so they already carry audio description — this only applies to the synthesized Category 2 prompts.) If the user pasted their own prompts and it's an audio run, and they omitted audio direction, note that the audio may be weak without it.
+
+### 3 — Held-out
+
+Captions from samples that were never seen during training. Tests true generalization, not memorization.
+
+- Source: `<workspace>/<run-name>/dataset/holdout.jsonl` (written by `prepare-dataset` Step 5).
+- Use all entries if there are ≤5; otherwise sample 5 with seed 42.
+- **If holdout doesn't exist** (dataset was too small, or user-skipped during prepare): print a clear note in the output summary — *"Held-out evaluation skipped: no holdout set was reserved for this run. The post-train eval only covers Categories 1 and 2."* Don't synthesize substitutes.
+
+## Rendering mechanism
+
+Use the trainer's existing validation infrastructure rather than wiring up `ltx-pipelines` from scratch. Create a temporary "validate-only" config and run the trainer with it.
+
+### Step 1 — Build eval config
+
+Copy `<workspace>/<run-name>/config.yaml` to `<workspace>/<run-name>/eval-config.yaml`. Patch:
+
+```yaml
+model:
+  load_checkpoint: "<absolute path to outputs/checkpoints/<final-lora>.safetensors>"
+
+optimization:
+  steps: 1                     # we don't want to train; we want validation to fire
+  # Keep batch_size/grad_accum at the run's autotuned values to match its VRAM footprint.
+
+validation:
+  skip_initial_validation: false
+  interval: 1                  # run validation at step 0 (and at the only training step)
+  samples:
+    # Inject all collected prompts here, tagged by category in the prompt itself
+    # so the output filenames make the category obvious.
+    - prompt: "[CAT1-IND] <caption from dataset.json>"
+    # ... repeat for each prompt in all three categories
+  # Keep video_dims, frame_rate, guidance/STG settings as the trained config.
+  # Keep generate_audio consistent with the trained modality config.
+
+checkpoints:
+  interval: null               # do not save more checkpoints
+  no_resume: true              # load the LoRA's weights but do not restore optimizer/scheduler/step state
+
+output_dir: "<workspace>/<run-name>/outputs/eval"
+```
+
+The `[CAT1-IND]`, `[CAT2-OOD]`, `[CAT3-HELDOUT]` tags in the prompt strings make the output MP4 filenames self-describing in the trainer's validation sample directory.
+
+**Attach the mode's conditions to each sample.** A bare `prompt` only validates a pure text-to-X mode (T2V, T2A). For any conditioned mode, the trained model expects the same conditioning at validation time — a prompt with no conditions tests a different task than what was trained, and conditioned modes may fail outright. Add the `conditions` list that matches the run's mode (the trained `config.yaml` `training_strategy` and the example config for the mode are the reference):
+
+| Mode | Add to each sample |
+|------|--------------------|
+| I2V | `conditions: [{type: first_frame, image_or_video: <frame/clip path>}]` |
+| Video extension / suffix | `conditions: [{type: prefix|suffix, ...}]` |
+| V2V / AV2AV IC-LoRA | `conditions: [{type: reference, ...}]` (point at a held-out reference) |
+| V2A (foley) | `conditions: [{type: video_to_audio, ...}]` |
+| A2V | `conditions: [{type: audio_to_video, ...}]` |
+| Inpainting (video/audio) | `conditions: [{type: mask, ...}]` |
+| Outpainting | `conditions: [{type: spatial_crop, ...}]` |
+| A2A IC-LoRA | `conditions: [{type: reference, ...}]` (held-out reference audio) |
+| T2V, T2A | none — a bare `prompt` is correct |
+
+Mirror the condition shapes used in the mode's example config under `packages/ltx-trainer/configs/`. For held-out (Category 3) and OOD (Category 2) samples on conditioned modes, draw the conditioning media from the held-out set so the eval stays out-of-distribution.
+
+### Step 2 — Run the trainer in validate-only mode
+
+```bash
+cd packages/ltx-trainer
+uv run python scripts/train.py "<workspace>/<run-name>/eval-config.yaml"
+```
+
+The trainer will load the LoRA, run initial validation against all the prompts, do one trivial training step (which we discard), and exit. Validation samples land in `<workspace>/<run-name>/outputs/eval/samples/`.
+
+### Step 3 — Organize outputs and clean up trainer cruft
+
+The validate-only run is a trainer run, so it inevitably writes throwaway artifacts: an indexed `samples/` dir, a forced final checkpoint (the trainer **always** saves one at the end, regardless of `checkpoints.interval`), and a `training_config.yaml`. Don't leave these around or duplicate the renders.
+
+1. **Move** (don't copy) each generated MP4 from the trainer's indexed `samples/` dir into the category layout, naming by category + a slug of the prompt. Use the index→category mapping you built when constructing `validation.samples`.
+2. **Delete the trainer cruft** from the eval dir once the renders are moved: the indexed `samples/` dir, the forced `checkpoints/` dir, and `training_config.yaml`. (These are byproducts of the validate-only hack — there's no config flag to suppress the final-checkpoint save, so clean it up here.)
+3. Write **one** manifest, `outputs/eval/prompts.json` (filename → full prompt + category). Don't leave a second copy elsewhere.
+4. Remove the temporary `eval-config.yaml` (or keep it under the run's scratch, not in `outputs/`).
+
+Final eval layout — exactly this, nothing else:
+
+```
+<workspace>/<run-name>/outputs/eval/
+  in-distribution/    <NN>_<prompt-slug>.mp4 ...
+  out-of-distribution/ <NN>_<prompt-slug>.mp4 ...
+  held-out/           <NN>_<prompt-slug>.mp4 ...   # only if a holdout set existed
+  prompts.json        # filename -> full prompt + category (single manifest)
+```
+
+No `eval/samples/`, no `eval/checkpoints/`, no `eval/training_config.yaml`, no duplicate manifest.
+
+### Step 4 — Surface paths
+
+Print a tight block, no judgement, no follow-up question:
+
+```
+Post-train evaluation complete.
+
+In-distribution renders (<K> samples):
+  <workspace>/<run-name>/outputs/eval/in-distribution/
+
+Out-of-distribution renders (<M> samples):
+  <workspace>/<run-name>/outputs/eval/out-of-distribution/
+
+Held-out renders (<N> samples):
+  <workspace>/<run-name>/outputs/eval/held-out/    # or: "(skipped — no holdout set)"
+
+Open the MP4s and decide for yourself whether the model is good. Soft
+training quality is judged by watching the videos, not by a checklist —
+there's no substitute for your own eyes here.
+```
+
+Return control to the orchestrator. The orchestrator's run is now complete.
+
+## What this phase does NOT do
+
+- Does not ask "is this good?" / "pass / partial / fail?".
+- Does not infer failure causes.
+- Does not suggest fixes, follow-up runs, hyperparameter changes, dataset changes.
+- Does not write any verdict to `run-summary.md` or elsewhere.
+- Does not delete or modify training checkpoints.
+- Does not push to any remote / cloud / registry.
+
+The user looks at the videos and makes their own call. If they want to iterate, they re-invoke the orchestrator with a new run-name.
+
+## Failure modes
+
+- **Final checkpoint missing.** Surface and stop. Don't render against an intermediate checkpoint without explicit user consent.
+- **`load_checkpoint` OOM at inference time.** Lower `validation.video_dims` in the eval config (smaller renders are still useful for a sanity look). Retry once. If still OOM, surface the failure and let the user run inference manually via `packages/ltx-pipelines/`.
+- **All renders look broken/black.** May be an inference-pipeline-side issue rather than a training failure. Mention in the output block: *"If renders look broken across all categories, try `packages/ltx-pipelines/` directly to rule out a pipeline issue."* Then exit. Do not investigate further.
+
+## Do not
+
+- Do not skip Category 1 or 2. They're cheap and informative.
+- Do not invent a held-out set if `holdout.jsonl` is missing — the prepare-dataset step decides that.
+- Do not coach the user on what "good" means for their use case.

.claude/skills/train-model/phases/prepare-dataset.md

@@ -0,0 +1,245 @@
+# Phase 5 — Prepare Dataset
+
+Procedure document for the `train-model` orchestrator (Phase 5). Read this file in full before acting on the prepare-dataset phase.
+
+Goal: produce a captioned, complete dataset metadata file at `<workspace>/<run-name>/dataset/dataset.json` consumable by `process_dataset.py`. Idempotent — re-runs skip work already done.
+
+The orchestrator's hard invariants apply (see `../SKILL.md`), especially: **no file mutation outside the workspace without explicit user approval.**
+
+## Inputs
+
+The orchestrator passes:
+- Source path (directory of videos / single video / pre-existing metadata file).
+- Target mode (T2V, I2V, V2V IC-LoRA, V2A, etc.) — determines which columns are required.
+- Captioner backend choice (Qwen3-Omni local vLLM server / Gemini Flash cloud / skip).
+- Workspace path `<workspace>/<run-name>/`.
+
+## Required Columns by Mode
+
+`process_dataset.py` detects columns by convention and resolves each to a role. The media column may be `video` **or** `audio`. When a dataset has a `video` column with an audio track and no separate `audio` column, audio is **auto-extracted** from the video (unless `--skip-audio`), so an explicit `audio` column is only needed when the audio lives in separate files.
+
+**Video-generating modes** (need a `video` column):
+
+| Mode | Required | Optional |
+|------|----------|----------|
+| T2V, I2V, video extension/suffix | `video`, `caption` | `audio` (else auto-extracted) |
+| Video outpainting | `video`, `caption` | |
+| Video inpainting | `video`, `caption`, `video_mask` | |
+| V2A (foley) | `video`, `caption` | `audio` (target; else auto-extracted) |
+| A2V | `video`, `caption` | `audio` (else auto-extracted from video) |
+| V2V IC-LoRA | `video`, `caption`, `reference_video` | |
+| AV2AV IC-LoRA | `video`, `caption`, `reference_video`, `reference_audio` | `audio` (else auto-extracted) |
+
+**Audio-only modes** (no `video` column — the media column is `audio`):
+
+| Mode | Required | Optional |
+|------|----------|----------|
+| T2A | `audio`, `caption` | |
+| Audio extension/suffix | `audio`, `caption` | |
+| Audio inpainting | `audio`, `caption`, `audio_mask` | |
+| A2A IC-LoRA | `audio`, `caption`, `reference_audio` | |
+
+Aliases: `media_path` for `video`, `ref_media_path` for `reference_video`.
+
+## Workflow
+
+### Step 1 — Classify source
+
+```bash
+# If source is a file, identify type:
+file "<source>"
+# If source is a directory, count media:
+find "<source>" -maxdepth 1 -type f \( -name "*.mp4" -o -name "*.mov" -o -name "*.webm" \) | wc -l
+```
+
+Cases:
+- **Pre-existing metadata file** (CSV/JSON/JSONL) → copy to `<workspace>/<run-name>/dataset/dataset.json`, audit columns. Skip to Step 4.
+- **Directory of short scenes** → skip Step 2, go to Step 3.
+- **Directory containing long videos** → run Step 2.
+- **Single long video** → run Step 2.
+
+**Stage the media under `dataset/` before captioning — don't discover this by failing.** Both `caption_videos.py` and `process_dataset.py` reference media by paths **relative to the metadata file's own directory**, so the media must live under `<workspace>/<run-name>/dataset/`. Stage it up front into `dataset/videos/` (and write metadata paths relative to `dataset/`, e.g. `videos/1.mp4`):
+- Prefer **symlinks** (instant, no disk cost): `ln -s <abs-source>/<clip> <workspace>/<run-name>/dataset/videos/<clip>`. Symlinks pointing at the original source location work correctly.
+- Use a **copy** instead if the workspace and source are on different filesystems or the source may move/change.
+- Never caption or preprocess directly against an out-of-tree source path (e.g. `/path/to/source-videos`) — it will fail the relative-path resolution. The original source is left untouched either way.
+(Scene-split output in Step 2 already lands under `dataset/scenes/`, which satisfies this.)
+
+### Step 2 — Scene splitting (only for long videos)
+
+`split_scenes.py` takes **one video file** at a time (`video_path` and `output_dir` are both positional arguments). When the source is a directory of long videos, iterate over each file. To drop scenes shorter than 2 seconds, use `--filter-shorter-than 2s` (the `--min-scene-length` option is an integer **frame** count, not seconds — don't pass a float).
+
+```bash
+cd packages/ltx-trainer
+# Single file:
+uv run python scripts/split_scenes.py \
+  "<video-file>" \
+  "<workspace>/<run-name>/dataset/scenes" \
+  --filter-shorter-than 2s
+
+# Directory of long videos — iterate:
+for f in "<source>"/*.mp4 "<source>"/*.mov "<source>"/*.webm; do
+  [ -e "$f" ] || continue
+  uv run python scripts/split_scenes.py "$f" \
+    "<workspace>/<run-name>/dataset/scenes" \
+    --filter-shorter-than 2s
+done
+```
+
+Result: scenes saved to `<workspace>/<run-name>/dataset/scenes/`. Pass that directory to Step 3.
+
+### Step 3 — Captioning
+
+Skip entirely if a metadata file with all required `caption` entries already exists.
+
+**Use the captioner's default instruction.** `caption_videos.py` ships a well-tuned default caption prompt — use it as-is (do **not** pass `--instruction`). Captioning runs in two phases: a small **spot-check pass** so the user can confirm the captions look sane, then a **full pass** on the rest.
+
+A custom `--instruction` is the exception, not the norm. Only use one when:
+- the **nature of the dataset genuinely demands it** (e.g. a narrow domain the default prompt won't describe well), or
+- the **user, after seeing the spot-check captions, explicitly asks** for a change (e.g. "too much background detail").
+
+Do not invent a custom instruction pre-emptively, and in particular **do not bake a subject name / trigger word into the captions via `--instruction`** — the trigger word is handled separately at preprocessing (see "Trigger word" below).
+
+#### Choosing a backend
+
+Two backends, with very different hardware needs:
+
+- **`qwen_omni` (local, default):** Qwen3-Omni-30B-A3B-Thinking served by a local vLLM HTTP server (`serve_captioner.py`). ~65 GiB model download. Default **FP8** quantization uses ~31 GiB of weights and **fits on a 40 GiB GPU** (plus KV cache); **bf16** uses ~60 GiB and needs **≥66 GiB free VRAM**.
+- **`gemini_flash` (cloud):** Google `gemini-3.5-flash`. No local model, runs anywhere, parallelisable with `--num-workers`.
+
+**Steer modest hardware to Gemini.** If the GPU is below ~40 GiB (i.e. typical consumer cards — 24 GB / 32 GB), it can't host even the FP8 server, so the local captioner isn't an option — recommend `gemini_flash` and tell the user they'll need Gemini auth: either a `GEMINI_API_KEY`/`GOOGLE_API_KEY` (get one at <https://aistudio.google.com/apikey>) or working gcloud/Vertex AI credentials. If they can't or won't set that up and the hardware can't run Qwen3, the only remaining path is bringing their own captions in the dataset metadata (skip captioning entirely). On a 40 GiB+ GPU the local server is viable (FP8); bf16 needs an 80GB-class card.
+
+#### Qwen server prerequisite (qwen_omni only)
+
+The local backend talks to a vLLM server that must already be running. Launch it once in a **separate terminal** (it stays loaded across captioning runs):
+
+```bash
+cd packages/ltx-trainer
+uv run python scripts/serve_captioner.py        # FP8 by default, serves on http://127.0.0.1:8001/v1
+# bf16 (needs >= 66 GiB free VRAM):  --quantization bf16
+# different port/interface:           --port 9000 --host 0.0.0.0
+```
+
+First launch downloads the model (~65 GiB). `caption_videos.py` reaches it via `--vllm-url` (default `http://127.0.0.1:8001/v1`). Skip this entirely when using `gemini_flash`.
+
+#### 3a — Spot-check pass (3 samples)
+
+Caption 3 samples with the **default prompt** (no `--instruction`) to confirm the captioner is producing sane output before committing to the whole set.
+
+```bash
+cd packages/ltx-trainer
+# qwen_omni (server from the previous step must be running):
+uv run python scripts/caption_videos.py \
+  "<workspace>/<run-name>/dataset/videos/<one-staged-clip>" \
+  --output "<workspace>/<run-name>/dataset/preview-captions.json" \
+  --captioner-type qwen_omni
+  # Point at 3 staged clips under dataset/videos/ (a small subdir or 3 explicit files) — not the out-of-tree source.
+  # Optional: --vllm-url http://127.0.0.1:9000/v1   (if the server uses a non-default port)
+```
+
+Print the 3 captions **in full** to the user, then **STOP and wait** for their explicit verdict:
+
+> "Here are sample captions from the default prompt. Please review them — reply 'good' to caption the rest, or tell me what to change."
+
+**This is a hard gate. Do NOT caption the full set until the user explicitly approves the samples.** Do not auto-proceed, do not assume "looks fine," do not batch this with other questions. The user must either approve or give tuning instructions first — the whole point of the spot-check is to let them judge caption quality and content before paying for the full pass.
+
+If the user requests changes, introduce a custom `--instruction` (or switch backend), re-run the spot-check on the same 3 samples, show the new captions, and **stop for approval again**. Loop until the user approves. If a custom instruction still isn't converging after a few rounds, switch captioner backend or have the user supply a few manual captions as examples — but still don't proceed to the full set without their OK.
+
+#### 3b — Full pass
+
+Run on the **staged media dir** (`dataset/videos/` from Step 1, or `dataset/scenes/` from Step 2) with the **default prompt** (or the same `--instruction` only if one was explicitly agreed in 3a). Caption the staged in-tree media — not the original out-of-tree source path.
+
+**Qwen3-Omni (local — server must be running):**
+
+```bash
+cd packages/ltx-trainer
+uv run python scripts/caption_videos.py \
+  "<workspace>/<run-name>/dataset/videos" \
+  --output "<workspace>/<run-name>/dataset/dataset.json" \
+  --captioner-type qwen_omni
+```
+
+**Gemini Flash (cloud — runs anywhere, parallelisable):**
+
+```bash
+# Auth: GEMINI_API_KEY / GOOGLE_API_KEY env var, or gcloud / Vertex AI credentials.
+cd packages/ltx-trainer
+uv run python scripts/caption_videos.py \
+  "<workspace>/<run-name>/dataset/videos" \
+  --output "<workspace>/<run-name>/dataset/dataset.json" \
+  --captioner-type gemini_flash \
+  --num-workers 5
+```
+
+Output: JSON list of `{caption, media_path}` with paths **relative to the output file location**. The 3 spot-check captions can be merged in to avoid re-captioning them.
+
+#### Trigger word (handled at preprocessing, not in captions)
+
+For style/concept LoRAs the trigger word is **not** written into the captions here. It is prepended to every caption at preprocessing: pass `--lora-trigger "<word>"` to `process_dataset.py` in Phase 7, which forwards it to the caption-processing step (`process_captions.py`, where the prepend actually happens). That is the canonical mechanism — keep the captions describing what's actually on screen (via the default prompt), and let the trigger flag bind the concept to the token. Record the chosen trigger word in the plan so Phase 7 passes it through. Do not also bake the word into captions (it would double up).
+
+**The injection mechanism is a fixed implementation detail — never make it a user-facing question.** The *only* trigger-word thing to ask the user is the **word itself** (or whether they want a trigger word at all). Do **not** ask, mention, or present as an option *how* it gets injected (e.g. "inject into the caption vs via `process_dataset`") — it is always `--lora-trigger`, full stop. Surfacing the method as a choice creates unnecessary confusion.
+
+### Step 4 — Conditioning inputs (modes that need references or masks)
+
+Some modes need a per-sample conditioning input beyond the video/audio and caption:
+
+| Mode | Required extra input | Column |
+|------|----------------------|--------|
+| V2V IC-LoRA | reference video | `reference_video` (alias `ref_media_path`) |
+| AV2AV IC-LoRA | reference video + reference audio | `reference_video`, `reference_audio` |
+| A2A IC-LoRA | reference audio | `reference_audio` |
+| Video inpainting | per-frame video mask | `video_mask` |
+| Audio inpainting | audio mask | `audio_mask` |
+
+These inputs encode **the user's specific idea** for the LoRA (what the reference represents, which regions the mask covers). There is no universal recipe, so **do not invent or default to a particular method** (e.g. don't assume Canny edges, depth, pose, or some generic box/border mask). The agent must not pick the conditioning semantics for the user.
+
+Workflow when the chosen mode needs one of these and the dataset doesn't already provide it:
+
+1. **Check first** — if the user already supplied the column (and the files exist), use it as-is and move on.
+2. **Otherwise, ask the user to provide it**, explaining concretely what's needed: the column name, that it's one file per sample aligned to each clip, and that the *content/semantics are their call* (what the reference should depict, what the mask should cover). Make clear this reflects their specific use-case — you won't guess it.
+3. **Help generate only if the user asks.** If they say "can you generate the references/masks by doing X" (X = their described method), then help: write or run a small script for *their* approach, or use a repo tool if it fits. One such tool exists — `scripts/compute_reference.py` generates **Canny edge** reference videos — but only mention/use it if the user specifically wants Canny; never offer it as the default.
+4. **Hard gate:** do not proceed to preprocessing for a conditioning mode until the required column is present with real files. Surface clearly if it's missing.
+
+**Column-naming note (if references are generated):** `compute_reference.py` writes a `reference_video` field, which
+`process_dataset.py` detects automatically. Legacy datasets using `ref_media_path` also work.
+
+### Step 5 — Holdout split
+
+Reserve a subset of samples as a **held-out set** never seen during training. This is what Phase 9 (post-train validate) renders against to test true generalization rather than memorization.
+
+Decision tree:
+
+1. **User already supplied a held-out set** (separate file or directory they explicitly nominated): do not split. Copy/reference their file to `<workspace>/<run-name>/dataset/holdout.jsonl` and leave `dataset.json` as-is.
+2. **Small dataset** (judge qualitatively; tens of samples or fewer): holding samples out meaningfully reduces training capacity. Ask the user:
+   > "Dataset has <N> samples. Reserving a holdout meaningfully reduces what's available for training. Options: (a) reserve 1–2 for holdout, (b) skip holdout — post-train eval will only test in-distribution. Your call."
+3. **Otherwise:** auto-split. Reserve a small fraction (this skill's default: roughly 10% of samples, bounded so the holdout doesn't grow huge — a handful of held-out samples is usually enough). Use seed 42 for the split so it's reproducible. Surface the count and the picked IDs in the plan.
+
+After splitting, write `<workspace>/<run-name>/dataset/holdout.jsonl` (one JSON object per line with the same columns as `dataset.json`). **Remove the held-out entries from `dataset.json`** so they don't enter preprocessing or training.
+
+Always print:
+> "Held out <K> of <N> samples for post-train evaluation. Held-out IDs: <list>."
+
+If holdout is skipped, surface in the plan: *"Skipping holdout — dataset is too small. Post-train eval will only render in-distribution prompts; true generalization isn't testable for this run."*
+
+### Step 6 — Audit
+
+Before returning to the orchestrator, verify the metadata file has all required columns for the chosen mode. Print a one-line summary:
+
+> "Prepared <N> training samples (+ <K> held out) for <mode>. Columns: <list>. Saved to `<workspace>/<run-name>/dataset/dataset.json` (+ `holdout.jsonl`)."
+
+## Idempotency
+
+- If `dataset.json` already exists and all required columns are present: skip captioning. Confirm reuse with the user only if the file was supplied by them outside the workspace (per the orchestrator's file-safety invariant).
+- If captioning was partial (some entries missing `caption`), re-run captioning only on the missing entries by filtering the metadata file before passing to `caption_videos.py`.
+
+## Failure Modes
+
+- Qwen server won't start / OOMs on launch → use the default `--quantization fp8` (not `bf16`), lower `--gpu-memory-utilization`, or reduce `--max-model-len` on `serve_captioner.py`. If the GPU simply can't host a 30B model, switch to `gemini_flash`.
+- `caption_videos.py` can't connect (qwen_omni) → the vLLM server isn't running or `--vllm-url` is wrong. Start `serve_captioner.py` first and confirm the URL/port match.
+- Gemini rate-limit → reduce `--num-workers`, retry.
+- Scene splitter produces 0 scenes → the detector found no cuts, or `--filter-shorter-than` removed everything. Lower/remove `--filter-shorter-than`, or adjust the detector threshold. (`--min-scene-length` is an integer minimum-frames-per-scene for the detector, not a short-scene filter.)
+- IC-LoRA reference compute fails on some frames → script logs the failures; report counts to the user and ask whether to proceed with the remaining samples or stop.
+
+## Do Not
+
+- Do not move or rename the user's source files. The skill reads them in place; the workspace contains only **derived** artifacts.
+- Do not delete `scenes/` or partial captioning outputs without approval — they may be expensive to regenerate.

.claude/skills/train-model/phases/preprocess-dataset.md

@@ -0,0 +1,130 @@
+# Phases 6 (one-sample) & 7 (full) — Preprocess Dataset
+
+Procedure document for the `train-model` orchestrator. Read this file in full before acting on the preprocess phase.
+
+Goal: run `process_dataset.py` to produce VAE latents, audio latents, and text embeddings. Two modes:
+1. **One-sample** (Phase 6 sanity check) — preprocess a single sample to `<workspace>/<run-name>/overfit/.precomputed/`.
+2. **Full** (Phase 7) — preprocess the whole dataset to `<workspace>/<run-name>/dataset/.precomputed/`.
+
+The orchestrator's hard invariants apply (see `../SKILL.md`), especially: **never silently overwrite existing user data.**
+
+## Required Subdirectories by Mode
+
+Under `.precomputed/`:
+
+| Subdir | Required for |
+|--------|--------------|
+| `latents/` | Video-bearing modes only (T2V, I2V, video extend/inpaint/outpaint, V2V/AV2AV IC-LoRA, A2V, V2A). **Not** produced for audio-only modes. |
+| `conditions/` | Always (text embeddings) |
+| `audio_latents/` | Any mode with audio: video modes carrying audio, plus all audio-only modes (T2A, audio extend/suffix/inpaint, A2A IC-LoRA) |
+| `reference_latents/` | V2V IC-LoRA, AV2AV IC-LoRA |
+| `reference_audio_latents/` | A2A IC-LoRA, AV2AV IC-LoRA |
+| `video_masks/` | Video inpainting |
+| `audio_masks/` | Audio inpainting |
+
+**Audio-only modes** (T2A, audio extend/suffix, audio inpainting, A2A IC-LoRA) produce `audio_latents/` + `conditions/` (plus `audio_masks/` or `reference_audio_latents/` as applicable) and **no `latents/`**. Do not flag a missing `latents/` as incomplete for these modes.
+
+## Workflow — Full Preprocess
+
+### Step 1 — Verify existing `.precomputed/` (if present)
+
+If `<workspace>/<run-name>/dataset/.precomputed/` already exists:
+
+1. List subdirectories present. Confirm all required for the chosen mode are present.
+2. Load one sample per modality and check tensor shapes:
+   ```bash
+   uv run python -c "import torch; t = torch.load('<path>'); print(t.shape if hasattr(t, 'shape') else {k: v.shape for k, v in t.items()})"
+   ```
+3. Compare shapes against the target resolution from the plan.
+
+**On any mismatch or missing subdirectory: STOP. Do not run `process_dataset.py`.** Ask the user via `AskUserQuestion`:
+- Reuse the existing data at its current resolution (update plan + config accordingly).
+- Re-preprocess to a new directory (`<workspace>/<run-name>/dataset/.precomputed-v2/` etc.) — preserves the existing data.
+- Abort.
+
+**Never pass `--overwrite` without explicit user approval** for this exact action.
+
+### Step 2 — Invoke `process_dataset.py`
+
+```bash
+cd packages/ltx-trainer
+uv run python scripts/process_dataset.py \
+  "<workspace>/<run-name>/dataset/dataset.json" \
+  --resolution-buckets "<W>x<H>x<F>" \
+  --model-path "<absolute-model-path>" \
+  --text-encoder-path "<absolute-gemma-path>" \
+  --output-dir "<workspace>/<run-name>/dataset/.precomputed" \
+  --load-text-encoder-in-8bit         # on 32GB tier (low-VRAM config), per t2v_lora_low_vram.yaml
+```
+
+Add as needed:
+- `--skip-audio` — if mode doesn't use audio (T2V video-only variants).
+- `--audio-durations "<list>"` — for T2A from a captions-only file.
+- `--lora-trigger "<trigger>"` — for style/concept LoRAs.
+- `--reference-downscale-factor <N>` — for IC-LoRA modes if downscaled references are desired.
+- `--video-column`, `--caption-column` — only if the metadata file uses non-standard column names.
+
+**Do not pass `--overwrite`** unless re-preprocessing was explicitly approved in Step 1.
+
+### Step 3 — Audit output
+
+After completion, verify:
+
+```bash
+ls "<workspace>/<run-name>/dataset/.precomputed/"
+# Expected subdirs per the mode table above.
+
+# Count files in each:
+for d in latents conditions audio_latents reference_latents video_masks audio_masks; do
+  if [ -d "<workspace>/<run-name>/dataset/.precomputed/$d" ]; then
+    echo "$d: $(ls "<workspace>/<run-name>/dataset/.precomputed/$d" | wc -l)"
+  fi
+done
+```
+
+Counts in each required subdir should equal the dataset sample count.
+
+**Reconcile counts — do this for every run, any dataset size.** Compare the `latents/` (and `audio_latents/`) count against the caption/sample count. If fewer latents were produced, `process_dataset.py` **silently skipped** clips — most commonly because they were **shorter than the target frame bucket** (it logs each skip). When counts don't match:
+
+1. Identify which clips were dropped (grep the preprocess log for skip/"fewer frames" lines, or diff the produced `.pt` stems against the metadata).
+2. **Surface it to the user** with the count and the specific clips — never silently proceed on a shrunk dataset.
+3. Offer options: re-preprocess at a **smaller frame bucket** the clips support, add a **second (shorter) bucket** to keep the short clips (multi-bucket requires `batch_size: 1`), or accept the loss. Let the user decide.
+
+**Audio gate (hard stop for audio runs).** For any run with an audio modality (joint audio+video, A2V, V2A, T2A, audio-only modes), verify `audio_latents/` is **present and non-empty** with one `.pt` per sample. `process_dataset.py` **swallows audio-decode errors and continues** — it logs "0 videos with audio" and produces empty `audio_latents/` rather than failing. If an audio run produced no audio latents, **stop** — do not proceed to training (it would silently train audio-free). The usual cause is a broken audio decode path (e.g. missing/incompatible `torchcodec`); confirm `uv run python -c "import torchaudio; torchaudio.load('<a clip>')"` works (see `references/troubleshooting.md`), fix it, then re-preprocess with `--overwrite`.
+
+## Workflow — One-Sample (Phase 6)
+
+Same as full preprocess, but operate on a single-sample metadata file:
+
+1. Pick the first sample from `<workspace>/<run-name>/dataset/dataset.json` and write a one-sample metadata file **inside the dataset dir** — e.g. `<workspace>/<run-name>/dataset/_one_sample.json` — copying the entry **verbatim, keeping its relative `media_path`**. `process_dataset.py` resolves media paths relative to the metadata file's own directory, so the one-sample file must sit beside the real media (i.e. in `dataset/`, the same dir as `dataset.json`). **Do not** place it in `overfit/` and **do not** rewrite the path to an absolute one — an absolute path produces mirrored nested output dirs (`.precomputed/latents/absolute/path/.../x.pt`) instead of a clean `latents/x.pt`.
+2. Run `process_dataset.py` on that file with `--output-dir "<workspace>/<run-name>/overfit/.precomputed"` (output still goes to `overfit/`, only the metadata lives in `dataset/`).
+3. Use the **same `--resolution-buckets`** as the planned full run. Critical: a small-shape sanity check is misleading because resolution is the dominant memory factor.
+4. Clean up the temporary `dataset/_one_sample.json` afterward (it's scratch; don't leave it in the dataset dir).
+
+## Decode-and-Verify (optional debug aid)
+
+If the user reports validation samples look wrong or training diverges, decode one preprocessed sample back to media:
+
+`decode_latents.py` takes the **latents directory** and an **output directory** as positional arguments (it decodes the whole directory, not a single `.pt` file). Add `--with-audio` and `--audio-latents-dir` if the dataset has audio.
+
+```bash
+cd packages/ltx-trainer
+uv run python scripts/decode_latents.py \
+  "<workspace>/<run-name>/dataset/.precomputed/latents" \
+  "<workspace>/<run-name>/dataset/.precomputed/decoded_check" \
+  --model-path "<absolute-model-path>"
+```
+
+If decoded output is garbled, preprocessing itself is suspect (wrong model, wrong VAE).
+
+## Failure Modes
+
+- **"shape mismatch" on resume:** Step 1's verification check. Ask user before any mutation.
+- **`frames % 8 != 1`** error from process_dataset.py: the requested frame count is invalid; correct in the plan and re-launch.
+- **VRAM OOM during preprocessing:** add `--load-text-encoder-in-8bit`. If still OOM, reduce `--batch-size`.
+- **Disk full:** preprocessed latents can be large (especially audio). Surface to user with a `du -sh` summary of `.precomputed/`.
+
+## Do Not
+
+- Do not delete or overwrite existing `.precomputed/` data without explicit user approval for that exact action.
+- Do not preprocess at a smaller resolution to "save time" — the sanity check exists specifically to validate the planned resolution.

.claude/skills/train-model/references/config-patching.md

@@ -0,0 +1,73 @@
+# Config Patching
+
+How to safely produce `<workspace>/<run-name>/config.yaml` from an example in `packages/ltx-trainer/configs/`. The trainer's config schema is Pydantic with `extra="forbid"` — unknown fields are rejected. Full field reference: [`packages/ltx-trainer/docs/configuration-reference.md`](../../../../packages/ltx-trainer/docs/configuration-reference.md).
+
+## Workflow
+
+1. Copy the example config matching the selected mode (see `mode-selector.md`) to `<workspace>/<run-name>/config.yaml`.
+2. Patch fields as described below. Preserve YAML comments where possible — they help the user audit the run later.
+3. **Never** edit the example config in `packages/ltx-trainer/configs/`. That's the user's reference library.
+
+## Required Patches (every run)
+
+| Field | Value |
+|-------|-------|
+| `model.model_path` | Absolute path to local `.safetensors` (from probe or user). |
+| `model.text_encoder_path` | Absolute path to local Gemma directory (from probe or user). |
+| `data.preprocessed_data_root` | `<workspace>/<run-name>/dataset/.precomputed` (absolute). |
+| `output_dir` | `<workspace>/<run-name>/outputs` (absolute). |
+
+## Hardware-Driven Patches
+
+Apply per the matched VRAM tier in `references/hardware-profiles.md`. After autotune (Phase 6), patch the winning trial's deltas in.
+
+## Schema Constraints (validate before launch)
+
+These will cause Pydantic errors or runtime failures; check before invoking the trainer.
+
+- **Frame count:** `validation.video_dims[2]` must satisfy `frames % 8 == 1` (1, 9, 17, 25, 33, 41, 49, 57, 65, 73, 81, 89, ...).
+- **Resolution:** `validation.video_dims[0]` and `[1]` must be divisible by 32.
+- **Multi-bucket training:** if dataset uses multiple resolution buckets, set `optimization.batch_size: 1`.
+- **At least one generated modality:** `training_strategy` must have at least one of `video.is_generated` or `audio.is_generated` set to `true`.
+- **Audio condition restrictions:** the audio modality cannot use `first_frame` or `spatial_crop` conditions.
+- **Strategy name:** prefer `training_strategy.name: "flexible"`. `text_to_video` and `video_to_video` still work but emit deprecation warnings.
+
+## LoRA Patches
+
+For style/concept LoRAs:
+
+- `lora.rank` and `lora.alpha`: set from the matched VRAM tier and use case. **32GB tier** pins rank 16 per `t2v_lora_low_vram.yaml`; **80GB+ tier** uses rank 32 per `t2v_lora.yaml`. Keep `alpha == rank`. See `mode-selector.md` for use-case-driven rank guidance.
+- `lora.target_modules`: short patterns like `"to_k"`, `"to_q"`, `"to_v"`, `"to_out.0"` match all attention modules (video + audio + cross-modal). Add `"ff.net.0.proj"`, `"ff.net.2"` only if user explicitly wants higher capacity.
+- **Audio-only LoRA targets** (T2A, audio inpainting): use `"audio_attn1.to_*"`, `"audio_attn2.to_*"` patterns to avoid touching video weights. See `configs/t2a_lora.yaml` for the exact list.
+
+## Validation Sample Prompts
+
+The example configs ship with placeholder validation prompts. Validation condition fields are documented in
+[`configuration-reference.md#validation-condition-types`](../../../../packages/ltx-trainer/docs/configuration-reference.md#validation-condition-types).
+For style/concept LoRAs:
+- Replace at least one `validation.samples[].prompt` with a prompt that uses the user's trigger word or describes the target concept. Tells the user something useful at the first validation interval.
+- Keep `validation.video_dims` consistent with the training resolution to make samples comparable.
+- **Describe the audio, for any run with a generated audio modality** (joint audio+video, T2A, V2A, etc.). The validation prompts must describe the audio the **same way the training captions do** — if the training captions transcribe speech or characterise sound (e.g. *"he says: ‘…’"*, *"calm spoken voice, quiet room tone"*, *"upbeat acoustic guitar"*), the validation prompts must include comparable audio direction. A prompt with no audio description gives the model no guidance for the audio branch and the generated audio comes out poor. This is not speech-specific — any audio (music, ambience, foley) needs describing. Mirror the structure/level of audio detail found in the dataset captions (inspect a few before writing the prompts).
+
+## W&B Patches
+
+- If the W&B credential check passes (`uv run python -c "import wandb; print(bool(wandb.Api().api_key))"` → `True`): `wandb.enabled: true`, `wandb.project` = `ltx2-<mode>`, `wandb.tags` includes the mode. (Do not use `wandb status` — it falsely reports `api_key: null` when logged in via netrc.)
+- If `False`: `wandb.enabled: false`. Surface in plan: *"Not logged in to W&B — run `wandb login` before training to enable tracking."* If the check errored/was ambiguous, ask the user rather than assuming off.
+
+## Output Dir Behaviour
+
+The trainer resumes optimizer/scheduler/step state **only when `model.load_checkpoint` is set** to a checkpoint file; it then looks for a matching `training_state_step_*.pt` next to that file. It does **not** auto-detect prior checkpoints in `output_dir/checkpoints/`. To resume, patch `model.load_checkpoint` to the latest checkpoint. To load weights but skip state restore: `checkpoints.no_resume: true`.
+
+For the orchestrator's resume flow: when resuming an interrupted run, patch `model.load_checkpoint` to the latest checkpoint under `output_dir/checkpoints/`. Leaving it unset starts a fresh run from step 0 even if checkpoints exist on disk.
+
+## Self-Check Before Launch
+
+Before any `python scripts/train.py` invocation:
+
+1. All `model_path`, `text_encoder_path`, `preprocessed_data_root` exist on disk.
+2. Frame and resolution constraints satisfied (see above).
+3. Generated modalities have matching latents directories under `.precomputed/`.
+4. For modes with `reference` condition: `reference_latents/` (and/or `reference_audio_latents/`) exists.
+5. For modes with `mask` condition: `video_masks/` (and/or `audio_masks/`) exists.
+
+A failed check at this point is much cheaper than a failed training start.

.claude/skills/train-model/references/hardware-profiles.md

@@ -0,0 +1,98 @@
+# VRAM Tiers
+
+Map probed GPU(s) to a starting training config. The autotune sweep in Phase 6 then empirically improves on this baseline. Use these **tier names** in `plan.md` and user-facing chat — not letter codes.
+
+Source of truth: the two configs shipped in the trainer repo —
+`packages/ltx-trainer/configs/t2v_lora.yaml` (standard) and
+`packages/ltx-trainer/configs/t2v_lora_low_vram.yaml` (low VRAM).
+Per `packages/ltx-trainer/docs/quick-start.md`, the trainer documents
+**80GB recommended** and **32GB minimum**. Anything below 32GB is
+unsupported by the project.
+
+## Probe
+
+```bash
+nvidia-smi --query-gpu=name,memory.total --format=csv,noheader
+```
+
+Pick the **smallest** VRAM tier across the visible GPUs. Multi-GPU only adds throughput at the same per-GPU memory budget — it doesn't relax per-GPU limits.
+
+## Minimum Gate
+
+If per-GPU VRAM is **< 32 GB**, stop the run. Surface to the user:
+
+> "This GPU has <N>GB VRAM. The LTX-2 trainer requires a minimum of 32GB (see `packages/ltx-trainer/docs/quick-start.md`). Training is unlikely to fit even with maximum memory savings, and we don't ship a tested config below 32GB. Options: (a) abort, (b) try anyway with the low-VRAM config and accept it may OOM — purely at your own risk."
+
+Do not invent a sub-32GB tier. The trainer team doesn't ship one.
+
+## 32GB tier — low-VRAM config
+
+**VRAM range:** 32 GB per GPU (trainer minimum).
+
+**Typical GPUs:** RTX 5090, V100 32GB.
+
+Start from `packages/ltx-trainer/configs/t2v_lora_low_vram.yaml` verbatim. Key choices already in that file (do not re-specify in `<workspace>/<run-name>/config.yaml` — copy the file and patch only the paths from `references/config-patching.md`):
+
+- `optimizer_type: "adamw8bit"`
+- `enable_gradient_checkpointing: true`
+- `batch_size: 1`, `gradient_accumulation_steps: 1`
+- `quantization: "int8-quanto"`
+- `load_text_encoder_in_8bit: true`
+- `offload_optimizer_during_validation: true`
+- `lora.rank: 16`, `lora.alpha: 16`
+
+Autotune (Phase 6) will sweep `quantization` off, `optimizer_type` → adamw, and `batch_size` up — but at 32GB the sweep often hits OOM on trial 2 or 3. That's fine; the conservative baseline still works.
+
+## 80GB+ tier — standard config
+
+**VRAM range:** 80 GB per GPU and above (trainer recommended).
+
+**Typical GPUs:** A100 80GB, H100 80GB, H200, B200.
+
+Start from `packages/ltx-trainer/configs/t2v_lora.yaml` verbatim. Key choices already in that file:
+
+- `optimizer_type: "adamw"`
+- `enable_gradient_checkpointing: true` (autotune may turn it off if headroom allows)
+- `batch_size: 1`, `gradient_accumulation_steps: 1`
+- `quantization: null`
+- `load_text_encoder_in_8bit: false`
+- `lora.rank: 32`, `lora.alpha: 32`
+
+On the 80GB+ tier, the autotune baseline already equals this config (adamw, no quantization), so the quantization/optimizer trials are no-ops; the only real lever is gradient checkpointing off — but for the 22B model that **usually OOMs even with tens of GB of apparent headroom**, so treat a win there as unlikely. The trainer reports its own step-time and peak-VRAM at the end of each run — use those rather than an external timer.
+
+For ≥140GB GPUs (H200, B200), the same 80GB+ tier baseline applies. FA3/FA4 attention backends are viable on Hopper/Blackwell and can speed up training, but they're optional — the trainer's defaults work on PyTorch SDPA without extra setup.
+
+## 40–60GB tier — mid-range (autotune from low-VRAM)
+
+**VRAM range:** 40–60 GB per GPU. The trainer doesn't ship a tested config for this range.
+
+**Typical GPUs:** A40, A6000 48GB, L40, RTX 6000 Ada.
+
+Start from the **32GB tier** (low-VRAM config) and let autotune relax `quantization`, `optimizer_type`, and `batch_size` based on actual headroom. Don't pre-bake intermediate YAML values that haven't been measured. Surface this as **40–60GB tier** in the plan.
+
+## Multi-GPU
+
+If `nvidia-smi` reports N ≥ 2 GPUs of the same model:
+
+- Launch with `uv run accelerate launch scripts/train.py <config>`.
+- Use `packages/ltx-trainer/configs/accelerate/fsdp.yaml` for full fine-tune.
+- DDP (default `accelerate launch` without a config file) is fine for LoRA.
+- Effective batch = `batch_size * gradient_accumulation_steps * num_gpus`. Reduce `gradient_accumulation_steps` proportionally to keep the effective batch consistent with the plan.
+
+## Full Fine-Tune
+
+If the user chose full fine-tune (`model.training_mode: "full"`):
+
+- Require multi-GPU + FSDP on 80GB+ tier GPUs. Otherwise warn in the plan that single-GPU full FT is unlikely to fit and propose LoRA instead.
+- Set `acceleration.offload_optimizer_during_validation: true` always (optimizer state is huge under full FT).
+
+## Model Path Constraints
+
+- `model.model_path`: local `.safetensors` only. No URLs.
+- `model.text_encoder_path`: local Gemma model directory. No URLs.
+
+If probe didn't find these in conventional locations (`/models/`, `~/models/`, `$LTX_MODELS_DIR`), ask the user in Phase 3 (or offer to download per `references/onboarding.md`).
+
+## Notes on Loss-of-Generality
+
+The two anchor tiers (32GB and 80GB+) correspond directly to the two configs the trainer ships. The autotune sweep is the empirical layer — if a particular GPU consistently lands on a different stable config, **update the relevant trainer config first**, not this file. This skill follows the trainer's choices, not the other way around.

.claude/skills/train-model/references/mode-selector.md

@@ -0,0 +1,74 @@
+# Mode Selector
+
+Map the user's stated intent to a `flexible`-strategy configuration. All modes are supported via a single strategy (`training_strategy.name: "flexible"`); the difference is which modality is generated and which `conditions` are attached.
+
+> Full reference: [`packages/ltx-trainer/docs/training-modes.md`](../../../../packages/ltx-trainer/docs/training-modes.md). This file is the **lookup table** for translating user intent.
+
+## Decision Table
+
+| User says (roughly)... | Mode | Example config | Modalities | Conditions |
+|------------------------|------|----------------|------------|------------|
+| "generate videos from text", "T2V LoRA" | T2V | `configs/t2v_lora.yaml` | video gen, audio gen | none |
+| "generate videos from a starting image", "I2V" | I2V | `configs/i2v_lora.yaml` | video gen, audio gen | `first_frame` (video) |
+| **plain concept/style LoRA** ("train a LoRA on X", no specific task) | **I2V by default** (see note) | `configs/i2v_lora.yaml` | video gen, audio gen | `first_frame` (video), `probability: 0.5` |
+| "extend a video forward in time" | Video extension (prefix) | `configs/video_extend_lora.yaml` | video gen, audio gen | `prefix` (video) |
+| "extend a video backward in time" | Video extension (suffix) | `configs/video_suffix_lora.yaml` | video gen, audio gen | `suffix` (video) |
+| "fill in masked regions of a video" | Video inpainting | `configs/video_inpainting_lora.yaml` | video gen | `mask` (video) |
+| "expand a video beyond its borders" | Video outpainting | `configs/video_outpainting_lora.yaml` | video gen | `spatial_crop` (video) |
+| "style transfer from reference video", "IC-LoRA", "depth/pose/canny control" | V2V IC-LoRA | `configs/v2v_ic_lora.yaml` | video gen | `reference` (video) |
+| "generate video to match an audio track" | A2V | `configs/a2v_lora.yaml` | video gen, audio frozen | none (audio `is_generated: false`) |
+| "add sound effects to silent video", "foley", "V2A" | V2A | `configs/v2a_lora.yaml` | video frozen, audio gen | none (video `is_generated: false`) |
+| "generate audio from text", "T2A" | T2A | `configs/t2a_lora.yaml` | audio gen | none |
+| "extend audio forward / backward" | Audio extension | `configs/audio_extend_lora.yaml`, `configs/audio_suffix_lora.yaml` | audio gen | `prefix` / `suffix` (audio) |
+| "fill in masked regions of audio" | Audio inpainting | `configs/audio_inpainting_lora.yaml` | audio gen | `mask` (audio) |
+| "audio style transfer from reference", "A2A IC-LoRA" | A2A IC-LoRA | `configs/a2a_ic_lora.yaml` | audio gen | `reference` (audio) |
+| "joint video+audio reference control" | AV2AV IC-LoRA | `configs/av2av_ic_lora.yaml` | video gen, audio gen | `reference` (both) |
+| Any of the above with full fine-tune | Full FT variant | as above, set `model.training_mode: "full"` | (mode-specific) | (mode-specific) |
+
+### Why I2V is the default for a plain concept/style LoRA
+
+A "train a LoRA on X" request isn't tied to one inference mode: **LoRA weights are pipeline-agnostic** — the same checkpoint loads in both T2V and I2V inference (both use `TI2VidOneStagePipeline`/`TwoStages`). The `i2v_lora` config trains `first_frame` with **`probability: 0.5`**, so the model learns both first-frame-conditioned (I2V) and unconditioned (T2V) generation in one run, and the first frame comes from each training clip automatically (no extra data prep). That makes I2V a versatile **superset** — usable for both at inference at no extra cost — which is why it's the default for a plain LoRA. Ask the user how they'll use it (text-only / from an image / both) and only drop to `t2v_lora` if they're sure it's text-only. (See the orchestrator `SKILL.md` Phase 1.)
+
+## Disambiguation Questions
+
+When the user's first answer is ambiguous, ask **one** follow-up:
+
+- "extend a video" → forward or backward in time?
+- "control with a reference" → video reference (depth/pose/canny/etc.) or audio reference?
+- "fill in regions" → video regions (masked frames) or audio regions (masked time)?
+- "T2V" → joint video+audio (default) or video-only?
+- LoRA or full fine-tune? Default to LoRA unless the user has multi-GPU + clear reason for full.
+
+## Combining Modes
+
+The flexible strategy allows stacking conditions. Common combinations:
+
+- **I2V + V2A** (start frame + generate audio for the resulting video) — not directly expressible; would need two passes.
+- **Video extension + audio extension** — both modalities generate, both have `prefix` condition. Express in one config.
+- **IC-LoRA + I2V** — `first_frame` + `reference` conditions on the video modality.
+
+If the user asks for a combination not listed, check `packages/ltx-trainer/src/ltx_trainer/training_strategies/flexible.py` for which conditions can co-exist on a modality. Audio modality cannot use `first_frame` or `spatial_crop`.
+
+## When the Intent Doesn't Map
+
+If after disambiguation there is no entry in the table and no combination of `flexible` conditions covers the user's request, go to the **Escape Hatch** section in the orchestrator `SKILL.md`. Do not silently pick the closest mode.
+
+## LoRA Rank by Use Case
+
+Once the mode is picked, choose `lora.rank` (and matching `lora.alpha`) based on what the LoRA is supposed to capture. These are starting points; autotune doesn't sweep rank because it's a quality knob, not a step-time one.
+
+| Use case | Suggested rank | Notes |
+|----------|----------------|-------|
+| Single character, single object, single style | 32–64 | Default for most concept LoRAs. Start at 32; bump to 64 if validation samples underfit. |
+| Multi-character world, dense series, complex multi-concept | 96–128 | More capacity for distinguishing several concepts inside one LoRA. |
+| Camera move, motion, transition (i.e. behavioural, not visual) | 8–16 | Motion is a thin signal — high ranks just memorise frame content. |
+| IC-LoRA control (V2V depth/pose/Canny/etc., A2A audio reference) | 16–32 | Start at 16 for structural control (depth, pose, edges); 24–32 if the reference carries richer style/texture. Video IC-LoRA often lands lower than concept LoRAs. |
+| LTX-2 trainer's default if unsure | 32 | Safe baseline. |
+
+On the **32GB tier** (`hardware-profiles.md`), the low-VRAM config already pins `lora.rank: 16`. If the user picks a higher-rank use case on 32GB, surface the trade-off in the plan but let them decide — the autotune sweep doesn't touch rank, so a too-high rank will simply OOM at training time.
+
+Keep `alpha == rank` unless the user has a specific reason otherwise (effective scaling = `alpha / rank`).
+
+## Starting Config
+
+Copy the matching example config — the exact filename from the **Example config** column of the decision table above (e.g. T2V → `configs/t2v_lora.yaml`, I2V → `configs/i2v_lora.yaml`, A2A IC-LoRA → `configs/a2a_ic_lora.yaml`) — into `<workspace>/<run-name>/config.yaml` as the starting point. Then patch per `references/config-patching.md`.

.claude/skills/train-model/references/onboarding.md

@@ -0,0 +1,115 @@
+# First-Run Onboarding
+
+What the orchestrator's Phase 2 probe checks for, what to do when something is missing, and what the skill is allowed to set up automatically (with explicit user approval).
+
+## Prerequisites Checked in Phase 2
+
+| Prerequisite | How to detect | If missing |
+|--------------|---------------|------------|
+| CUDA GPU visible | `nvidia-smi` returns ≥1 GPU | Stop. Training requires CUDA — point the user at non-LTX-2 docs. |
+| Linux | `uname -s` returns `Linux` | Stop. Trainer uses Triton (Linux-only). |
+| `uv` installed | `command -v uv` | Offer to install (see "Auto-setup" below). |
+| Workspace synced | `[ -f uv.lock ] && uv pip list \| grep -q ltx-trainer` | Offer to run `uv sync` from repo root. |
+| LTX-2 model weights | Search `/models/`, `~/models/`, `$LTX_MODELS_DIR` for a `.safetensors` matching `*ltx*2*` | Offer to download (see "Model downloads"). |
+| Gemma text encoder dir | Search same locations for a directory containing Gemma config | Offer to download. |
+| Captioner backend | Gemini auth (`GEMINI_API_KEY`/`GOOGLE_API_KEY` or gcloud/Vertex), OR a ≥40 GiB GPU to host the Qwen3-Omni-30B vLLM server (FP8), OR captions already in the dataset | See "Captioner graceful degradation" below. **Check the HF cache for an already-downloaded Qwen model before assuming a download is needed** (see note below the table). |
+| W&B login (optional) | `uv run python -c "import wandb; print(bool(wandb.Api().api_key))"` → `True` means logged in. Uses wandb's own credential resolution (env/netrc/settings). **Don't** use `wandb status` (reports `api_key: null` even when logged in). | Not a blocker. If `False`: disabled in config + flagged in plan. If the check errors/ambiguous: ask the user, don't assume off. |
+| Disk space | `df -h $WORKSPACE` | Surface available space alongside what a run consumes (preprocessed latents, several-GB checkpoints, validation samples). Flag concerns to the user; don't enforce a hard threshold. |
+
+Surface findings as a compact table in chat. For each missing item, present the user with a concrete next step (download command, install command, or "skip this — here's the consequence").
+
+**Check the HF cache before declaring the local captioner unavailable.** The Qwen3-Omni model is served from the HuggingFace cache, not `~/models/`. Before concluding it must be downloaded (or ruling it out on free-disk grounds), check whether it's already cached:
+
+```bash
+ls -d "${HF_HOME:-$HOME/.cache/huggingface}"/hub/models--Qwen--Qwen3-Omni* 2>/dev/null \
+  && du -sh "${HF_HOME:-$HOME/.cache/huggingface}"/hub/models--Qwen--Qwen3-Omni* 2>/dev/null
+```
+
+If it's cached (~60 GiB, all shards present), no download is needed — don't rule out the local captioner because of low free disk on the *home* partition; the weights already exist. Only the GPU-VRAM constraint (≥40 GiB for FP8) then applies.
+
+## Auto-Setup (with explicit user approval)
+
+The skill may, only after the user explicitly says yes, do these setup actions. Each action is a single discrete question.
+
+### Run `uv sync`
+
+```bash
+cd <repo-root>
+uv sync
+```
+
+Ask: *"Repo not synced. Run `uv sync` now? It will download the project's Python dependencies."*
+
+### Install `uv`
+
+Ask: *"`uv` not installed. Install via the official one-liner now?"*
+
+```bash
+curl -LsSf https://astral.sh/uv/install.sh | sh
+```
+
+After install, ask the user to restart their shell or `source ~/.bashrc` before continuing.
+
+### Model downloads
+
+Use `huggingface-cli` (comes with `huggingface-hub`, transitively pulled by `uv sync`). Default destination: `$LTX_MODELS_DIR` if set, else `~/models/` (create if missing). Surface destination in the prompt — never download into the repo or into the workspace.
+
+**LTX-2 base model:**
+
+```bash
+huggingface-cli download Lightricks/LTX-2.3 \
+  ltx-2.3-22b-dev.safetensors \
+  --local-dir ~/models/ltx-2.3
+```
+
+Public reference: <https://huggingface.co/Lightricks/LTX-2.3>.
+
+**Gemma text encoder:**
+
+```bash
+huggingface-cli download google/gemma-3-12b-it-qat-q4_0-unquantized \
+  --local-dir ~/models/gemma-3-12b
+```
+
+**Qwen3-Omni captioner (only if the GPU can host it):**
+
+The local captioner is Qwen3-Omni-30B-A3B-Thinking served by a vLLM server (`scripts/serve_captioner.py`), which downloads the model (~65 GiB) on first launch via `uvx vllm` — there's no separate `huggingface-cli` step. Default **FP8** (~31 GiB weights) fits on a **40 GiB** GPU; **bf16** (~60 GiB) needs **≥66 GiB free VRAM**. On a GPU below ~40 GiB (typical consumer 24/32 GB cards), don't use it — use Gemini instead (see "Captioner graceful degradation").
+
+The base model + text encoder are large (multi-GB) downloads. Ask the user before each one — `huggingface-cli` reports the actual size at the start of the transfer. Do **not** batch them into a single "yes/no"; the user may want only what's missing.
+
+### Hugging Face login (if any downloads fail with 401)
+
+Ask: *"Hugging Face download requires login (some Lightricks models are gated). Run `huggingface-cli login` now? You'll need a token from <https://huggingface.co/settings/tokens>."*
+
+## Captioner Graceful Degradation
+
+The captioner is the trickiest prerequisite. The local backend (`qwen_omni`) is now a **30B model served by a vLLM server** — ~65 GiB download; FP8 fits on a 40 GiB GPU, bf16 needs ≥66 GiB. Typical consumer cards (24/32 GB) can't host it, so for most users **prefer Gemini**.
+
+Decision tree:
+
+1. **User already has captions in their dataset metadata** → skip the captioner entirely (Step 3 of `prepare-dataset` skips when captions are present).
+2. **GPU ≥40 GiB (FP8) / ≥66 GiB (bf16)** → `qwen_omni` is viable: launch `scripts/serve_captioner.py` first, then caption. Gemini is still fine here too.
+3. **GPU below ~40 GiB (the common consumer case)** → `qwen_omni` is not an option. Recommend **`gemini_flash`** and tell the user they need Gemini auth: a `GEMINI_API_KEY`/`GOOGLE_API_KEY` (get one at <https://aistudio.google.com/apikey>) **or** working gcloud/Vertex AI credentials.
+4. **No Gemini auth and can't run Qwen3** → the only remaining path is bringing their own captions: add a `caption` column to the dataset metadata, then re-invoke the skill.
+
+Wait for the user's choice. Don't pick one automatically — but make the hardware reality explicit so they don't try to run the server on a card that can't host it.
+
+## What Auto-Setup Does NOT Touch
+
+- Does not modify the user's shell rc files except by explicit instruction (e.g., "add `export LTX_MODELS_DIR=...` to your `.bashrc`?" with the user agreeing).
+- Does not modify `~/.gitconfig`, `~/.ssh/`, or any auth-related files.
+- Does not install GPU drivers, CUDA, or system packages.
+- Does not delete or move existing model files. If a model is found but at an unexpected path, surface and let the user decide.
+- Does not download into the repo or workspace. Models live at `~/models/` (or `$LTX_MODELS_DIR`).
+
+## Configuration Inheritance
+
+After downloads, the skill records the resolved paths into the plan's Assumptions section and into `config.yaml`:
+
+```yaml
+model:
+  model_path: "/home/<user>/models/ltx-2.3/ltx-2.3-22b-dev.safetensors"
+  text_encoder_path: "/home/<user>/models/gemma-3-12b"
+```
+
+Suggest (don't enforce) setting `LTX_MODELS_DIR=~/models` in their shell rc for future runs.

.claude/skills/train-model/references/plan-template.md

@@ -0,0 +1,107 @@
+# Plan Template
+
+Write `<workspace>/<run-name>/plan.md` following this template. The plan is the user's contract with the agent — it's the gate before any heavy work runs.
+
+Scale each section to its relevance. Skip subsections that don't apply (e.g., no captioning if data is already captioned), but never collapse "Assumptions" or "Cost/time estimate".
+
+```markdown
+# Training Plan — <run-name>
+
+## Goal
+
+<One paragraph restating the user's intent in their own terms.>
+
+## Mode
+
+**<Mode name>** — <one-line rationale linking user intent to mode>.
+
+- Config base: the concrete example config selected by the mode (e.g. `packages/ltx-trainer/configs/t2v_lora.yaml`) — use the actual filename, not a placeholder
+- Conditions: <list, or "none">
+- Training mode: <`lora` | `full`>
+
+## Dataset
+
+- Source: `<absolute path>` (<N samples>)
+- Captions: <`already present` | `will be generated with <backend>`>
+- Audio: <`present` | `absent — using --skip-audio` | `to be paired with --audio-durations`>
+- IC-LoRA references: <`present` | `to be generated via compute_reference.py` | `n/a`>
+
+## Preprocessing
+
+- Target resolution buckets: `<W>x<H>x<F>` (frames satisfy `frames % 8 == 1`; W,H divisible by 32)
+- Estimated time: ~<duration> on detected hardware
+- Output: `<workspace>/<run-name>/dataset/.precomputed/`
+
+## Training Config
+
+| Field | Value |
+|-------|-------|
+| Optimizer | <adamw / adamw8bit> |
+| Mixed precision | <bf16 / fp16> |
+| Quantization | <null / int8-quanto / ...> |
+| Gradient checkpointing | <on / off> |
+| Batch size | <N> |
+| Gradient accumulation | <N> (effective batch = <N>) |
+| Steps | <N> |
+| Learning rate | <value> |
+| LoRA rank / alpha | <N / N> (or "full FT") |
+| LoRA target modules | <list> (or "n/a") |
+| LoRA trigger word | <word> (or "n/a") |
+| Validation interval | every <N> steps |
+| Checkpoint interval | every <N> steps |
+
+## Hardware
+
+- GPU(s): <name> x <count>, <VRAM>GB per GPU
+- Launch: <`python scripts/train.py` (single) | `accelerate launch` (multi)>
+- VRAM tier: <32GB tier | 40–60GB tier | 80GB+ tier> — <low-VRAM config (`t2v_lora_low_vram.yaml`) | standard config (`t2v_lora.yaml`) | mid-range, autotuned from low-VRAM>
+
+## Sanity Check + Autotune
+
+*In plain terms: before committing to the full run, I do a quick dry run on a single clip at your target resolution to catch out-of-memory or config errors in ~2 minutes (rather than failing hours in), then try a few config variants to pick the fastest one that fits your GPU.*
+
+Mechanics:
+- 1 sample, full target resolution, 50 steps + 1 validation pass.
+- Autotune sweep: up to 5 trials varying quantization / optimizer / batch size.
+- Stops at first OOM or no-improvement.
+
+## Monitoring
+
+<One of:>
+- W&B: enabled, project `<name>`, entity `<entity-or-default>`. URL will be surfaced once training starts.
+- W&B: **not logged in** — run `wandb login` before training to enable tracking. Otherwise training proceeds without remote logging.
+
+## Outputs
+
+- Training config: `<workspace>/<run-name>/config.yaml`
+- Checkpoints: `<workspace>/<run-name>/outputs/checkpoints/`
+- Validation samples: `<workspace>/<run-name>/outputs/samples/`
+- Autotune log: `<workspace>/<run-name>/autotune.log`
+
+## Assumptions
+
+Defaults the agent chose silently. Override any by replying with the new value.
+
+- <list every non-trivial assumed value: precision, scheduler type, seed, validation prompts, checkpoint retention, etc.>
+
+## Cost / Time Estimate
+
+Give only estimates you can ground; label anything not yet measured as rough. Do **not** state a confident training duration before the sanity check has measured a real step time — say "training duration TBD until the sanity check measures step time" and fill it in afterward (per Hard Invariant #5: no fabricated predictions).
+
+- Captioning: ~<duration> (rough)
+- Preprocessing: ~<duration> (rough)
+- Sanity check + autotune: ~<duration> (rough)
+- Full training: **measured after sanity check** — then `<measured step-time> × <steps>`
+- **Total wall-clock estimate:** rough until step time is measured; refine after the sanity check.
+
+## Approve to Proceed
+
+Reply "approve" (or with edits) to start. No captioning, preprocessing, autotune, or training will run before approval.
+```
+
+## Notes on Writing the Plan
+
+- Show numbers, not adjectives. "~3 hours" beats "fairly long."
+- Surface every assumption that, if wrong, would cost the user time. Better to over-list than under-list — the user can skim.
+- If a section reveals you need to ask another question, **stop and ask** before finalizing the plan. The plan is the last gate, not the first.
+- If the user's hardware can't reasonably support the requested mode (e.g., single-GPU full FT on a 32GB consumer card), say so plainly in the plan and propose the alternative (LoRA, multi-GPU, etc.), rather than silently downgrading.

.claude/skills/train-model/references/troubleshooting.md

@@ -0,0 +1,84 @@
+# Troubleshooting
+
+Quick lookup for failures during sanity check, preprocessing, or training. For deeper coverage see `packages/ltx-trainer/docs/troubleshooting.md`.
+
+## OOM During Training Step
+
+Order of operations (cheapest first):
+
+1. `optimization.enable_gradient_checkpointing: true` (if not already on).
+2. `optimization.batch_size: 1` and increase `gradient_accumulation_steps` to preserve effective batch.
+3. `optimization.optimizer_type: "adamw8bit"`.
+4. `acceleration.quantization: "int8-quanto"`.
+5. Reduce `lora.rank` (32 → 16 → 8). Alpha follows rank.
+6. Reduce target resolution (`validation.video_dims` and re-preprocess the dataset at the new resolution).
+
+The last option is expensive — flag it clearly to the user before re-preprocessing.
+
+## OOM During Validation Sample Generation
+
+The validation pass loads decoders + runs CFG/STG inference; it can OOM even when the training step fits.
+
+1. `acceleration.load_text_encoder_in_8bit: true` (trainer config, not the dataset script).
+2. `acceleration.offload_optimizer_during_validation: true` (especially for full FT or high-rank LoRA).
+3. Reduce `validation.video_dims` (smaller validation than training is fine — it's only for visual feedback).
+4. Reduce `validation.inference_steps` (e.g. 30 → 20).
+5. Increase `validation.interval` to validate less often.
+
+## NaN Loss
+
+1. Check `acceleration.mixed_precision_mode`: prefer `"bf16"`. If `"fp16"`, switch.
+2. Verify dataset latents are well-formed: `uv run python scripts/decode_latents.py <latents-dir> <output-dir> --model-path <model>` (it decodes a whole latents directory, not a single `.pt`) should reconstruct sensibly.
+3. Lower `optimization.learning_rate` by 5x.
+4. Add `optimization.max_grad_norm: 1.0` (default; verify it's set).
+5. If using `quantization`, try `null` — INT8/INT4 quantization can interact badly with poorly-conditioned LoRA inits at high LR.
+
+## Validation Samples Look Wrong but Loss Is Fine
+
+Often not a bug — validation uses simplified inference. For real quality assessment, run a checkpoint through `packages/ltx-pipelines/` after training.
+
+## Trainer Won't Start: Config Validation Error
+
+Pydantic `extra="forbid"` means typos in field names fail loudly. Read the error carefully — it names the offending field and path. Fix and re-launch.
+
+Common offenders:
+- `latents_dir` typo or wrong relative path.
+- A required field genuinely missing after copying an example (most fields have defaults; check the error message for the exact field path).
+- `target_modules` listed at wrong nesting level (must be under `lora:`).
+
+## Trainer Won't Start: Missing Files
+
+- `model_path` not found → re-probe `/models/`, `~/models/`, `$LTX_MODELS_DIR`, or ask the user.
+- `text_encoder_path` directory missing the Gemma config → ensure the path is to the Gemma model dir, not its parent.
+- `preprocessed_data_root` doesn't contain expected subdirs → re-verify Phase 7 ran for the chosen mode (see `phases/preprocess-dataset.md`).
+
+## Resume Stops Working
+
+The trainer **does not auto-resume from `output_dir`**. Resume happens only when `model.load_checkpoint` is explicitly set to a checkpoint file; the trainer then loads those weights and looks for a `training_state_step_*.pt` next to that file to restore optimizer/scheduler/step. Common pitfalls:
+- `model.load_checkpoint` not set or set to the wrong path → fresh run from step 0 even when `outputs/checkpoints/` is full of artifacts. Patch `model.load_checkpoint` to the latest checkpoint.
+- `checkpoints.no_resume: true` is set → weights load but state is discarded. Remove the flag if you want a proper resume.
+- `training_state_step_*.pt` missing from next to the loaded checkpoint → weights load but step counter resets. Make sure the state file accompanies the checkpoint.
+- `training_state_step_*.pt` corrupted (size 0, fails `torch.load`) → trainer falls back to step 0 with a warning.
+
+## Autotune Trial Failed Mid-Sweep
+
+- If trial 2 (quantization off) OOMs: revert to trial 1 and stop the sweep. The 32GB tier baseline is correctly aggressive.
+- If trial 3 (adamw) OOMs: revert to adamw8bit. Continue with trial 4 if VRAM headroom allows.
+- If trial 4 (batch_size up) OOMs: revert and stop. We've found the ceiling.
+
+Never carry over a failing trial's deltas. Always revert to the last-known-good before the next change.
+
+## Captioning Is Slow
+
+- Local `qwen_omni` is a 30B model served by `serve_captioner.py` (vLLM). If the server won't start or OOMs on launch: keep the default `--quantization fp8` (don't use `bf16` unless ≥66 GiB free VRAM), lower `--gpu-memory-utilization`, or reduce `--max-model-len`. If the GPU can't host a 30B model at all, switch to `gemini_flash`.
+- `caption_videos.py --captioner-type qwen_omni` errors connecting → the vLLM server isn't running or `--vllm-url` doesn't match. Start `serve_captioner.py` first and confirm the port.
+- For most hardware and for larger datasets, prefer `--captioner-type gemini_flash --num-workers <N>` (needs Gemini auth: `GEMINI_API_KEY`/`GOOGLE_API_KEY` or gcloud/Vertex) — runs anywhere and parallelises; local Qwen needs a heavy GPU and a running server.
+
+## Process_Dataset Errors
+
+- "frames divisible by..." → the video doesn't have enough frames at the requested temporal resolution. Either shorten the requested frame count or use `split_scenes.py` to break long videos.
+- "shape mismatch" on existing `.precomputed/` → user requested a different resolution than the existing data. Per the invariants, **stop and ask** — do not overwrite. Offer: reuse at old resolution / re-preprocess to a new dir / abort.
+
+## When To Give Up and Ask The User
+
+If a fix isn't obvious from this file or `packages/ltx-trainer/docs/troubleshooting.md` within two attempts, stop and surface the full error + the steps already tried to the user. Don't loop indefinitely on autonomous fixes — the user has context the agent doesn't (which checkpoints are precious, what they care about preserving, etc.).

.gitattributes

@@ -2,7 +2,9 @@
 *.safetensors filter=lfs diff=lfs merge=lfs -text
 *.sft filter=lfs diff=lfs merge=lfs -text
 *.pt filter=lfs diff=lfs merge=lfs -text
+*.mp3 filter=lfs diff=lfs merge=lfs -text
 *.mp4 filter=lfs diff=lfs merge=lfs -text
+packages/ltx-pipelines/tests/assets/*.wav filter=lfs diff=lfs merge=lfs -text
 *.png filter=lfs diff=lfs merge=lfs -text
 *.jpeg filter=lfs diff=lfs merge=lfs -text
 *.jpg filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -17,8 +17,10 @@ checkpoints/
 
 # Other files
 .DS_Store
-tmp
 .wandb
+projects/
+tmp
+wandb/
 
 # Model checkpoints
 *.ckpt
@@ -36,14 +38,38 @@ tmp
 *.json
 *.m4a
 *.mov
+*.mp3
 *.mp4
 *.png
 *.wav
 *.webp
 
-# HDR IC-LoRA e2e test baseline (checked in via Git LFS)
+# Full-params expected results for the --full-e2e quality lane (checked in via Git LFS)
-!packages/ltx-pipelines/tests/assets/expected_hdr_ic_lora_exr/frame_*.exr
+!packages/ltx-pipelines/tests/assets/full_e2e/
+!packages/ltx-pipelines/tests/assets/full_e2e/*.mp4
+!packages/ltx-pipelines/tests/assets/full_e2e/*.wav
+!packages/ltx-pipelines/tests/assets/full_e2e/expected_hdr_ic_lora_exr/frame_*.exr
+
+# HDR IC-LoRA e2e test input clip (checked in via Git LFS)
 !packages/ltx-pipelines/tests/assets/hdr_ic_lora_test_input.mp4
 
+# Fast integration-profile goldens (bit-exact decoded video/audio; checked in via Git LFS)
+!packages/ltx-pipelines/tests/assets/integration/
+!packages/ltx-pipelines/tests/assets/integration/*.safetensors
+
+# ltx-bench Grafana dashboards (source of truth in the repo)
+!packages/ltx-bench/grafana/*.json
+
+# ltx-trainer E2E test dataset (committed via Git LFS).
+# The target_audio/ and reference_audio/ subdirectories are populated lazily at test runtime
+# (high-bitrate MP3 extracted from target_videos, plus a low-bitrate MP3 reference); their
+# contents stay gitignored via the root *.mp3 rule above.
+!packages/ltx-trainer/tests/assets/test_dataset/dataset.json
+!packages/ltx-trainer/tests/assets/test_dataset/target_videos/*.mp4
+!packages/ltx-trainer/tests/assets/test_dataset/reference_videos/*.mp4
+!packages/ltx-trainer/tests/assets/test_dataset/conditioning/first_frame.jpg
+!packages/ltx-trainer/tests/assets/test_dataset/conditioning/video_mask.png
+!packages/ltx-trainer/tests/assets/test_dataset/conditioning/audio_mask.pt
+
 # Binary files
 *.so

README.md

@@ -14,19 +14,43 @@
 
 ## 🚀 Quick Start
 
+Clone the repo
+
 ```bash
-# Clone the repository
 git clone https://github.com/Lightricks/LTX-2.git
 cd LTX-2
-
-# Set up the environment
-uv sync --frozen
-source .venv/bin/activate
 ```
 
-### Required Models
+Download the relevant [models](https://huggingface.co/Lightricks/LTX-2.3) or use the [Hugging Face CLI](https://huggingface.co/docs/huggingface_hub/guides/cli)
 
-Download the following models from the [LTX-2.3 HuggingFace repository](https://huggingface.co/Lightricks/LTX-2.3):
+```bash
+hf auth login
+hf download Lightricks/LTX-2.3 \
+    ltx-2.3-22b-distilled-1.1.safetensors ltx-2.3-spatial-upscaler-x2-1.1.safetensors --local-dir models/ltx-2.3
+hf download google/gemma-3-12b-it-qat-q4_0-unquantized --local-dir models/gemma-3-12b
+```
+
+If you get a 401/403, accept the model terms on Hugging Face and log in with a **Read** token (fine-grained tokens need the "read gated repos" scope enabled).
+
+Generate
+
+```bash
+uv run python -m ltx_pipelines.distilled \
+    --distilled-checkpoint-path models/ltx-2.3/ltx-2.3-22b-distilled-1.1.safetensors \
+    --spatial-upsampler-path    models/ltx-2.3/ltx-2.3-spatial-upscaler-x2-1.1.safetensors \
+    --gemma-root models/gemma-3-12b \
+    --seed 42 \
+    --output-path output.mp4 \
+    --prompt "A medium close-up shot features a Caucasian man with a beard, wearing a green and white baseball cap without any letters on the front, and a light blue shirt over a white t-shirt. He is positioned in the center of the frame, looking intently directly at the camera, his eyes focused on camera. His facial expression is one of deep concentration, with his brow slightly raised. As he looks straight at the camera, a quick sniff sound is heard, and then he speaks with a deep male voice and a satisfied tone, saying, 'I think it's so good.' The camera remains static throughout, maintaining a shallow depth of field, which keeps the man in sharp focus while the background is softly blurred, showing a beige wall behind him. After a brief pause, another short, audible sniff is heard. The man then continues to speak, his voice maintaining the same quality, as he states, 'So good. So good.' He elaborates further, emphasizing his point with a final statement, 'This got to be, it's got to be the best tool I've ever seen.'"
+```
+
+In cases of GPU memory constraints, consider `--quantization fp8-cast --offload {cpu, disk}`. See [additional flags](packages/ltx-pipelines/docs/installation.md#common-cli-flags).
+
+This uses the distilled model and pipeline for fast results. For better quality or other capabilities, see [Models](#full-model-list) and [Pipelines](#available-pipelines).
+
+### Full Model List
+
+For pipelines beyond the quickstart, download the relevant models from the [LTX-2.3 HuggingFace repository](https://huggingface.co/Lightricks/LTX-2.3):
 
 **LTX-2.3 Model Checkpoint** (choose and download one of the following)
   * [`ltx-2.3-22b-dev.safetensors`](https://huggingface.co/Lightricks/LTX-2.3/blob/main/ltx-2.3-22b-dev.safetensors) - [Download](https://huggingface.co/Lightricks/LTX-2.3/resolve/main/ltx-2.3-22b-dev.safetensors)
@@ -39,7 +63,7 @@ Download the following models from the [LTX-2.3 HuggingFace repository](https://
 **Temporal Upscaler** - Supported by the model and will be required for future pipeline implementations
   * [`ltx-2.3-temporal-upscaler-x2-1.0.safetensors`](https://huggingface.co/Lightricks/LTX-2.3/blob/main/ltx-2.3-temporal-upscaler-x2-1.0.safetensors) - [Download](https://huggingface.co/Lightricks/LTX-2.3/resolve/main/ltx-2.3-temporal-upscaler-x2-1.0.safetensors)
 
-**Distilled LoRA** - Required for current two-stage pipeline implementations in this repository (except DistilledPipeline and ICLoraPipeline)
+**Distilled LoRA** - Required for current two-stage pipeline implementations in this repository (except DistilledPipeline, ICLoraPipeline, and LipDubPipeline)
   * [`ltx-2.3-22b-distilled-lora-384-1.1.safetensors`](https://huggingface.co/Lightricks/LTX-2.3/blob/main/ltx-2.3-22b-distilled-lora-384-1.1.safetensors) - [Download](https://huggingface.co/Lightricks/LTX-2.3/resolve/main/ltx-2.3-22b-distilled-lora-384-1.1.safetensors)
 
 **Gemma Text Encoder** (download all assets from the repository)
@@ -58,6 +82,7 @@ Download the following models from the [LTX-2.3 HuggingFace repository](https://
   * [`LTX-2-19b-LoRA-Camera-Control-Jib-Up`](https://huggingface.co/Lightricks/LTX-2-19b-LoRA-Camera-Control-Jib-Up) - [Download](https://huggingface.co/Lightricks/LTX-2-19b-LoRA-Camera-Control-Jib-Up/resolve/main/ltx-2-19b-lora-camera-control-jib-up.safetensors)
   * [`LTX-2-19b-LoRA-Camera-Control-Static`](https://huggingface.co/Lightricks/LTX-2-19b-LoRA-Camera-Control-Static) - [Download](https://huggingface.co/Lightricks/LTX-2-19b-LoRA-Camera-Control-Static/resolve/main/ltx-2-19b-lora-camera-control-static.safetensors)
   * [`LTX-2.3-22b-IC-LoRA-HDR`](https://huggingface.co/Lightricks/LTX-2.3-22b-IC-LoRA-HDR) - HDR IC-LoRA and pre-computed text embeddings for `HDRICLoraPipeline`
+  * [`LTX-2.3-22b-IC-LoRA-LipDub`](https://huggingface.co/Lightricks/LTX-2.3-22b-IC-LoRA-LipDub) - [Download](https://huggingface.co/Lightricks/LTX-2.3-22b-IC-LoRA-LipDub/resolve/main/ltx-2.3-22b-ic-lora-lipdub-0.9.safetensors)
 
 ### Available Pipelines
 
@@ -70,13 +95,14 @@ Download the following models from the [LTX-2.3 HuggingFace repository](https://
 * **[A2VidPipelineTwoStage](packages/ltx-pipelines/src/ltx_pipelines/a2vid_two_stage.py)** - Audio-to-video generation conditioned on an input audio file
 * **[RetakePipeline](packages/ltx-pipelines/src/ltx_pipelines/retake.py)** - Regenerate a specific time region of an existing video
 * **[HDRICLoraPipeline](packages/ltx-pipelines/src/ltx_pipelines/hdr_ic_lora.py)** - Video-to-video with HDR output (linear float frames via LogC3 inverse decode, suitable for EXR export and tonemapping)
+* **[LipDubPipeline](packages/ltx-pipelines/src/ltx_pipelines/lipdub.py)** - Lip dubbing, rephrasing, matching speaker identity (distilled model, single IC-LoRA, Two stages).
 
 ### ⚡ Optimization Tips
 
 * **Use DistilledPipeline** - Fastest inference with only 8 predefined sigmas (8 steps stage 1, 4 steps stage 2)
-* **Enable FP8 quantization** - Enables lower memory footprint: `--quantization fp8-cast` (CLI) or `quantization=QuantizationPolicy.fp8_cast()` (Python). Fp8-cast should be used with bf16 checkpoints, it shall downcast them on the fly. For Hopper GPUs with TensorRT-LLM, use `--quantization fp8-scaled-mm` for FP8 scaled matrix multiplication. Fp8-scaled-mm should be used with fp8 checkpoints.
+* **Enable FP8 quantization** - Enables lower memory footprint: `--quantization fp8-cast` (CLI) or `quantization=QuantizationPolicy.fp8_cast()` (Python). Fp8-cast should be used with bf16 checkpoints, it shall downcast them on the fly. On Hopper+ GPUs with native FP8 support, use `--quantization fp8-scaled-mm` for FP8 scaled matrix multiplication. Fp8-scaled-mm should be used with fp8 checkpoints.
-* **Install attention optimizations** - Use xFormers (`uv sync --extra xformers`) or [Flash Attention 3](https://github.com/Dao-AILab/flash-attention) for Hopper GPUs
+* **Install attention optimizations** - On datacenter Blackwell GPUs (B200), install FlashAttention 4 manually: `uv pip install 'flash-attn-4==4.0.0b9'` (this specific revision is the one we have verified against torch 2.9.1+cu128; newer betas have known issues on consumer Blackwell). On Hopper GPUs, install the FlashAttention 3 wheel. On other CUDA GPUs, PyTorch SDPA is used automatically. An installed backend is selected automatically at runtime; forcing a specific one is a Python-API option (`AttentionFunction.FLASH_ATTENTION_3`/`FLASH_ATTENTION_4`), not a CLI flag.
-* **Use gradient estimation** - Reduce inference steps from 40 to 20-30 while maintaining quality (see [pipeline documentation](packages/ltx-pipelines/README.md#denoising-loop-optimization))
+* **Use gradient estimation** - Reduce inference steps from 40 to 20-30 while maintaining quality (see [pipeline documentation](packages/ltx-pipelines/docs/optimization.md#denoising-loop-optimization))
 * **Skip memory cleanup** - If you have sufficient VRAM, disable automatic memory cleanup between stages for faster processing
 * **Choose single-stage pipeline** - Use `TI2VidOneStagePipeline` for faster generation when high resolution isn't required
 
@@ -92,7 +118,7 @@ When writing prompts, focus on detailed, chronological descriptions of actions a
 - Describe lighting and colors
 - Note any changes or sudden events
 
-For additional guidance on writing a prompt please refer to <https://ltx.video/blog/how-to-prompt-for-ltx-2>
+For additional guidance on writing a prompt please refer to <https://ltx.io/blog/prompting-guide-for-ltx-2>
 
 ### Automatic Prompt Enhancement
 

docs/plan.md

@@ -0,0 +1,70 @@
+# SCAIL-2 → LTX-2 移植計畫
+
+將 SCAIL-2（arXiv:2606.10804，`zai-org/SCAIL-2`）的端到端角色動畫手法移植到 LTX-2。
+
+> **注意**：SCAIL-2 原始實作建構於 **Wan 2.1**，非 LTX-2。座標慣例與架構需翻譯到 LTX 的資料流。
+
+## 三個核心機制
+
+1. **Driving latent 直接串接** — 把驅動影片 latent 直接接進 DiT token 序列（不經骨架/pose 中介），用 width 軸座標偏移 ΔW 讓 driving 座標跟主 video 座標分開。
+2. **In-context mask channel** — 疊加額外輸入 channel（1 個環境開關 + K=6 個角色綁定槽，展開為 `4(K+1)=28` channel）到模型輸入，讓模型知道背景/角色對應。
+3. **Mode-specific RoPE** — Animation Mode 與 Replacement Mode 用不同的座標指派規則。
+
+## LTX-2 對應落點
+
+| 機制 | LTX-2 落點 | 改權重？ |
+|---|---|---|
+| 1. Driving 串接 + ΔW | 新 `ConditioningItem`（clone `reference_video_cond.py`），改 `positions` 偏移 | 否 |
+| 3. Mode-specific RoPE | 上述 item 加 `DrivingMode` enum | 否 |
+| 2. In-context mask channel | 加寬 `patchify_proj` 輸入 channel + `LatentState` 帶額外 channel + 投影前 concat | **是**（需微調） |
+
+**關鍵洞察**：LTX 的 `ConditioningItem.apply_to(latent_state, latent_tools) -> LatentState` 就是「串接進序列」的天然注入點，機制 1+3 **完全不用動 transformer / rope.py**。RoPE 由 `LatentState.positions` `[B,3,T,2]`（pixel 座標，axis1=(time,h,w)）驅動；ΔW = 對 `positions[:,2]`（width）加常數。width 正規化上界 `max_pos[2]=2048`，超過會 wrap。
+
+## SCAIL-2 座標規則（論文）
+
+序列 `[z_ref; z_t; z_driv]`，driving 永遠在 width 軸帶固定偏移 ΔW：
+
+| | Animation | Replacement |
+|---|---|---|
+| z_ref | T=0, H=[0,Hv), W=[0,Wv) | T=0, **H=[ΔH_ref, ΔH_ref+Hv)**, W=[0,Wv) |
+| z_t | T=[1,Tv], H=[0,Hv), W=[0,Wv) | T=[0,Tv−1], H=[0,Hv), W=[0,Wv) |
+| z_driv | T=[1,Tv], H=[0,Hv), **W=[ΔW, ΔW+Wv)** | T=[0,Tv−1], H=[0,Hv), **W=[ΔW, ΔW+Wv)** |
+
+## 分階段計畫
+
+### Phase 1 — Driving 串接 item（機制 1+3，推論期 PoC）✅ 已完成
+- 純推論、不改權重、不動 `patchify_proj`、不碰 ltx-trainer。
+- 用現有 checkpoint 驗證資料流正確性（序列長度、座標偏移、attention mask、denoise_mask、不 wrap）。
+- **PoC 僅驗證 plumbing**；LTX-2 未經此訓練，畫面不會是正確動畫。
+
+### Phase 2 — In-context mask channel（機制 2，checkpoint 手術）✅ 已完成（plumbing）
+- **決策**：時間編碼採忠實堆疊，`8×(K+1)=56` channel（LTX 時間因子 8，K=6），非 Wan 的 28。
+- `LTXModel(mask_conditioning_channels=0)` config-gated 加寬 `patchify_proj`；預設不變。
+- `LatentState`/`Modality` 加 `cond_channels` 欄位，跟著 clone/clear/append 流動；投影前在 `_apply_patchify_proj` concat（None 補零）。
+- 新輸入欄位 **zero-init**：`widen_patchify_proj_for_mask_channels` 轉換舊 checkpoint，行為不變、待微調才生效。
+- `VideoConditionByMaskChannels`：(K+1) 語意 mask → 空間下採樣 + 時間 8× 堆疊 → 寫入尾端 driving token（target 保持零，符合論文）。
+- 驗證通過（`verify_mask_channels.py`）：向後相容、zero-init 加寬 forward == baseline、mask pipeline 不 crash。**僅驗證 plumbing，畫質需 Phase 3 微調。**
+
+### Phase 3 — 訓練整合（ltx-trainer）✅ 已完成（程式碼路徑；實訓需 GPU）
+- **決策**：完整整合到 `FlexibleStrategy`；訓練 = LoRA + 解凍 `patchify_proj`（新 mask 欄位無法純 LoRA 訓練）。
+- 新 `DrivingConditionConfig` + `MaskChannelsConditionConfig`（`flexible.py`）；`_apply_driving_condition`(cond-first concat + ΔW) + `_build_mask_channels`(重用 `encode_mask_channels`) → `Modality.cond_channels`。
+- `load_transformer(mask_conditioning_channels=)` 用 `widen_module_patchify_proj_for_mask_channels` 加寬；`ModelConfig.mask_conditioning_channels`；trainer 在 LoRA 模式解凍 patchify_proj。
+- `configs/scail_animation_lora.yaml` + docs。CPU 單元驗證通過（`verify_phase3_trainer.py`）。
+- **本機無 GPU/Linux/checkpoint → 未跑實機訓練**；dataset 前處理（driving latents + 語意 mask）與 validation runner 接線未做。
+
+### Phase 4 — Pipeline + CLI 包裝 ✅ 已完成（程式碼路徑；實跑需 GPU）
+- `LTXModelConfigurator` 讀 `mask_conditioning_channels` → SCAIL checkpoint 自描述、載入自動加寬（不需 runtime wrapper）。
+- `scail_animation.py`：`ScailAnimationPipeline`（仿 `distilled.py` 兩階段）+ 可測純函式 `build_scail_conditionings`（driving + mask 條件組裝）+ `load_masks` + `main()`。
+- `utils/args.py` `scail_animation_arg_parser`（`--driving-video/--mask-path/--mode/--driving-strength`）。
+- CPU 驗證通過（`verify_phase4_pipeline.py`）；`ltx-pipelines/CLAUDE.md` 表格 row。
+- **實機端到端需 GPU + SCAIL-trained checkpoint**（config 含 `mask_conditioning_channels` + 訓練好的 patchify_proj + LoRA）。
+
+## Phase 1 簡化取捨（記錄，Phase 2 需回頭處理）
+- (a) driving 時間座標直接複製 target 的（token-wise），故 driving 需與 target 同 F/H/W。
+- (b) 單一 frozen driving group 的 `attention_mask` 維持 None（= 全連接，target 完全看得到 driving），與 reference cond 一致。
+- (c) **ANIMATION 與 REPLACEMENT 在 Phase 1 產生相同 driving 座標** — mode 差異（z_ref 的 ΔH_ref 高度位移、target 時間原點、mask channel）屬 Phase 2，enum 先保留佔位。
+
+## 參考
+- 論文：arXiv:2606.10804 — *SCAIL-2: Unifying Controlled Character Animation with End-to-end In-Context Conditioning*
+- 官方實作：`zai-org/SCAIL-2`（GitHub / HuggingFace），建構於 Wan 2.1
+- 藍本檔案：`packages/ltx-core/src/ltx_core/conditioning/types/reference_video_cond.py`

docs/tasks.md

@@ -0,0 +1,59 @@
+# SCAIL-2 → LTX-2 移植任務追蹤
+
+狀態圖例：✅ 完成 ｜ 🔄 進行中 ｜ ⬜ 未開始 ｜ ⏸️ 暫緩
+
+相關計畫見 [`plan.md`](./plan.md)。
+
+## Phase 1 — Driving 串接 item（推論期 PoC）
+
+| # | 任務 | 狀態 | 產出 / 備註 |
+|---|---|---|---|
+| 1.1 | 新增 `VideoConditionByDrivingLatent` + `DrivingMode` | ✅ | `packages/ltx-core/src/ltx_core/conditioning/types/driving_video_cond.py`，以 `reference_video_cond.py` 為藍本，實作 ΔW width 偏移、時間對齊 target、`max_pos` 防呆、token 數防呆 |
+| 1.2 | 匯出新 conditioning 類別 | ✅ | `conditioning/types/__init__.py`、`conditioning/__init__.py` |
+| 1.3 | 免-GPU 資料流驗證腳本 | ✅ | 序列長度、ΔW 不重疊、時間對齊、frozen denoise_mask、attention_mask=None、`clear_conditioning` 剝除、超界/token 數防呆 — 全通過；ruff clean |
+| 1.4 | （選配）端到端 smoke run | ✅ | 本機 CPU-only、無 checkpoint，改用小型真實 `LTXModel`（隨機權重）跑 pipeline 實走路徑：`create_noised_state`（含 driving cond）→ `modality_from_latent_state` → **真 transformer forward**（seq 160=target 80+driving 80）→ `clear_conditioning`（→80）。ANIMATION/REPLACEMENT 皆通過：不 crash、輸出 finite、driving frozen 且正確剝除。**僅驗證整合，不評估畫質** |
+
+## Phase 2 — In-context mask channel（checkpoint 手術）
+
+> **決策**：mask channel 時間編碼採**忠實堆疊**——每 latent 幀對應 8 個 pixel 子幀沿 channel 堆疊，`8×(K+1)=56` channel（LTX 時間因子 8，K=6）。非 Wan 的 4×(K+1)=28。
+
+| # | 任務 | 狀態 | 備註 |
+|---|---|---|---|
+| 2.1 | `patchify_proj` 加寬（config-gated） | ✅ | `LTXModel(mask_conditioning_channels=0)`；>0 時 `patchify_proj=Linear(in+mask, inner)`。預設不變 |
+| 2.2 | `LatentState`/`Modality` 帶 `cond_channels` 欄位 | ✅ | `types.py`、`modality.py` 加 `cond_channels: Tensor\|None=None`（patchified [B,T,C]）；`tools.clear_conditioning` 裁切；`helpers.modality_from_latent_state` 帶入 |
+| 2.3 | 投影前 concat/zero-pad cond_channels | ✅ | `transformer_args.py` `_apply_patchify_proj`：寬度不足時用 cond_channels 補齊，None 則補零 |
+| 2.4 | checkpoint zero-init 加寬轉換 | ✅ | `mask_channels_checkpoint.py` `widen_patchify_proj_for_mask_channels`：尾端補零欄，載入舊權重行為不變 |
+| 2.5 | `VideoConditionByMaskChannels`（56ch） | ✅ | `mask_channels_cond.py`：1 環境開關 + K=6 綁定槽 → 空間下採樣 + 時間 8× 堆疊(causal 首幀複製)=56ch，寫入尾端 driving token，target 保持零 |
+| 2.6 | token-append conditioning 延伸 cond_channels | ✅ | `cond_channels.py` `extend_cond_channels`；reference_video/reference_audio/driving/keyframe 皆接上 |
+| 2.7 | Phase 2 驗證（免訓練） | ✅ | `verify_mask_channels.py`：mask=0 向後相容；zero-init 加寬 forward == baseline（任意 cond_channels）；driving+mask pipeline forward 不 crash、cond_channels 形狀/placement 正確、`clear_conditioning` 剝除 |
+| 2.8 | Replacement Mode z_ref 高度位移 ΔH_ref | ⬜ | 仍暫緩（Phase 1 divergence，需獨立 reference token group） |
+
+## Phase 3 — 訓練整合（ltx-trainer）
+
+> **決策**：完整整合到 `FlexibleStrategy`；訓練方式 = **LoRA + 解凍 patchify_proj**（新 mask 欄位無法純 LoRA 訓練）。**本機無 GPU/Linux/checkpoint，只做到 CPU 單元驗證**，實訓需在 GPU 機器跑。
+
+| # | 任務 | 狀態 | 備註 |
+|---|---|---|---|
+| 3.1 | 新增 Driving/MaskChannels ConditionConfig | ✅ | `flexible.py`：`DrivingConditionConfig`(latents_dir/mode/width_offset) + `MaskChannelsConditionConfig`(mask_dir/num_slots)，加入 union + `get_data_sources` |
+| 3.2 | strategy 接線 driving concat + cond_channels | ✅ | `_apply_driving_condition`(cond-first concat + ΔW width 偏移) + `_build_mask_channels`(重用 `encode_mask_channels`，寫前 N driving token) → `Modality.cond_channels` |
+| 3.3 | model_loader + ModelConfig 支援加寬 | ✅ | `widen_module_patchify_proj_for_mask_channels`(live module zero-init) + `load_transformer(mask_conditioning_channels=)` + `ModelConfig.mask_conditioning_channels` |
+| 3.4 | LoRA 模式解凍 patchify_proj | ✅ | `trainer._unfreeze_patchify_proj`：mask_channels>0 時把 video patchify_proj 設 trainable，讓新欄位隨 LoRA 一起訓 |
+| 3.5 | 範例 config + docs | ✅ | `configs/scail_animation_lora.yaml`；`configs/README.md` 與 `docs/training-modes.md` 表格 row |
+| 3.6 | CPU 單元驗證 | ✅ | `verify_phase3_trainer.py`：config round-trip、prepare_training_inputs 建 cond_channels[B,T,56]、driving 前置/mask placement、widened model forward + compute_loss finite、widen helper zero-init 等價 |
+| 3.7 | dataset 前處理（產 driving latents + 語意 mask） | 🟡 | `char_masks/` 前處理已做：`scripts/process_char_masks.py`（label-map 影片/圖 → `[K+1,F_pix,H,W]`，對齊 target latent，nearest 保留整數 label，ch0 環境開關）。`driving_latents/` 沿用既有 `process_videos.py`（driving 影片走 video latent 路徑，同 target 形狀）。**仍缺**：從原始影片產生 label-map 的分割/追蹤步驟（SAM 等，資料集特定，未含） |
+| 3.8 | validation runner 接 driving/mask | ⬜ | 驗證期取樣尚未接 SCAIL 條件（config 內 validation 先停用），與 Phase 4 一起 |
+| 3.9 | 實機訓練跑通 | ⬜ | 需 Linux + GPU + checkpoint，本機無法 |
+
+## Phase 4 — Pipeline + CLI
+
+| # | 任務 | 狀態 | 備註 |
+|---|---|---|---|
+| 4.1 | configurator 讀 `mask_conditioning_channels` | ✅ | `LTXModelConfigurator` 兩分支 `config.get("mask_conditioning_channels",0)` → SCAIL checkpoint 自描述、載入自動加寬（不需 runtime widen wrapper） |
+| 4.2 | `ScailAnimationPipeline` + `build_scail_conditionings` | ✅ | `scail_animation.py`：仿 `distilled.py` 兩階段，注入 SCAIL driving+mask conditioning；`build_scail_conditionings` 為可測純函式；`load_masks` helper |
+| 4.3 | `scail_animation_arg_parser` + `main()` | ✅ | `utils/args.py`：在 `default_2_stage_distilled_arg_parser` 上加 `--driving-video/--mask-path/--mode/--driving-strength` |
+| 4.4 | CPU 驗證 + docs | ✅ | `verify_phase4_pipeline.py`：import、CLI round-trip、conditioning assembly(driving append + mask 在尾端)、configurator 自描述加寬；`ltx-pipelines/CLAUDE.md` pipeline 表格 row |
+| 4.5 | 實機端到端跑通 | ⬜ | 需 GPU + SCAIL-trained checkpoint（含 `mask_conditioning_channels` config + 訓練好的 patchify_proj + LoRA） |
+
+## 決議紀錄
+- **範圍**：先只做 Phase 1（推論期 PoC）。Phase 2+ 待 Phase 1 驗證後再討論。
+- **架構前提**：SCAIL-2 建構於 Wan 2.1，本移植為跨架構移植。

modal/README.md

@@ -0,0 +1,111 @@
+# SCAIL-2 on Modal — free testing environment
+
+Test the SCAIL-2 LTX-2 training/inference **code path** on Modal without a local
+GPU. The free path uses tiny random-init models + synthetic data (no checkpoint,
+no dataset), so it validates that the SCAIL wiring runs on real Linux/GPU — not
+model quality.
+
+> ⚠️ The real 19B LTX-2 model is **not** free to train/run. The `verify`/`smoke`
+> targets are near-free; `train` on a real checkpoint uses a paid GPU.
+
+## 1. One-time setup
+
+1. Sign up at [modal.com](https://modal.com) (the free Starter plan includes a
+   monthly credit allowance — enough for many `verify`/`smoke` runs).
+2. Install and authenticate:
+   ```bash
+   pip install modal
+   modal setup          # opens a browser to link your account / token
+   ```
+
+## 2. Free / near-free checks
+
+From the repo root:
+
+```bash
+# CPU: SCAIL-2 Phase 1-4 plumbing (driving concat, mask channels, zero-init widen,
+# a FlexibleStrategy training step + loss, inference conditioning assembly).
+modal run modal/app.py::verify
+
+# Same checks on a T4 GPU (validates the CUDA path). ~cents.
+modal run modal/app.py::smoke
+```
+
+The first run builds the image (installs the `ltx-core`/`ltx-pipelines`/`ltx-trainer`
+workspace via `uv sync`). `ltx-kernels` (CUDA-compiled) is intentionally skipped —
+attention falls back to PyTorch SDPA, so no CUDA toolchain is required.
+
+Expected tail:
+```
+[OK] Phase 1 driving concat: seq 48 -> 96
+[OK] Phase 2 zero-init widen: in_features=72, output preserved
+[OK] Phase 3 training step: cond_channels (1, 96, 56), loss ...
+[OK] Phase 4 build_scail_conditionings: [driving, mask_channels]
+All SCAIL-2 checks passed on cuda   # (or cpu)
+```
+
+## 3. Scaling up to real training (paid)
+
+`train` runs `packages/ltx-trainer/scripts/train.py` against a real checkpoint.
+You must supply the weights + data via the persistent `scail-data` Volume.
+
+1. Create/populate the Volume (checkpoint, Gemma encoder, preprocessed latents):
+   ```bash
+   modal volume create scail-data          # if not auto-created
+   modal volume put scail-data /local/ltx-2-model.safetensors /model/ltx-2.safetensors
+   modal volume put scail-data /local/gemma /model/gemma
+   modal volume put scail-data /local/preprocessed /data/preprocessed
+   ```
+2. Copy `configs/scail_animation_lora.yaml`, and point its paths at the mounted
+   Volume (everything lands under `/data` in the container):
+   ```yaml
+   model:
+     model_path: "/data/model/ltx-2.safetensors"
+     text_encoder_path: "/data/model/gemma"
+     mask_conditioning_channels: 56       # widens patchify_proj at load
+   data:
+     preprocessed_data_root: "/data/preprocessed"
+   ```
+   (Put your edited config on the Volume too, or bake it into the repo.)
+3. Launch (pick a GPU big enough for the model — the 19B needs an A100):
+   ```bash
+   # edit gpu="A10G" -> "A100" in modal/app.py::train for the full model
+   modal run modal/app.py::train --config-rel /data/scail_animation_lora.yaml
+   ```
+
+### Dataset preprocessing (not yet automated for SCAIL)
+
+The SCAIL training config expects, under `preprocessed_data_root/`:
+
+```
+latents/           # target video latents
+conditions/        # text embeddings
+driving_latents/   # driving video latents (same F/H/W as target)
+char_masks/        # per-sample "mask" = [K+1, F_pix, H, W]  (ch0 env switch, 1..K binding slots)
+```
+
+`latents/`, `conditions/`, and `driving_latents/` come from the existing
+`packages/ltx-trainer/scripts/process_dataset.py` (run it once per video set).
+`char_masks/` is produced by `packages/ltx-trainer/scripts/process_char_masks.py`
+from per-sample **label-map** videos/images (integer pixel labels: `0` =
+environment, `1..K` = characters → binding slots):
+
+```bash
+python packages/ltx-trainer/scripts/process_char_masks.py dataset.csv \
+    --mask-column char_labels --latents-dir ./latents \
+    --output-dir ./char_masks --num-slots 6 --main-media-column media_path
+```
+
+You still need a segmentation/tracking model (e.g. SAM) to *produce* those label
+maps from raw video — that upstream step is dataset-specific and not included.
+
+## Cost notes
+
+| Target  | GPU   | Rough cost | Use |
+|---------|-------|-----------|-----|
+| `verify`| none  | ~free      | validate code path on CPU |
+| `smoke` | T4    | cents      | validate CUDA path |
+| `train` | A10G/A100 | paid   | real fine-tuning (needs checkpoint + data) |
+
+Free credits are best spent on `verify`/`smoke` to catch integration issues before
+committing a paid GPU to a real run. Watch usage in the Modal dashboard.

modal/app.py

@@ -0,0 +1,84 @@
+"""Modal app for testing the SCAIL-2 LTX-2 training/inference path.
+
+Free-tier friendly: `verify` runs on CPU and `smoke` on a cheap T4, both using
+tiny random-init models + synthetic data (no checkpoint, no dataset) so a full run
+costs pennies. `train` is the scale-up entrypoint that runs the real trainer once
+you upload a checkpoint + preprocessed data to the `scail-data` Volume -- that one
+uses a real GPU and is NOT free.
+
+Setup (once):
+    pip install modal && modal setup
+
+Run:
+    modal run modal/app.py::verify     # CPU plumbing check (~free)
+    modal run modal/app.py::smoke      # same checks on a T4 GPU (cents)
+    modal run modal/app.py::train --config-rel configs/scail_animation_lora.yaml
+"""
+
+import subprocess
+from pathlib import Path
+
+import modal
+
+REPO = Path(__file__).parent.parent
+
+# Build the workspace once into the image. ltx-kernels (CUDA-compiled) is excluded
+# from the workspace and not needed -- attention falls back to SDPA.
+image = (
+    modal.Image.debian_slim(python_version="3.11")
+    .apt_install("git")
+    .pip_install("uv")
+    .env({"UV_LINK_MODE": "copy", "UV_PROJECT_ENVIRONMENT": "/root/LTX-2/.venv"})
+    .add_local_dir(
+        str(REPO),
+        "/root/LTX-2",
+        copy=True,
+        ignore=[".git", ".venv", "**/__pycache__", "**/*.pyc", "outputs", "wandb", "**/.pytest_cache"],
+    )
+    .run_commands("cd /root/LTX-2 && uv sync --package ltx-trainer")
+)
+
+app = modal.App("scail-ltx2", image=image)
+
+# Persistent volume for the (large) checkpoint + preprocessed dataset used by `train`.
+data_volume = modal.Volume.from_name("scail-data", create_if_missing=True)
+
+_UV_PY = ["uv", "run", "--package", "ltx-trainer", "python"]
+
+
+def _run(args: list[str]) -> None:
+    subprocess.run([*_UV_PY, *args], cwd="/root/LTX-2", check=True)
+
+
+@app.function(timeout=1800)
+def verify() -> None:
+    """Run the SCAIL-2 plumbing checks on CPU (near-free)."""
+    _run(["modal/checks.py"])
+
+
+@app.function(gpu="T4", timeout=1800)
+def smoke() -> None:
+    """Run the same checks on a T4 GPU to validate the CUDA path (cents)."""
+    _run(["modal/checks.py"])
+
+
+@app.function(gpu="A10G", timeout=60 * 60 * 6, volumes={"/data": data_volume})
+def train(config_rel: str = "configs/scail_animation_lora.yaml") -> None:
+    """Run the real SCAIL trainer. NOT free -- needs a real checkpoint + data.
+
+    Upload your base checkpoint, Gemma text encoder, and preprocessed data to the
+    ``scail-data`` Volume (mounted at /data), and point the config's paths at /data.
+    For the full 19B model use a bigger GPU (e.g. gpu="A100") and expect real cost.
+    """
+    _run(["packages/ltx-trainer/scripts/train.py", config_rel])
+    data_volume.commit()
+
+
+@app.local_entrypoint()
+def main(target: str = "verify", config_rel: str = "configs/scail_animation_lora.yaml") -> None:
+    if target == "smoke":
+        smoke.remote()
+    elif target == "train":
+        train.remote(config_rel)
+    else:
+        verify.remote()

modal/checks.py

@@ -0,0 +1,159 @@
+# ruff: noqa: T201
+"""Device-aware SCAIL-2 smoke checks (CPU or GPU), runnable anywhere.
+
+Consolidates the Phase 1-4 plumbing checks into one script that auto-selects CUDA
+when available. It uses tiny, randomly-initialised models and synthetic data, so it
+needs **no checkpoint and no dataset** -- ideal for a near-free Modal test that
+validates the SCAIL training/inference code path end to end in a real Linux/GPU
+environment before spending credits on the full 19B model.
+
+Run locally:   uv run --package ltx-trainer python modal/checks.py
+On Modal:      modal run modal/app.py::verify        (CPU)
+               modal run modal/app.py::smoke         (T4 GPU)
+"""
+
+from __future__ import annotations
+
+from dataclasses import replace
+
+import torch
+
+from ltx_core.components.noisers import GaussianNoiser
+from ltx_core.components.patchifiers import VideoLatentPatchifier
+from ltx_core.conditioning import DrivingMode, VideoConditionByDrivingLatent
+from ltx_core.model.transformer.mask_channels_checkpoint import (
+    widen_module_patchify_proj_for_mask_channels,
+)
+from ltx_core.model.transformer.model import LTXModel, LTXModelType
+from ltx_core.tools import VideoLatentTools
+from ltx_core.types import SpatioTemporalScaleFactors, VideoLatentShape
+from ltx_pipelines.scail_animation import build_scail_conditionings
+from ltx_pipelines.utils.helpers import create_noised_state, modality_from_latent_state
+from ltx_trainer.timestep_samplers import UniformTimestepSampler
+from ltx_trainer.training_strategies.flexible import (
+    DrivingConditionConfig,
+    FlexibleStrategy,
+    FlexibleStrategyConfig,
+    MaskChannelsConditionConfig,
+    ModalityConfig,
+)
+
+# --- config ---------------------------------------------------------------------
+B, C, F, H, W = 1, 16, 3, 4, 4
+HEADS, HEAD_DIM, LAYERS, INNER = 4, 16, 2, 64
+CTX_LEN, K = 8, 6
+SCALE = SpatioTemporalScaleFactors.default()
+MASK_CH = SCALE.time * (K + 1)  # 56
+F_PIX = (F - 1) * SCALE.time + 1
+N = F * H * W
+
+
+def _tiny_model(mask_channels: int, device: torch.device, dtype: torch.dtype) -> LTXModel:
+    model = LTXModel(
+        model_type=LTXModelType.VideoOnly,
+        num_attention_heads=HEADS,
+        attention_head_dim=HEAD_DIM,
+        in_channels=C,
+        out_channels=C,
+        num_layers=LAYERS,
+        cross_attention_dim=INNER,
+        mask_conditioning_channels=mask_channels,
+    )
+    for p in model.parameters():
+        torch.nn.init.normal_(p, std=0.02)
+    return model.to(device=device, dtype=dtype).eval()
+
+
+def main() -> None:
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    dtype = torch.float32
+    print(f"== SCAIL-2 checks on device={device} "
+          f"({torch.cuda.get_device_name(0) if device.type == 'cuda' else 'cpu'}) ==")
+    torch.manual_seed(0)
+
+    tools = VideoLatentTools(
+        patchifier=VideoLatentPatchifier(patch_size=1),
+        target_shape=VideoLatentShape(batch=B, channels=C, frames=F, height=H, width=W),
+        fps=24.0,
+        scale_factors=SCALE,
+    )
+
+    # Phase 1: driving-latent concat + ΔW RoPE (inference conditioning path).
+    driving = torch.randn(B, C, F, H, W, device=device, dtype=dtype)
+    noiser = GaussianNoiser(generator=torch.Generator(device=device).manual_seed(1))
+    state = create_noised_state(
+        tools, [VideoConditionByDrivingLatent(driving, mode=DrivingMode.ANIMATION)], noiser, dtype, device, 1.0
+    )
+    assert state.latent.shape[1] == 2 * N
+    print(f"[OK] Phase 1 driving concat: seq {N} -> {state.latent.shape[1]}")
+
+    # Phase 2: widen patchify_proj (zero-init) is output-preserving.
+    base = _tiny_model(0, device, dtype)
+    ctx = torch.randn(B, CTX_LEN, INNER, device=device, dtype=dtype)
+    sigma = torch.ones(B, device=device, dtype=dtype)
+    mod = modality_from_latent_state(state, context=ctx, sigma=sigma)
+    with torch.inference_mode():
+        out0, _ = base(video=mod, audio=None, perturbations=None)
+    widen_module_patchify_proj_for_mask_channels(base, MASK_CH)
+    mod_cc = replace(mod, cond_channels=torch.randn(B, 2 * N, MASK_CH, device=device, dtype=dtype))
+    with torch.inference_mode():
+        out1, _ = base(video=mod_cc, audio=None, perturbations=None)
+    assert torch.allclose(out0, out1, atol=1e-4), "zero-init widening changed output"
+    print(f"[OK] Phase 2 zero-init widen: in_features={base.patchify_proj.in_features}, output preserved")
+
+    # Phase 3: FlexibleStrategy driving + mask-channels training step.
+    cfg = FlexibleStrategyConfig(
+        name="flexible",
+        video=ModalityConfig(
+            is_generated=True,
+            latents_dir="video_latents",
+            conditions=[
+                DrivingConditionConfig(type="driving", latents_dir="driving_latents"),
+                MaskChannelsConditionConfig(type="mask_channels", mask_dir="char_masks", num_slots=K),
+            ],
+        ),
+    )
+    strategy = FlexibleStrategy(cfg)
+
+    def latents() -> dict:
+        return {
+            "latents": torch.randn(B, C, F, H, W, device=device),
+            "num_frames": torch.tensor([F]),
+            "height": torch.tensor([H]),
+            "width": torch.tensor([W]),
+            "fps": torch.tensor([24.0]),
+        }
+
+    mask_pix = torch.rand(B, K + 1, F_PIX, H * SCALE.height, W * SCALE.width, device=device)
+    batch = {
+        "video_latents": latents(),
+        "driving_latents": latents(),
+        "char_masks": {"mask": (mask_pix > 0.5).float()},
+        "conditions": {
+            "video_prompt_embeds": torch.randn(B, CTX_LEN, INNER, device=device),
+            "audio_prompt_embeds": torch.randn(B, CTX_LEN, INNER, device=device),
+            "prompt_attention_mask": torch.ones(B, CTX_LEN, device=device),
+        },
+    }
+    inputs = strategy.prepare_training_inputs(batch, UniformTimestepSampler(0.0, 1.0))
+    assert inputs.video.cond_channels.shape == (B, 2 * N, MASK_CH)
+    model = _tiny_model(MASK_CH, device, dtype)
+    with torch.inference_mode():
+        vpred, _ = model(video=inputs.video, audio=None, perturbations=None)
+    loss = strategy.compute_loss(vpred, None, inputs)
+    assert loss.shape == (B,)
+    assert torch.isfinite(loss).all()
+    cc_shape = tuple(inputs.video.cond_channels.shape)
+    print(f"[OK] Phase 3 training step: cond_channels {cc_shape}, loss {loss.item():.4f}")
+
+    # Phase 4: inference conditioning assembly.
+    masks = batch["char_masks"]["mask"]
+    conds = build_scail_conditionings(driving, masks, mode=DrivingMode.ANIMATION)
+    assert len(conds) == 2
+    print("[OK] Phase 4 build_scail_conditionings: [driving, mask_channels]")
+
+    print("\nAll SCAIL-2 checks passed on", device)
+
+
+if __name__ == "__main__":
+    main()

packages/ltx-core/README.md

@@ -8,9 +8,10 @@ The foundational library for the LTX-2 Audio-Video generation model. This packag
 - **`conditioning/`**: Tools for preparing latent states and applying conditioning (image, video, keyframes)
 - **`guidance/`**: Perturbation system for fine-grained control over attention mechanisms
 - **`loader/`**: Utilities for loading weights from `.safetensors`, fusing LoRAs, and managing memory
+- **`block_streaming/`**: Memory-efficient inference that streams transformer blocks through the GPU one at a time (from pinned CPU buffers or directly from disk)
 - **`model/`**: PyTorch implementations of the LTX-2 Transformer, Video VAE, Audio VAE, Vocoder and Upscaler
 - **`text_encoders/gemma`**: Gemma text encoder implementation with tokenizers, feature extractors, and separate encoders for audio-video and video-only generation
-- **`quantization/`**: FP8 quantization backends (FP8-TensorRT-LLM scaled MM, FP8 cast) for reduced memory footprint.
+- **`quantization/`**: FP8 quantization backends (FP8 scaled MM, FP8 cast) for reduced memory footprint.
 
 ## 🚀 Quick Start
 
@@ -55,6 +56,7 @@ pip install -e packages/ltx-core
 
 - **Loader** ([`loader/`](src/ltx_core/loader/)): Model loading from `.safetensors`, LoRA fusion, weight remapping, and memory management
 - **Quantization** ([`quantization/`](src/ltx_core/quantization/)): FP8 quantization backends for reduced memory footprint and faster inference
+- **Block Streaming** ([`block_streaming/`](src/ltx_core/block_streaming/)): Streams transformer blocks through the GPU one block at a time, so the full model runs on machines without enough memory to hold all its weights at once
 
 ### Loader
 
@@ -77,13 +79,17 @@ model = builder.build(device=torch.device("cuda"))
 Use the `.lora()` method to attach one or more LoRA adapters before calling `.build()`:
 
 ```python
+from ltx_core.loader import SDOps
+
+lora_sd_ops = SDOps(name="identity").with_matching()  # or a model-specific key-renaming SDOps
+
 builder = (
     SingleGPUModelBuilder(
         model_class_configurator=MyModelConfigurator,
         model_path="/path/to/model.safetensors",
     )
-    .lora("/path/to/lora_a.safetensors", strength=0.8)
+    .lora("/path/to/lora_a.safetensors", 0.8, lora_sd_ops)
-    .lora("/path/to/lora_b.safetensors", strength=0.5)
+    .lora("/path/to/lora_b.safetensors", 0.5, lora_sd_ops)
 )
 model = builder.build(device=torch.device("cuda"))
 ```
@@ -103,7 +109,7 @@ builder = SingleGPUModelBuilder(
     model_class_configurator=MyModelConfigurator,
     model_path="/path/to/model.safetensors",
     lora_load_device=torch.device("cuda"),
-).lora("/path/to/lora.safetensors", strength=1.0)
+).lora("/path/to/lora.safetensors", 1.0, lora_sd_ops)
 
 model = builder.build(device=torch.device("cuda"))
 ```
@@ -112,48 +118,33 @@ model = builder.build(device=torch.device("cuda"))
 
 The `quantization/` module provides FP8 quantization support for the LTX-2 transformer, significantly reducing memory usage while maintaining quality. Two backends are available:
 
-#### FP8 Scaled MM (TensorRT-LLM)
+#### FP8 Scaled MM
 
-Uses NVIDIA TensorRT-LLM's `cublas_scaled_mm` for efficient FP8 matrix multiplication. Weights are stored in FP8 format with per-tensor scaling, and inputs are quantized dynamically (or statically with calibration data).
+Uses PyTorch's `torch._scaled_mm` for efficient FP8 matrix multiplication. Weights are stored in FP8 format with per-tensor scaling, and inputs are quantized dynamically.
-
-**Requirements**: `uv sync --frozen --extra fp8-trtllm`
 
 **Usage with QuantizationPolicy:**
 
 ```python
-from ltx_core.quantization import QuantizationPolicy
+from ltx_core.quantization.fp8_scaled_mm import build_policy as build_fp8_scaled_mm_policy
 
-# Dynamic input quantization (no calibration needed)
+# Discovers the layer set from the checkpoint's .weight_scale tensors
-policy = QuantizationPolicy.fp8_scaled_mm()
+policy = build_fp8_scaled_mm_policy("/path/to/checkpoint.safetensors")
-
-# Static input quantization with calibration file
-policy = QuantizationPolicy.fp8_scaled_mm(calibration_amax_path="/path/to/amax.json")
 ```
 
-The policy provides `sd_ops` and `module_ops` that can be passed to the model builder:
+The policy carries `sd_ops`, `module_ops`, and `fuse_rule` that are passed to the model builder:
 
 ```python
+import torch
 from ltx_core.loader import SingleGPUModelBuilder
 
 builder = SingleGPUModelBuilder(
-    model=model,
+    model_class_configurator=MyModelConfigurator,
-    device=device,
+    model_path="/path/to/checkpoint.safetensors",
-    sd_ops=policy.sd_ops,
+    model_sd_ops=policy.sd_ops,
     module_ops=policy.module_ops,
+    fuse_rule=policy.fuse_rule,
 )
-builder.load(checkpoint_path)
+model = builder.build(device=torch.device("cuda"))
-```
-
-**Calibration File Format** (for static input quantization):
-
-```json
-{
-  "amax_values": {
-    "transformer_blocks.0.attn.to_q.input_quantizer": 12.5,
-    "transformer_blocks.0.attn.to_k.input_quantizer": 8.3,
-    ...
-  }
-}
 ```
 
 #### FP8 Cast
@@ -161,7 +152,42 @@ builder.load(checkpoint_path)
 A simpler approach that casts weights to FP8 for storage and upcasts during inference:
 
 ```python
-policy = QuantizationPolicy.fp8_cast()
+from ltx_core.quantization.fp8_cast import build_policy as build_fp8_cast_policy
+
+policy = build_fp8_cast_policy("/path/to/checkpoint.safetensors")
+```
+
+### Block Streaming
+
+The `block_streaming/` module ([`src/ltx_core/block_streaming/`](src/ltx_core/block_streaming/)) lets the full model run on machines that lack the memory to hold all of its weights at once. It streams the transformer's blocks through a small rolling set of GPU buffers, loading each block's weights just before it runs and recycling them afterwards, so only a few blocks are resident on the GPU at any moment. Construct it with `StreamingModelBuilder`, which returns a `BlockStreamingWrapper` -- an `nn.Module` drop-in for the wrapped model.
+
+#### Strategies
+
+The strategy is chosen automatically from `cpu_slots_count` relative to the number of blocks:
+
+- **RAM streaming** (default, `cpu_slots_count` omitted or `>= num_blocks`): all blocks are pre-loaded into pinned CPU buffers (with LoRA fusion) at build time, then copied to the GPU on demand. Fast; higher CPU memory.
+- **Disk streaming** (`cpu_slots_count < num_blocks`): blocks are read from the `.safetensors` file on demand on a background worker thread. Slower; lowest CPU memory.
+
+#### Basic usage
+
+```python
+import torch
+from ltx_core.block_streaming import StreamingModelBuilder
+
+builder = StreamingModelBuilder(
+    model_class_configurator=MyModelConfigurator,
+    model_path="/path/to/model.safetensors",
+    blocks_attr="transformer_blocks",      # dotted path to the nn.ModuleList
+    blocks_prefix="transformer_blocks",    # state-dict key prefix for block weights
+)
+
+# Omit cpu_slots_count for RAM streaming; pass a value < num_blocks for disk streaming.
+model = builder.build(
+    device=torch.device("cuda"),
+    dtype=torch.bfloat16,
+    cpu_slots_count=4,
+    gpu_slots_count=2,
+)
 ```
 
 For complete, production-ready pipeline implementations that combine these building blocks, see the [`ltx-pipelines`](../ltx-pipelines/) package.

packages/ltx-core/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "ltx-core"
-version = "1.1.2"
+version = "1.1.7"
 description = "Core implementation of Lightricks' LTX-2 model"
 readme = "README.md"
 requires-python = ">=3.10"
@@ -13,38 +13,14 @@ dependencies = [
     "safetensors",
     "accelerate",
     "scipy>=1.14",
+    # Apple Silicon only: Apple's fused MPSGraph SDPA, the AUTOMATIC attention
+    # backend on MPS. The marker installs it on Apple Silicon and prunes it
+    # everywhere else (Linux/CUDA), so it is a hard requirement exactly where it
+    # is the only viable attention kernel. Requires torch>=2.11 (within the
+    # torch~=2.7 floor); the resolver forks torch to >=2.11 on macOS.
+    "mps-sdpa>=0.2.0; sys_platform == 'darwin' and platform_machine == 'arm64'",
 ]
 
-[project.optional-dependencies]
-xformers = ["xformers"]
-fp8-trtllm = [
-    "tensorrt-llm==1.0.0",
-    "onnx>=1.16.0,<1.20.0",
-    "openmpi",
-]
-
-[tool.uv]
-conflicts = [
-  [
-    { extra = "xformers" },
-    { extra = "fp8-trtllm" },
-  ],
-]
-
-[tool.uv.sources]
-xformers = { index = "pytorch" }
-tensorrt-llm = { index = "nvidia" }
-
-[[tool.uv.index]]
-name = "pytorch"
-url = "https://download.pytorch.org/whl/cu129"
-explicit = true
-
-[[tool.uv.index]]
-name = "nvidia"
-url = "https://pypi.nvidia.com/"
-explicit = true
-
 [build-system]
 requires = ["uv_build>=0.9.8,<0.10.0"]
 build-backend = "uv_build"

packages/ltx-core/src/ltx_core/batch_split.py

@@ -25,8 +25,12 @@ from ltx_core.model.transformer.modality import Modality
 
 def _split_perturbations(config: BatchedPerturbationConfig, sizes: list[int]) -> list[BatchedPerturbationConfig]:
     """Split a ``BatchedPerturbationConfig`` along the batch dimension."""
-    it = iter(config.perturbations)
+    chunks = []
-    return [BatchedPerturbationConfig([next(it) for _ in range(s)]) for s in sizes]
+    offset = 0
+    for size in sizes:
+        chunks.append(config.batch_slice(offset, offset + size))
+        offset += size
+    return chunks
 
 
 def _merge_tensors(tensors: list[torch.Tensor | None]) -> torch.Tensor | None:
@@ -54,6 +58,10 @@ class BatchSplitAdapter(nn.Module):
         self._model = model
         self._max_batch_size = max_batch_size
 
+    @property
+    def num_blocks(self) -> int:
+        return self._model.num_blocks
+
     def _get_chunk_sizes(self, batch_size: int) -> list[int]:
         full, remainder = divmod(batch_size, self._max_batch_size)
         sizes = [self._max_batch_size] * full
@@ -65,7 +73,7 @@ class BatchSplitAdapter(nn.Module):
         self,
         video: Modality | None,
         audio: Modality | None,
-        perturbations: BatchedPerturbationConfig,
+        perturbations: BatchedPerturbationConfig | None,
     ) -> tuple[torch.Tensor | None, torch.Tensor | None]:
         batch_size = (video or audio).latent.shape[0]
 
@@ -77,7 +85,9 @@ class BatchSplitAdapter(nn.Module):
 
         v_chunks = video.split(sizes) if video is not None else [None] * n
         a_chunks = audio.split(sizes) if audio is not None else [None] * n
-        p_chunks = _split_perturbations(perturbations, sizes)
+        # A None config means "perturb nothing"; forward it per chunk so the inner model
+        # builds a per-chunk all-keep mask (splitting None has nothing to slice).
+        p_chunks = _split_perturbations(perturbations, sizes) if perturbations is not None else [None] * n
 
         chunk_results = [
             self._model(video=vc, audio=ac, perturbations=pc)

packages/ltx-core/src/ltx_core/block_streaming/__init__.py

@@ -5,8 +5,9 @@ CPU-to-GPU copies, caching, and stream synchronization.  Two weight
 source strategies are available:
 - **RAM streaming** (default): all blocks pre-loaded into pinned CPU
   buffers with LoRA fusion at build time.  Fast, higher CPU memory.
-- **Disk streaming** (``cpu_slots < num_blocks``): blocks read from
+- **Disk streaming** (``cpu_slots < blocks_number``): blocks are read from
-  disk on demand with FIFO eviction.  Slower, lower CPU memory.
+  disk on demand by a :class:`DiskWeightSource`, on a background worker
+  thread.  Slower, lower CPU memory.
 """
 
 from ltx_core.block_streaming.builder import DISK_CPU_SLOTS, StreamingModelBuilder

packages/ltx-core/src/ltx_core/block_streaming/block_fetcher.py

@@ -0,0 +1,83 @@
+"""BlockFetcher: async disk reads on a worker thread."""
+
+from __future__ import annotations
+
+import logging
+import queue
+import threading
+from dataclasses import dataclass
+
+import torch
+
+from ltx_core.block_streaming.disk import DiskBlockReader
+
+logger = logging.getLogger(__name__)
+
+Buffer = dict[str, torch.Tensor]
+
+
+@dataclass(slots=True)
+class FetchHandle:
+    """Caller-facing handle for an outstanding read returned by :meth:`BlockFetcher.submit`.
+    Carries only what the caller needs: a completion event the worker sets and the
+    read's error. The worker updates these once the read finishes.
+    """
+
+    _done: threading.Event
+    _error: BaseException | None = None
+
+    def wait(self) -> BaseException | None:
+        """Block until the read finishes; return its error, or ``None`` on success."""
+        self._done.wait()
+        return self._error
+
+
+@dataclass(slots=True)
+class _ReadRequest:
+    """One outstanding read, internal to :class:`BlockFetcher`.
+    The fetcher's worker reads block ``idx`` into the caller-carved ``buffer`` and
+    updates ``handle`` (its error, then its event) once the read has finished.
+    """
+
+    idx: int
+    buffer: Buffer
+    handle: FetchHandle
+
+
+class BlockFetcher:
+    """Fills caller-supplied buffers on a worker thread."""
+
+    def __init__(self, reader: DiskBlockReader) -> None:
+        self._reader = reader
+        self._request_queue: queue.SimpleQueue[_ReadRequest | None] = queue.SimpleQueue()
+        self._worker = threading.Thread(target=self._run, name="BlockFetcher-IO", daemon=True)
+        self._worker.start()
+
+    def submit(self, idx: int, buffer: Buffer) -> FetchHandle:
+        """Enqueue a read of block *idx* into the caller-carved *buffer*, return its handle."""
+        handle = FetchHandle(_done=threading.Event())
+        request = _ReadRequest(idx=idx, buffer=buffer, handle=handle)
+        self._request_queue.put(request)
+        return handle
+
+    def cleanup(self) -> None:
+        """Drain pending reads, join the worker, close the reader."""
+        self._request_queue.put(None)
+        self._worker.join()
+        self._reader.cleanup()
+
+    def _run(self) -> None:
+        # Pinned buffers are allocated under the caller's inference_mode, so
+        # in-place copy_ from this thread requires inference_mode here too.
+        with torch.inference_mode():
+            while True:
+                request = self._request_queue.get()
+                if request is None:
+                    return
+
+                try:
+                    self._reader.read_into(request.buffer, request.idx)
+                except Exception as exc:
+                    logger.exception("BlockFetcher: fetch failed for item %d", request.idx)
+                    request.handle._error = exc
+                request.handle._done.set()

packages/ltx-core/src/ltx_core/block_streaming/builder.py

@@ -2,45 +2,63 @@
 
 from __future__ import annotations
 
+import copy
 import logging
-from collections.abc import Callable
+from dataclasses import replace
-from dataclasses import dataclass, field, replace
+from typing import TYPE_CHECKING, Final, Generic
-from typing import Generic
 
+import safetensors
 import torch
 from torch import nn
 
+from ltx_core.block_streaming import utils as bs_utils
+from ltx_core.block_streaming.block_fetcher import BlockFetcher
 from ltx_core.block_streaming.disk import DiskBlockReader, DiskTensorReader, LoraSource
-from ltx_core.block_streaming.pool import BlockLayout, WeightPool
+from ltx_core.block_streaming.pool import BufferPool
 from ltx_core.block_streaming.provider import WeightsProvider
-from ltx_core.block_streaming.source import DiskWeightSource, PinnedWeightSource, WeightSource
+from ltx_core.block_streaming.source import DiskWeightSource, PinnedBlock, PinnedWeightSource, WeightSource
-from ltx_core.block_streaming.utils import build_pool_layout, resolve_attr
+from ltx_core.block_streaming.stream_sync import create_stream_sync
+from ltx_core.block_streaming.utils import (
+    carve_buffer,
+    derive_layout,
+    layout_nbytes,
+    make_block_key,
+    resolve_attr,
+)
 from ltx_core.block_streaming.wrapper import BlockStreamingWrapper
-from ltx_core.loader.fuse_loras import apply_loras
+from ltx_core.devices import synchronize_device
+from ltx_core.loader.fuse_loras import FuseRule, bf16_fuse_rule, fuse_lora_weights
 from ltx_core.loader.helpers import create_meta_model, load_state_dict, read_model_config
 from ltx_core.loader.module_ops import ModuleOps
 from ltx_core.loader.primitives import (
     LoraPathStrengthAndSDOps,
     LoraStateDictWithStrength,
     ModelBuilderProtocol,
+    StateDict,
     StateDictLoader,
+    TensorLayout,
 )
 from ltx_core.loader.registry import DummyRegistry, Registry
 from ltx_core.loader.sd_ops import SDOps
 from ltx_core.loader.sft_loader import SafetensorsModelStateDictLoader
 from ltx_core.model.model_protocol import ModelConfigurator, ModelType
 
+if TYPE_CHECKING:
+    from typing_extensions import Self
+
 logger = logging.getLogger(__name__)
 
 DISK_CPU_SLOTS = 2
 _DEFAULT_GPU_SLOTS = 2
+_PREFETCH_DEPTH = 2
 
 
-@dataclass(frozen=True)
 class StreamingModelBuilder(Generic[ModelType], ModelBuilderProtocol[ModelType]):
     """Immutable builder for :class:`BlockStreamingWrapper`.
     Reads block weights from safetensors on demand.  ``cpu_slots`` and
     ``gpu_slots`` control the memory/speed trade-off (see :meth:`build`).
+    The builder is immutable (``with_*`` return modified copies) and exposes
+    its state via read-only properties backed by private attributes.
     Args:
         model_class_configurator: Creates the model from a config dict.
         model_path: One or more ``.safetensors`` checkpoint paths.
@@ -49,38 +67,123 @@ class StreamingModelBuilder(Generic[ModelType], ModelBuilderProtocol[ModelType])
         loras: LoRA adapters fused into weights at load time.
         model_loader: Strategy for reading checkpoint metadata.
         registry: Shared cache for loaded state dicts.
+        fuse_rule: Per-policy LoRA merge rule. Defaults to ``bf16_fuse_rule``;
+            use ``fp8_cast_fuse_rule`` for fp8_cast streaming so the pinned
+            buffers receive correctly-quantized weights.
         blocks_attr: Dotted path to the ``nn.ModuleList`` (e.g.
-            ``"velocity_model.transformer_blocks"``).
+            ``"transformer_blocks"``).
         blocks_prefix: State-dict key prefix for block weights
             (e.g. ``"transformer_blocks"``).
-        state_dict_prefix: Key prefix for non-block weights
+        cpu_slots_count: Default number of pinned CPU buffer slots used by
-            (e.g. ``"velocity_model."``).
+            :meth:`build` when it is not given an explicit ``cpu_slots_count``.
-        model_wrapper: Optional callable wrapping the model
+            ``None`` = RAM streaming (all blocks pinned); a small value (e.g.
-            (e.g. ``X0Model``).
+            ``DISK_CPU_SLOTS``) selects disk streaming. Lets a builder fully
+            encode its offload behaviour so callers need not re-specify it.
     """
 
-    model_class_configurator: type[ModelConfigurator[ModelType]]
+    def __init__(  # noqa: PLR0913
-    model_path: str | tuple[str, ...]
+        self,
-    model_sd_ops: SDOps | None = None
+        model_class_configurator: type[ModelConfigurator[ModelType]],
-    module_ops: tuple[ModuleOps, ...] = field(default_factory=tuple)
+        model_path: str | tuple[str, ...],
-    loras: tuple[LoraPathStrengthAndSDOps, ...] = field(default_factory=tuple)
+        model_sd_ops: SDOps | None = None,
-    model_loader: StateDictLoader = field(default_factory=SafetensorsModelStateDictLoader)
+        module_ops: tuple[ModuleOps, ...] = (),
-    registry: Registry = field(default_factory=DummyRegistry)
+        loras: tuple[LoraPathStrengthAndSDOps, ...] = (),
+        model_loader: StateDictLoader | None = None,
+        registry: Registry | None = None,
+        fuse_rule: FuseRule = bf16_fuse_rule,
+        blocks_attr: str = "",
+        blocks_prefix: str = "",
+        cpu_slots_count: int | None = None,
+    ) -> None:
+        # Read-only: typed with the covariant ModelType, so it must not be a mutable attribute.
+        self._model_class_configurator: Final = model_class_configurator
+        self._model_path = model_path
+        self._model_sd_ops = model_sd_ops
+        self._module_ops = module_ops
+        self._loras = loras
+        self._model_loader = model_loader if model_loader is not None else SafetensorsModelStateDictLoader()
+        self._registry = registry if registry is not None else DummyRegistry()
+        self._fuse_rule = fuse_rule
+        self._blocks_attr = blocks_attr
+        self._blocks_prefix = blocks_prefix
+        self._cpu_slots_count = cpu_slots_count
 
-    # Streaming-specific
+    @property
-    blocks_attr: str = ""
+    def model_class_configurator(self) -> type[ModelConfigurator[ModelType]]:
-    blocks_prefix: str = ""
+        return self._model_class_configurator
-    state_dict_prefix: str = ""
-    model_wrapper: Callable[[ModelType], nn.Module] | None = None
 
-    def with_sd_ops(self, sd_ops: SDOps | None) -> StreamingModelBuilder:
+    @property
-        return replace(self, model_sd_ops=sd_ops)
+    def model_path(self) -> str | tuple[str, ...]:
+        return self._model_path
 
-    def with_module_ops(self, module_ops: tuple[ModuleOps, ...]) -> StreamingModelBuilder:
+    @property
-        return replace(self, module_ops=module_ops)
+    def checkpoint(self) -> str | tuple[str, ...]:
+        return self._model_path
 
-    def with_loras(self, loras: tuple[LoraPathStrengthAndSDOps, ...]) -> StreamingModelBuilder:
+    @property
-        return replace(self, loras=loras)
+    def model_sd_ops(self) -> SDOps | None:
+        return self._model_sd_ops
+
+    @property
+    def module_ops(self) -> tuple[ModuleOps, ...]:
+        return self._module_ops
+
+    @property
+    def loras(self) -> tuple[LoraPathStrengthAndSDOps, ...]:
+        return self._loras
+
+    @property
+    def model_loader(self) -> StateDictLoader:
+        return self._model_loader
+
+    @property
+    def registry(self) -> Registry:
+        return self._registry
+
+    @property
+    def fuse_rule(self) -> FuseRule:
+        return self._fuse_rule
+
+    @property
+    def blocks_attr(self) -> str:
+        return self._blocks_attr
+
+    @property
+    def blocks_prefix(self) -> str:
+        return self._blocks_prefix
+
+    @property
+    def cpu_slots_count(self) -> int | None:
+        return self._cpu_slots_count
+
+    def with_sd_ops(self, sd_ops: SDOps | None) -> Self:
+        clone = copy.copy(self)
+        clone._model_sd_ops = sd_ops
+        return clone
+
+    def with_module_ops(self, module_ops: tuple[ModuleOps, ...]) -> Self:
+        clone = copy.copy(self)
+        clone._module_ops = module_ops
+        return clone
+
+    def with_loras(self, loras: tuple[LoraPathStrengthAndSDOps, ...]) -> Self:
+        clone = copy.copy(self)
+        clone._loras = loras
+        return clone
+
+    def with_registry(self, registry: Registry) -> Self:
+        clone = copy.copy(self)
+        clone._registry = registry
+        return clone
+
+    def with_lora_load_device(self, device: torch.device) -> Self:
+        # Streaming fuses LoRAs into pinned CPU buffers; no other staging device is meaningful.
+        raise NotImplementedError("StreamingModelBuilder loads LoRA weights on CPU only.")
+
+    def with_fuse_rule(self, fuse_rule: FuseRule) -> Self:
+        clone = copy.copy(self)
+        clone._fuse_rule = fuse_rule
+        return clone
 
     def model_config(self) -> dict:
         """Read model configuration from the checkpoint metadata."""
@@ -92,137 +195,290 @@ class StreamingModelBuilder(Generic[ModelType], ModelBuilderProtocol[ModelType])
 
     def build(
         self,
-        target_device: torch.device,
+        device: torch.device | None = None,
-        dtype: torch.dtype,
+        dtype: torch.dtype | None = None,
         cpu_slots_count: int | None = None,
         gpu_slots_count: int | None = None,
         **_kwargs: object,
     ) -> BlockStreamingWrapper:
         """Build and return a ready-to-use :class:`BlockStreamingWrapper`.
         Args:
-            target_device: GPU device for compute.
+            device: GPU device for compute. ``None`` defaults to ``cuda``.
-            dtype: Weight dtype (e.g. ``torch.bfloat16``).
+            dtype: Weight dtype (e.g. ``torch.bfloat16``). Required.
-            cpu_slots_count: Number of pinned CPU buffer slots.
+            cpu_slots_count: Number of pinned CPU buffer slots. ``None`` falls
-                ``None`` = RAM streaming (all blocks pre-loaded with LoRA fusion).
+                back to the builder's configured ``cpu_slots_count``, and if that
+                is also ``None``, to RAM streaming (all blocks pre-loaded with
+                LoRA fusion).
             gpu_slots_count: Number of GPU buffer slots.
                 ``None`` = ``_DEFAULT_GPU_SLOTS`` (2).
         """
         if not self.blocks_prefix:
             raise ValueError("blocks_prefix must be non-empty for streaming")
+        if dtype is None:
+            raise ValueError("StreamingModelBuilder.build requires an explicit dtype")
+        device = device if device is not None else torch.device("cuda")
 
-        # 1. Create meta model (no weights allocated).
         config = read_model_config(self.model_path, self.model_loader)
         meta_model: nn.Module = create_meta_model(self.model_class_configurator, config, self.module_ops)
-        if self.model_wrapper is not None:
-            meta_model = self.model_wrapper(meta_model)
         meta_model.eval()
 
         blocks = resolve_attr(meta_model, self.blocks_attr)
-        layout = build_pool_layout(blocks[0], dtype)
 
-        # 2. Determine slot counts.
+        checkpoint_paths = list(self.model_path) if isinstance(self.model_path, tuple) else [self.model_path]
+        block_key_map, non_block_keys = _scan_checkpoint_keys(checkpoint_paths, self.model_sd_ops, self.blocks_prefix)
+        expected_indices = set(range(len(blocks)))
+        if set(block_key_map) != expected_indices:
+            missing = sorted(expected_indices - set(block_key_map))
+            extra = sorted(set(block_key_map) - expected_indices)
+            raise ValueError(
+                f"Block weights under prefix '{self.blocks_prefix}.' do not match the {len(blocks)} model blocks: "
+                f"missing indices {missing}, unexpected indices {extra}"
+            )
+
+        cpu_slots_count = cpu_slots_count if cpu_slots_count is not None else self._cpu_slots_count
         cpu_slots_count = cpu_slots_count if cpu_slots_count is not None else len(blocks)
         gpu_slots_count = gpu_slots_count if gpu_slots_count is not None else _DEFAULT_GPU_SLOTS
 
-        # 3. Build source and load non-block weights.
         if cpu_slots_count >= len(blocks):
-            source, lora_sources = self._build_pinned_source(meta_model, target_device, dtype, cpu_slots_count)
+            lora_sd_and_strengths = self._load_lora_sds()
-        else:
+            source, lora_sources = self._build_pinned_source(
-            source, lora_sources = self._build_disk_source(meta_model, layout, target_device, dtype, cpu_slots_count)
+                blocks, dtype, cpu_slots_count, block_key_map, lora_sd_and_strengths
-
+            )
-        # 4. Create provider and wrapper.
+            non_block_loras = lora_sd_and_strengths
-        copy_stream = torch.cuda.Stream(device=target_device)
+        else:
-        gpu_pool = WeightPool(
+            reader = DiskTensorReader(checkpoint_paths)
-            layout, gpu_slots_count, target_device, reuse_barrier=lambda event: copy_stream.wait_event(event)
+            source, lora_sources = self._build_disk_source(
+                blocks, dtype, cpu_slots_count, reader, block_key_map, prefetch_depth=_PREFETCH_DEPTH
+            )
+            non_block_loras = [src.as_state_dict_with_strength() for src in lora_sources]
+
+        self._load_non_block_weights(meta_model, non_block_keys, device, dtype, non_block_loras)
+
+        sync = create_stream_sync(device)
+        gpu_pool = BufferPool(source.slot_nbytes, gpu_slots_count, device, reuse_barrier=sync.reuse_barrier)
+        provider = WeightsProvider(
+            gpu_pool,
+            sync,
+            device,
+            source,
+            lora_sources,
+            self.blocks_prefix,
+            fuse_rule=self.fuse_rule,
         )
-        provider = WeightsProvider(gpu_pool, copy_stream, target_device, source, lora_sources, self.blocks_prefix)
         return BlockStreamingWrapper(
             model=meta_model,
             blocks=blocks,
             provider=provider,
-            target_device=target_device,
+            target_device=device,
         )
 
+    def _load_lora_sds(self) -> list[LoraStateDictWithStrength]:
+        """Load each configured LoRA into a state dict for fusion (pinned path)."""
+        return [
+            LoraStateDictWithStrength(
+                load_state_dict([lora.path], self.model_loader, self.registry, torch.device("cpu"), lora.sd_ops),
+                lora.strength,
+            )
+            for lora in self.loras
+        ]
+
+    def _filtered_sd_ops(self, name_suffix: str, allowed_model_keys: frozenset[str]) -> SDOps:
+        """``model_sd_ops`` restricted to *allowed_model_keys* (post-rename keys).
+        The loader skips keys filtered to None before reading them, so a restricted
+        load never materializes the excluded partition. The distinct ``name`` avoids
+        a registry cache-id collision with the other partition.
+        """
+        base = self.model_sd_ops if self.model_sd_ops is not None else SDOps("streaming").with_matching()
+        allowed = allowed_model_keys if base.allowed_keys is None else (allowed_model_keys & base.allowed_keys)
+        return replace(base, name=f"{base.name}__{name_suffix}", allowed_keys=allowed)
+
     def _build_pinned_source(
         self,
-        meta_model: nn.Module,
+        blocks: nn.ModuleList,
-        target_device: torch.device,
         dtype: torch.dtype,
         cpu_slots_count: int,
+        block_key_map: dict[int, list[tuple[str, str]]],
+        lora_sd_and_strengths: list[LoraStateDictWithStrength],
     ) -> tuple[WeightSource, list[LoraSource]]:
-        """Pre-load all blocks into pinned CPU buffers with LoRA fusion."""
+        """Pre-load each block into its own contiguous pinned CPU buffer with LoRA fusion."""
-        model_sd = load_state_dict(
+        for block_idx in block_key_map:
-            self.model_path, self.model_loader, self.registry, torch.device("cpu"), self.model_sd_ops
+            if block_idx >= cpu_slots_count:
+                raise ValueError(
+                    f"Pinned source requires one CPU slot per block; "
+                    f"got block index {block_idx} with only {cpu_slots_count} slots."
                 )
 
-        if self.loras:
+        # One contiguous pinned buffer per block, carved into per-param views. The
-            lora_sds = [
+        # views (flattened by full key) are filled in place; the source then keeps
-                load_state_dict([lora.path], self.model_loader, self.registry, torch.device("cpu"), lora.sd_ops)
+        # only the contiguous buffer and the layout to re-carve it on read.
-                for lora in self.loras
+        pinned_buffers: dict[int, torch.Tensor] = {}
-            ]
+        block_layouts: dict[int, TensorLayout] = {}
-            lora_sd_and_strengths = [
+        fill_views: dict[str, torch.Tensor] = {}
-                LoraStateDictWithStrength(sd, lora.strength) for sd, lora in zip(lora_sds, self.loras, strict=True)
+        for block_idx, entries in block_key_map.items():
-            ]
+            block_state = _block_state(blocks[block_idx])
-            model_sd = apply_loras(
+            layout = derive_layout({param_name: block_state[param_name] for _sft_key, param_name in entries}, dtype)
-                model_sd=model_sd,
+            buffer = bs_utils.alloc_buffer(layout_nbytes(layout), torch.device("cpu"), pin_memory=True)
-                lora_sd_and_strengths=lora_sd_and_strengths,
+            views = carve_buffer(buffer, layout)
-                dtype=dtype,
+            pinned_buffers[block_idx] = buffer
-                destination_sd=model_sd if isinstance(self.registry, DummyRegistry) else None,
+            block_layouts[block_idx] = layout
+            for param_name, view in views.items():
+                fill_views[make_block_key(self.blocks_prefix, block_idx, param_name)] = view
+
+        block_sd = load_state_dict(
+            self.model_path,
+            self.model_loader,
+            self.registry,
+            torch.device("cpu"),
+            self._filtered_sd_ops("blocks", frozenset(fill_views)),
         )
 
-        # Partition: non-block weights go to GPU, block weights go directly
+        should_sync = False
-        # to pinned buffers.  This avoids holding the full state dict and
+        for key, fused in fuse_lora_weights(
-        # pinned copies simultaneously.
+            block_sd, lora_sd_and_strengths, fuse_rule=self.fuse_rule, preserve_input_device=False
-        non_block_sd: dict[str, torch.Tensor] = {}
+        ):
-        block_tensors: dict[int, dict[str, torch.Tensor]] = {}
+            if key not in fill_views:
-        prefix_dot = self.blocks_prefix + "."
+                raise ValueError(f"Block-restricted load produced {key!r}, which is not a pinned block weight")
+            fill_views[key].copy_(fused, non_blocking=True)
+            block_sd.sd[key] = None
+            should_sync = True
+        if should_sync:
+            synchronize_device()
 
-        for key, tensor in model_sd.sd.items():
+        # Fill remaining pinned keys from the source state dict.
-            if key.startswith(prefix_dot):
+        for key, view in fill_views.items():
-                rest = key[len(prefix_dot) :]
+            if block_sd.sd[key] is None:
-                idx_str, _, param_name = rest.partition(".")
-                try:
-                    block_idx = int(idx_str)
-                except ValueError:
-                    non_block_sd[self.state_dict_prefix + key] = tensor.to(device=target_device, dtype=dtype)
                 continue
-                block_tensors.setdefault(block_idx, {})[param_name] = tensor
+            view.copy_(block_sd.sd[key])
-            else:
+            block_sd.sd[key] = None
-                non_block_sd[self.state_dict_prefix + key] = tensor.to(device=target_device, dtype=dtype)
-
-        meta_model.load_state_dict(non_block_sd, strict=False, assign=True)
-        del model_sd, non_block_sd
-
-        # Pin block weights one block at a time, freeing the source tensors as we go.
-        pinned: dict[int, dict[str, torch.Tensor]] = {}
-        for idx in range(cpu_slots_count):
-            src = block_tensors.pop(idx)
-            pinned[idx] = {name: tensor.to(dtype=dtype).pin_memory() for name, tensor in src.items()}
 
+        pinned = {idx: PinnedBlock(pinned_buffers[idx], block_layouts[idx]) for idx in pinned_buffers}
         return PinnedWeightSource(pinned), []
 
     def _build_disk_source(
         self,
-        meta_model: nn.Module,
+        blocks: nn.ModuleList,
-        layout: BlockLayout,
-        target_device: torch.device,
         dtype: torch.dtype,
         cpu_slots_count: int,
+        reader: DiskTensorReader,
+        block_key_map: dict[int, list[tuple[str, str]]],
+        prefetch_depth: int,
     ) -> tuple[WeightSource, list[LoraSource]]:
-        """Create a DiskWeightSource backed by a DiskBlockReader for lazy loading."""
+        """Create a DiskWeightSource backed by a DiskBlockReader.
-        lora_sources = [LoraSource(lora.path, lora.sd_ops, lora.strength) for lora in self.loras]
+        Pool slots are sized to the largest block and carved per block on read, so
-        checkpoint_paths = list(self.model_path) if isinstance(self.model_path, tuple) else [self.model_path]
+        heterogeneous blocks (e.g. layers with differing attention layouts) share
-        reader = DiskTensorReader(checkpoint_paths)
+        one pool. Pool capacity is ``cpu_slots_count + prefetch_depth`` so the
+        lookahead loop in ``DiskWeightSource.get`` never evicts its own target.
+        Layouts come from the meta model; assumes module_ops keep the meta param
+        dtype in sync with the post-sd_ops checkpoint dtype.
+        """
+        block_layouts = _block_layouts(blocks, block_key_map, dtype)
+        slot_nbytes = max(layout_nbytes(layout) for layout in block_layouts.values())
 
+        cpu_pool = BufferPool(
+            slot_nbytes,
+            cpu_slots_count + prefetch_depth,
+            torch.device("cpu"),
+            reuse_barrier=lambda event: event.synchronize(),
+            pin_memory=True,
+        )
+        block_reader = DiskBlockReader(
+            reader=reader,
+            block_key_map=block_key_map,
+            sd_ops=self.model_sd_ops,
+            blocks_prefix=self.blocks_prefix,
+        )
+        fetcher = BlockFetcher(block_reader)
+        source = DiskWeightSource(
+            cpu_pool,
+            fetcher,
+            block_layouts,
+            blocks_number=len(blocks),
+            prefetch_depth=prefetch_depth,
+        )
+        lora_sources = [LoraSource(lora.path, lora.sd_ops, lora.strength) for lora in self.loras]
+
+        return source, lora_sources
+
+    @torch.inference_mode()
+    def _load_non_block_weights(
+        self,
+        model: nn.Module,
+        non_block_keys: list[tuple[str, str]],
+        device: torch.device,
+        dtype: torch.dtype,
+        lora_sd_and_strengths: list[LoraStateDictWithStrength],
+    ) -> None:
+        """Load the non-block weights onto *device* and fuse LoRAs -- both paths.
+        Reads through the loader with ``model_sd_ops`` restricted to the
+        non-block keys, so ``sd_ops`` (incl. kv-ops such as Gemma's ``lm_head``
+        duplication) is applied exactly once and block tensors are never read.
+        """
+        non_block_sd_ops = self._filtered_sd_ops("non_block", frozenset(mk for _sft_key, mk in non_block_keys))
+        loaded = load_state_dict(self.model_path, self.model_loader, self.registry, device, non_block_sd_ops)
+        non_block_sd = {key: tensor.to(dtype=dtype) for key, tensor in loaded.sd.items()}
+
+        if lora_sd_and_strengths:
+            non_block_state = StateDict(sd=non_block_sd, device=device, size=0, dtype={dtype})
+            for key, fused in fuse_lora_weights(
+                non_block_state,
+                lora_sd_and_strengths,
+                fuse_rule=self.fuse_rule,
+                preserve_input_device=True,
+            ):
+                non_block_sd[key] = fused
+
+        model.load_state_dict(non_block_sd, strict=False, assign=True)
+
+
+def _block_state(block: nn.Module) -> dict[str, torch.Tensor]:
+    """Streamed-eligible tensors of a block: parameters then buffers.
+    Block streaming swaps both params and checkpoint-backed buffers (e.g. Gemma4's
+    per-layer ``layer_scalar``), so the layout, pinned packing, and meta-ordering
+    all consult parameters and buffers together. Non-checkpoint (computed) buffers
+    are harmless here -- only keys present in ``block_key_map`` are ever streamed.
+    """
+    return {**dict(block.named_parameters()), **dict(block.named_buffers())}
+
+
+def _block_layouts(
+    blocks: nn.ModuleList,
+    block_key_map: dict[int, list[tuple[str, str]]],
+    dtype: torch.dtype,
+) -> dict[int, TensorLayout]:
+    """Per-block layout of the streamed tensors, taken from the meta model.
+    Blocks may differ in shape and even in which tensors they have (e.g. Gemma4's
+    full-attention layers drop ``v_proj``), so each block gets its own layout in
+    ``block_key_map`` order. The pinned packing, the disk reader, and the GPU carve
+    all key off this same per-block layout, so the provider's contiguous H2D copy
+    is valid for any entry order (no cross-block ordering required).
+    """
+    layouts: dict[int, TensorLayout] = {}
+    for idx, entries in block_key_map.items():
+        state = _block_state(blocks[idx])
+        layouts[idx] = derive_layout({param_name: state[param_name] for _sft_key, param_name in entries}, dtype)
+    return layouts
+
+
+def _scan_checkpoint_keys(
+    checkpoint_paths: list[str],
+    sd_ops: SDOps | None,
+    blocks_prefix: str,
+) -> tuple[dict[int, list[tuple[str, str]]], list[tuple[str, str]]]:
+    """Partition checkpoint keys into per-block and non-block lists.
+    Opens the safetensors files for header-only key enumeration; no tensor data
+    is read.
+    """
     block_key_map: dict[int, list[tuple[str, str]]] = {}
     non_block_keys: list[tuple[str, str]] = []
-
+    prefix_dot = blocks_prefix + "."
-        for sft_key in reader.keys():  # noqa: SIM118
+    for path in checkpoint_paths:
-            model_key = self.model_sd_ops.apply_to_key(sft_key) if self.model_sd_ops else sft_key
+        with safetensors.safe_open(path, framework="pt", device="cpu") as handle:
+            for sft_key in handle.keys():  # noqa: SIM118
+                model_key = sd_ops.apply_to_key(sft_key) if sd_ops else sft_key
                 if model_key is None:
                     continue
-            if model_key.startswith(self.blocks_prefix + "."):
+                if model_key.startswith(prefix_dot):
-                rest = model_key[len(self.blocks_prefix) + 1 :]
+                    rest = model_key[len(prefix_dot) :]
                     idx_str, _, param_name = rest.partition(".")
                     try:
                         block_idx = int(idx_str)
@@ -232,74 +488,4 @@ class StreamingModelBuilder(Generic[ModelType], ModelBuilderProtocol[ModelType])
                     block_key_map.setdefault(block_idx, []).append((sft_key, param_name))
                 else:
                     non_block_keys.append((sft_key, model_key))
-
+    return block_key_map, non_block_keys
-        self._load_non_block_weights(
-            reader,
-            non_block_keys,
-            meta_model,
-            target_device,
-            dtype,
-            sd_ops=self.model_sd_ops,
-            key_prefix=self.state_dict_prefix,
-            lora_sources=lora_sources,
-            matmul_device=target_device,
-        )
-
-        cpu_pool = WeightPool(
-            layout,
-            cpu_slots_count,
-            torch.device("cpu"),
-            reuse_barrier=lambda event: event.synchronize(),
-            pin_memory=True,
-        )
-        block_reader = DiskBlockReader(reader=reader, block_key_map=block_key_map, dtype=dtype)
-        source = DiskWeightSource(cpu_pool, block_reader)
-        return source, lora_sources
-
-    # ------------------------------------------------------------------
-    # Helpers
-    # ------------------------------------------------------------------
-
-    @staticmethod
-    def _fuse_lora_delta(
-        model_key: str,
-        tensor: torch.Tensor,
-        lora_sources: list[LoraSource],
-        matmul_device: torch.device | None = None,
-    ) -> torch.Tensor:
-        """Add all matching LoRA deltas to *tensor* in-place."""
-        if not lora_sources or not model_key.endswith(".weight"):
-            return tensor
-        prefix = model_key[: -len(".weight")]
-        device = tensor.device if tensor.device.type == "cuda" else matmul_device
-        for source in lora_sources:
-            delta = source.get_delta(prefix, device=device)
-            if delta is not None:
-                tensor = tensor.add_(delta.to(device=tensor.device, dtype=tensor.dtype))
-        return tensor
-
-    @staticmethod
-    @torch.inference_mode()
-    def _load_non_block_weights(
-        reader: DiskTensorReader,
-        non_block_keys: list[tuple[str, str]],
-        model: nn.Module,
-        device: torch.device,
-        dtype: torch.dtype,
-        sd_ops: SDOps | None = None,
-        key_prefix: str = "",
-        lora_sources: list[LoraSource] | None = None,
-        matmul_device: torch.device | None = None,
-    ) -> None:
-        """Load non-block weights into *model* on *device*."""
-        state_dict: dict[str, torch.Tensor] = {}
-        sources = lora_sources or []
-        for sft_key, model_key in non_block_keys:
-            tensor = reader.get_tensor(sft_key).to(device=device, dtype=dtype)
-            tensor = StreamingModelBuilder._fuse_lora_delta(model_key, tensor, sources, matmul_device)
-            if sd_ops is not None:
-                for kv in sd_ops.apply_to_key_value(model_key, tensor):
-                    state_dict[key_prefix + kv.new_key] = kv.new_value
-                continue
-            state_dict[key_prefix + model_key] = tensor
-        model.load_state_dict(state_dict, strict=False, assign=True)

packages/ltx-core/src/ltx_core/block_streaming/disk.py

@@ -2,11 +2,23 @@
 
 from __future__ import annotations
 
+from collections.abc import Iterator
+
 import safetensors
 import torch
 
+from ltx_core.block_streaming.utils import allocate_layout_views, make_block_key
+from ltx_core.loader.fuse_loras import LoraProduct
+from ltx_core.loader.primitives import LoraStateDictWithStrength, StateDict
 from ltx_core.loader.sd_ops import SDOps
 
+_SAFETENSORS_DTYPE_TO_TORCH: dict[str, torch.dtype] = {
+    "F64": torch.float64,
+    "F32": torch.float32,
+    "F16": torch.float16,
+    "BF16": torch.bfloat16,
+}
+
 
 class DiskTensorReader:
     """Key-based tensor accessor over one or more safetensors files."""
@@ -21,9 +33,6 @@ class DiskTensorReader:
             for sft_key in handle.keys():  # noqa: SIM118
                 self._key_to_handle_idx[sft_key] = handle_idx
 
-    def keys(self) -> list[str]:
-        return list(self._key_to_handle_idx.keys())
-
     def get_tensor(self, key: str) -> torch.Tensor:
         return self._handles[self._key_to_handle_idx[key]].get_tensor(key)
 
@@ -31,49 +40,58 @@ class DiskTensorReader:
         self._handles.clear()
         self._key_to_handle_idx.clear()
 
+    def __contains__(self, key: str) -> bool:
+        return key in self._key_to_handle_idx
+
+    def __iter__(self) -> Iterator[str]:
+        return iter(self._key_to_handle_idx)
+
 
 class DiskBlockReader:
-    """Reads one block at a time from safetensors into provided buffers.
+    """Reads one block at a time from safetensors into provided buffers."""
-    Maps block indices to safetensors keys via a pre-computed key map.
-    """
 
     def __init__(
         self,
         reader: DiskTensorReader,
         block_key_map: dict[int, list[tuple[str, str]]],
-        dtype: torch.dtype,
+        sd_ops: SDOps | None = None,
+        blocks_prefix: str = "",
     ) -> None:
         self._reader = reader
         self._block_key_map = block_key_map
-        self._dtype = dtype
+        self._sd_ops = sd_ops
+        self._blocks_prefix = blocks_prefix
 
     def read_into(self, target: dict[str, torch.Tensor], block_idx: int) -> None:
+        block_prefix = make_block_key(self._blocks_prefix, block_idx, "")
         for sft_key, param_name in self._block_key_map[block_idx]:
             tensor = self._reader.get_tensor(sft_key)
-            if tensor.dtype != self._dtype:
+            if self._sd_ops is None:
-                tensor = tensor.to(self._dtype)
                 target[param_name].copy_(tensor)
+                continue
+            full_key = make_block_key(self._blocks_prefix, block_idx, param_name)
+            for result in self._sd_ops.apply_to_key_value(full_key, tensor):
+                if not result.new_key.startswith(block_prefix):
+                    raise ValueError(
+                        f"SDOps output key '{result.new_key}' is outside block {block_idx} "
+                        f"(expected prefix '{block_prefix}'); cannot route to a per-block buffer."
+                    )
+                target[result.new_key[len(block_prefix) :]].copy_(result.new_value)
 
     def cleanup(self) -> None:
         self._reader.close()
 
 
 class LoraSource:
-    """Pinned-memory cache of LoRA A/B matrices for on-the-fly fusion.
+    """Pinned-memory cache of matched LoRA A/B factors backed by a single buffer."""
-    At init, loads all matched A/B pairs into pinned CPU memory.
-    :meth:`get_delta` computes ``(B * strength) @ A`` on the given device.
-    """
 
     def __init__(self, path: str, sd_ops: SDOps | None, strength: float) -> None:
         self.strength = strength
-
-        # param_prefix -> (pinned_a, pinned_b)
         self._pinned_ab: dict[str, tuple[torch.Tensor, torch.Tensor]] = {}
 
         a_keys: dict[str, str] = {}
         b_keys: dict[str, str] = {}
         with safetensors.safe_open(path, framework="pt", device="cpu") as handle:
-            # First pass: build key map.
             for sft_key in handle.keys():  # noqa: SIM118
                 model_key = sd_ops.apply_to_key(sft_key) if sd_ops is not None else sft_key
                 if model_key is None:
@@ -83,24 +101,67 @@ class LoraSource:
                 elif model_key.endswith(".lora_B.weight"):
                     b_keys[model_key[: -len(".lora_B.weight")]] = sft_key
 
-            # Second pass: load and pin matched A+B pairs (orphans silently skipped).
+            matched_prefixes = list(a_keys.keys() & b_keys.keys())
-            for prefix in a_keys.keys() & b_keys.keys():
+
-                self._pinned_ab[prefix] = (
+            # Build the layout from safetensors header metadata only — no tensor data is read.
-                    handle.get_tensor(a_keys[prefix]).pin_memory(),
+            layout: dict[str, tuple[torch.Size, torch.dtype]] = {}
-                    handle.get_tensor(b_keys[prefix]).pin_memory(),
+            for prefix in matched_prefixes:
+                a_slice_view = handle.get_slice(a_keys[prefix])
+                b_slice_view = handle.get_slice(b_keys[prefix])
+                layout[f"{prefix}.A"] = (
+                    torch.Size(a_slice_view.get_shape()),
+                    _SAFETENSORS_DTYPE_TO_TORCH[a_slice_view.get_dtype()],
+                )
+                layout[f"{prefix}.B"] = (
+                    torch.Size(b_slice_view.get_shape()),
+                    _SAFETENSORS_DTYPE_TO_TORCH[b_slice_view.get_dtype()],
                 )
 
-    def get_delta(self, param_prefix: str, device: torch.device | None = None) -> torch.Tensor | None:
+            all_views = allocate_layout_views(layout, pin_memory=True)
-        """Return ``(B * strength) @ A`` for *param_prefix*, or ``None``."""
+
+            for prefix in matched_prefixes:
+                a_view = all_views[f"{prefix}.A"]
+                b_view = all_views[f"{prefix}.B"]
+                a_view.copy_(handle.get_tensor(a_keys[prefix]))
+                b_view.copy_(handle.get_tensor(b_keys[prefix]))
+                self._pinned_ab[prefix] = (a_view, b_view)
+
+    def as_state_dict_with_strength(self) -> LoraStateDictWithStrength:
+        """Return a :class:`LoraStateDictWithStrength` view of the pinned A/B factors.
+        Lets non-block fusion consume the already-loaded disk-streaming LoRA
+        without re-reading the safetensors file or re-applying ``sd_ops``.
+        """
+        sd: dict[str, torch.Tensor] = {}
+        for prefix, (a, b) in self._pinned_ab.items():
+            sd[f"{prefix}.lora_A.weight"] = a
+            sd[f"{prefix}.lora_B.weight"] = b
+        size = sum(t.numel() * t.element_size() for t in sd.values())
+        dtypes = {t.dtype for t in sd.values()}
+        return LoraStateDictWithStrength(
+            StateDict(sd=sd, device=torch.device("cpu"), size=size, dtype=dtypes),
+            self.strength,
+        )
+
+    def get_ab(
+        self,
+        param_prefix: str,
+        device: torch.device | None = None,
+        dtype: torch.dtype | None = None,
+    ) -> LoraProduct | None:
+        """Return the :class:`LoraProduct` for *param_prefix*, or ``None``."""
         pair = self._pinned_ab.get(param_prefix)
         if pair is None:
             return None
         a, b = pair
-        if device is not None and device.type == "cuda":
+        # Move A/B to a GPU-class target (CUDA/MPS) so the B@A aggregation runs on
-            a = a.to(device=device)
+        # the device; on a CPU target they stay put. non_blocking only helps CUDA.
-            b = b.to(device=device)
+        if device is not None and device.type in ("cuda", "mps"):
-        delta = torch.matmul(b * self.strength, a)
+            a = a.to(device=device, non_blocking=device.type == "cuda")
-        return delta
+            b = b.to(device=device, non_blocking=device.type == "cuda")
+        if dtype is not None:
+            a = a.to(dtype=dtype)
+            b = b.to(dtype=dtype)
+        return LoraProduct(a, b, self.strength)
 
     def cleanup(self) -> None:
         self._pinned_ab.clear()

packages/ltx-core/src/ltx_core/block_streaming/pool.py

@@ -1,4 +1,4 @@
-"""Weight buffer pool for block streaming."""
+"""Raw buffer pool for block streaming."""
 
 from __future__ import annotations
 
@@ -7,58 +7,65 @@ from typing import Callable
 
 import torch
 
-from ltx_core.block_streaming.utils import allocate_buffer
+from ltx_core.block_streaming import utils
-
+from ltx_core.block_streaming.stream_sync import StreamEvent
-# Type alias for the buffer layout used by slot allocation.
-BlockLayout = dict[str, tuple[torch.Size, torch.dtype]]
 
 
-class WeightPool:
+class BufferPool:
-    """Fixed pool of pre-allocated weight buffers with event-based reuse safety.
+    """Fixed pool of pre-allocated raw buffer slots with event-based reuse.
-    Buffers are allocated once at construction.  :meth:`acquire` pops a
+    Slots are carved from a single contiguous ``uint8`` buffer; each is
-    free buffer (waiting any pending event first).  :meth:`release`
+    ``slot_nbytes`` long and handed out as a raw 1-D ``uint8`` tensor.
-    returns it, optionally attaching an event that must complete before
-    the buffer can be reused.
     Args:
-        layout: ``{name: (shape, dtype)}`` for each buffer.
+        slot_nbytes: Byte size of each slot.
-        capacity: Number of buffers to pre-allocate.
+        capacity: Number of slots to pre-allocate.
         device: Device for allocation.
-        reuse_barrier: Called with the pending event before a buffer is reused.
+        reuse_barrier: Called with the pending event before a slot is reused.
         pin_memory: Pin buffers (for async H2D copies from CPU).
     """
 
     def __init__(
         self,
-        layout: BlockLayout,
+        slot_nbytes: int,
         capacity: int,
         device: torch.device,
-        reuse_barrier: Callable[[torch.cuda.Event], None],
+        reuse_barrier: Callable[[StreamEvent], None],
         pin_memory: bool = False,
     ) -> None:
+        self._slot_nbytes = slot_nbytes
         self._capacity = capacity
-        self._free: deque[dict[str, torch.Tensor]] = deque()
+        self._free: deque[torch.Tensor] = deque()
-        self._events: dict[int, torch.cuda.Event] = {}
+        self._events: dict[int, StreamEvent] = {}
         self._reuse_barrier = reuse_barrier
-        for _ in range(capacity):
+        buffer = utils.alloc_buffer(max(slot_nbytes * capacity, 1), device, pin_memory)
-            self._free.append(allocate_buffer(layout, device, pin_memory))
+        for slot in range(capacity):
+            self._free.append(buffer[slot * slot_nbytes : (slot + 1) * slot_nbytes])
 
     @property
     def capacity(self) -> int:
         return self._capacity
 
-    def acquire(self) -> dict[str, torch.Tensor]:
+    @property
-        """Take a free buffer, waiting any pending event before returning."""
+    def slot_nbytes(self) -> int:
-        weights = self._free.popleft()
+        return self._slot_nbytes
-        event = self._events.pop(id(weights), None)
+
+    def acquire(self) -> torch.Tensor:
+        """Take a free raw slot, waiting any pending event before returning.
+        Raises :class:`RuntimeError` if every slot is currently in use.
+        """
+        if not self._free:
+            raise RuntimeError(f"BufferPool exhausted: all {self._capacity} buffers are in use")
+        buffer = self._free.popleft()
+        event = self._events.pop(id(buffer), None)
         if event is not None:
             self._reuse_barrier(event)
-        return weights
+        return buffer
 
-    def release(self, weights: dict[str, torch.Tensor], event: torch.cuda.Event | None = None) -> None:
+    def release(self, buffer: torch.Tensor, event: StreamEvent | None = None) -> None:
-        """Return a buffer to the free list.
+        """Return a raw slot to the free list.
-        If *event* is given it is waited on the next :meth:`acquire`
+        The *buffer* must be the exact tensor object returned by :meth:`acquire`
-        of this buffer, ensuring the prior operation has completed.
+        (reuse is keyed on its identity). If *event* is given it is waited on the
+        next :meth:`acquire` of this slot, ensuring the prior operation finished.
         """
         if event is not None:
-            self._events[id(weights)] = event
+            self._events[id(buffer)] = event
-        self._free.append(weights)
+        self._free.append(buffer)

packages/ltx-core/src/ltx_core/block_streaming/provider.py

@@ -3,91 +3,130 @@
 from __future__ import annotations
 
 from collections import OrderedDict
+from typing import NamedTuple
 
 import torch
 
 from ltx_core.block_streaming.disk import LoraSource
-from ltx_core.block_streaming.pool import WeightPool
+from ltx_core.block_streaming.pool import BufferPool
 from ltx_core.block_streaming.source import WeightSource
+from ltx_core.block_streaming.stream_sync import StreamEvent, StreamSync
+from ltx_core.block_streaming.utils import carve_buffer, layout_nbytes
+from ltx_core.loader.fuse_loras import FuseRule, aggregate_lora_products, bf16_fuse_rule, device_fuse_rule
+from ltx_core.loader.primitives import StateDict
+
+_EMPTY_STATE_DICT = StateDict(sd={}, device=torch.device("cpu"), size=0, dtype=set())
+
+
+class CachedBlock(NamedTuple):
+    """A cached GPU block: the raw pool slot plus the carved per-key views.
+    The raw slot is what is returned to the pool on eviction; the views are
+    what callers consume.
+    """
+
+    raw: torch.Tensor
+    views: dict[str, torch.Tensor]
 
 
 class WeightsProvider:
     """Provides GPU-ready block weights via H2D copy from a pinned CPU weight source.
     Args:
         pool: Pre-allocated GPU weight buffer pool.
-        copy_stream: Dedicated CUDA stream for async H2D copies.
+        sync: Coordinates copy-vs-compute ordering for the backend
-        target_device: GPU device for compute.
+            (see :class:`StreamSync`).
+        target_device: device for compute.
         source: Pinned CPU weight source.
         lora_sources: LoRA adapters fused on H2D copy.
         blocks_prefix: State-dict prefix for LoRA key matching.
+        fuse_rule: Per-policy LoRA merge rule (must be streaming-compatible:
+            no companion-key emission). Defaults to ``bf16_fuse_rule``.
     """
 
     def __init__(
         self,
-        pool: WeightPool,
+        pool: BufferPool,
-        copy_stream: torch.cuda.Stream,
+        sync: StreamSync,
         target_device: torch.device,
         source: WeightSource,
         lora_sources: list[LoraSource] | None = None,
         blocks_prefix: str = "",
+        fuse_rule: FuseRule = bf16_fuse_rule,
     ) -> None:
-        self._copy_stream = copy_stream
+        self._sync = sync
         self._pool = pool
-        self._cache: OrderedDict[int, dict[str, torch.Tensor]] = OrderedDict()
+        self._cache: OrderedDict[int, CachedBlock] = OrderedDict()
-        self._events: dict[int, torch.cuda.Event] = {}
+        self._events: dict[int, StreamEvent | None] = {}
         self._target_device = target_device
         self._source = source
         self._lora_sources = lora_sources or []
         self._blocks_prefix = blocks_prefix
+        self._fuse_rule = fuse_rule
 
     def get(self, idx: int) -> dict[str, torch.Tensor]:
         """Return GPU weights for block *idx*. Does H2D copy on miss."""
         if idx in self._cache:
-            return self._cache[idx]
+            return self._cache[idx].views
 
         # Evict oldest GPU buffer if at capacity.
         if len(self._cache) >= self._pool.capacity:
-            evicted_idx, evicted_weights = self._cache.popitem(last=False)
+            evicted_idx, evicted = self._cache.popitem(last=False)
-            self._pool.release(evicted_weights, event=self._events.pop(evicted_idx, None))
+            self._pool.release(evicted.raw, event=self._events.pop(evicted_idx, None))
 
-        gpu_weights = self._pool.acquire()
+        layout = self._source.block_layout(idx)
-        cpu_weights = self._source.get(idx)
+        raw = self._pool.acquire()
+        gpu_weights = carve_buffer(raw, layout)
+        cpu_buffer = self._source.get(idx)
 
-        h2d_event = self._copy_to_gpu(idx, gpu_weights, cpu_weights)
+        h2d_event = self._copy_to_gpu(idx, raw, gpu_weights, cpu_buffer, layout_nbytes(layout))
         self._source.release(idx, event=h2d_event)
 
-        self._cache[idx] = gpu_weights
+        self._cache[idx] = CachedBlock(raw, gpu_weights)
         return gpu_weights
 
     def _copy_to_gpu(
         self,
         idx: int,
+        raw: torch.Tensor,
         gpu_weights: dict[str, torch.Tensor],
-        cpu_weights: dict[str, torch.Tensor],
+        cpu_buffer: torch.Tensor,
-    ) -> torch.cuda.Event:
+        nbytes: int,
-        """Enqueue H2D copy + LoRA fusion on the copy stream and wait on compute.
+    ) -> StreamEvent | None:
-        The wait is intentionally inside this method so callers -- and
+        """Copy block weights to the target device and fuse LoRAs.
-        instrumentation regions wrapping it -- observe the full transfer time.
+        *cpu_buffer* is one contiguous source buffer carved by the same layout as
+        *raw*, so a single byte copy of its leading *nbytes* reproduces every view
+        in *gpu_weights*.
+        The copy + fusion run under :meth:`StreamSync.copy_scope`, then
+        :meth:`StreamSync.commit_copy` orders the copy before compute and returns
+        a guard event for the source to reuse (the ordering is committed inside
+        this method so callers -- and instrumentation regions wrapping it --
+        observe the full transfer time).
         """
-        with torch.cuda.stream(self._copy_stream):
+        if not cpu_buffer.is_contiguous() or cpu_buffer.dtype != torch.uint8 or cpu_buffer.numel() < nbytes:
-            for name, gpu_tensor in gpu_weights.items():
+            raise ValueError(
-                gpu_tensor.copy_(cpu_weights[name], non_blocking=True)
+                f"source buffer for block {idx} must be a contiguous uint8 buffer of >= {nbytes} bytes, "
+                f"got {cpu_buffer.dim()}-D {cpu_buffer.dtype} with {cpu_buffer.numel()} elements"
+            )
+        with self._sync.copy_scope():
+            raw[:nbytes].copy_(cpu_buffer[:nbytes], non_blocking=self._sync.is_async_copy)
             if self._lora_sources:
                 self._fuse_block_loras(idx, gpu_weights)
-            h2d_event = torch.cuda.Event()
-            h2d_event.record(self._copy_stream)
 
-        torch.cuda.current_stream(self._target_device).wait_event(h2d_event)
+        return self._sync.commit_copy()
-        return h2d_event
 
-    def release(self, idx: int, event: torch.cuda.Event) -> None:
+    def release(self, idx: int, event: StreamEvent | None) -> None:
-        """Attach a compute-done event -- waited before this buffer is recycled."""
+        """Attach a compute-done guard, waited before this buffer is recycled
+        (``None`` when the backend needs no guard)."""
         self._events[idx] = event
 
+    def mark_block_done(self, idx: int) -> None:
+        """Record a compute-done guard for block *idx* and queue it for slot reuse.
+        Called once the block's forward pass has been enqueued, so the buffer is
+        not overwritten by a later copy until this compute completes."""
+        self.release(idx, self._sync.record_compute_done())
+
     def cleanup(self) -> None:
-        """Synchronize streams and release all resources."""
+        """Drain outstanding copy/compute work and release all resources."""
-        self._copy_stream.synchronize()
+        self._sync.synchronize()
-        torch.cuda.current_stream(self._target_device).synchronize()
         self._cache.clear()
         self._events.clear()
         self._source.cleanup()
@@ -98,13 +137,29 @@ class WeightsProvider:
         return len(self._cache)
 
     def _fuse_block_loras(self, idx: int, weights: dict[str, torch.Tensor]) -> None:
-        """Fuse LoRA deltas directly into GPU block weights."""
+        """Fuse LoRA deltas directly into GPU block weights via ``fuse_rule``.
+        The fusion device+dtype come from :func:`device_fuse_rule`: on MPS it
+        aggregates the ``B@A`` on the GPU in fp32 (fast, and fp32 avoids the
+        bf16-on-MPS unreliability), with the rule casting back to the weight
+        dtype; CUDA/CPU keep the rule's dtype. ``get_ab`` places A/B on the
+        target device for CUDA/MPS, so aggregation and the in-place fuse stay
+        co-located there.
+        """
+        rule = device_fuse_rule(self._target_device, self._fuse_rule)
         for name, tensor in weights.items():
             if not name.endswith(".weight"):
                 continue
-            full_key = f"{self._blocks_prefix}.{idx}.{name}"
+            prefix = f"{self._blocks_prefix}.{idx}.{name}".removesuffix(".weight")
-            prefix = full_key[: -len(".weight")]
+            products = (
-            for source in self._lora_sources:
+                ab
-                delta = source.get_delta(prefix, device=self._target_device)
+                for ab in (
-                if delta is not None:
+                    s.get_ab(prefix, device=self._target_device, dtype=rule.aggregation_dtype)
-                    tensor.add_(delta.to(dtype=tensor.dtype))
+                    for s in self._lora_sources
+                )
+                if ab is not None
+            )
+            deltas = aggregate_lora_products(products, rule.aggregation_dtype)
+            if deltas is None:
+                continue
+            fused = rule(name, tensor, deltas, _EMPTY_STATE_DICT)
+            tensor.copy_(fused[name])

packages/ltx-core/src/ltx_core/block_streaming/source.py

@@ -2,23 +2,39 @@
 
 from __future__ import annotations
 
-from collections import OrderedDict
+from typing import NamedTuple, Protocol
-from typing import Protocol
 
 import torch
 
-from ltx_core.block_streaming.disk import DiskBlockReader
+from ltx_core.block_streaming.block_fetcher import BlockFetcher, FetchHandle
-from ltx_core.block_streaming.pool import WeightPool
+from ltx_core.block_streaming.pool import BufferPool
+from ltx_core.block_streaming.stream_sync import StreamEvent
+from ltx_core.block_streaming.utils import carve_buffer, layout_nbytes
+from ltx_core.loader.primitives import TensorLayout
 
 
 class WeightSource(Protocol):
-    """Provides pinned CPU weights for a given block index."""
+    """Provides pinned CPU weights for a given block index.
+    Blocks may be heterogeneous: each has its own layout, so the source exposes a
+    per-block layout and the byte size of the largest block (which sizes the
+    pool slots -- a smaller block is carved into the front of a max-sized slot).
+    The source is the single source of truth for each block's layout.
+    """
 
-    def get(self, idx: int) -> dict[str, torch.Tensor]:
+    def block_layout(self, idx: int) -> TensorLayout:
-        """Return CPU weights for block *idx*."""
+        """Per-block buffer layout (shape + dtype for each param)."""
         ...
 
-    def release(self, idx: int, event: torch.cuda.Event) -> None:
+    @property
+    def slot_nbytes(self) -> int:
+        """Byte size of the largest block; sizes a pool slot (16-byte aligned)."""
+        ...
+
+    def get(self, idx: int) -> torch.Tensor:
+        """Return one contiguous CPU buffer for block *idx*."""
+        ...
+
+    def release(self, idx: int, event: StreamEvent | None) -> None:
         """Signal that an async operation using these weights is guarded by *event*."""
         ...
 
@@ -27,57 +43,133 @@ class WeightSource(Protocol):
         ...
 
 
+class _Scheduled(NamedTuple):
+    """A scheduled (possibly in-flight) read: the raw pool slot and its fetch handle."""
+
+    raw: torch.Tensor
+    status: FetchHandle
+
+
 class DiskWeightSource(WeightSource):
-    """Reads block weights from disk into pinned CPU buffers on demand."""
+    """WeightSource that streams blocks from disk via a :class:`BlockFetcher`.
+    ``get(idx)`` must be paired with a ``release(idx)`` before *idx* is fetched
+    again; getting a block that is already in flight raises. Each read acquires a
+    raw pool slot, carves it to block *idx*'s layout, and hands the carved views to
+    the fetcher to fill, so one max-sized slot can serve blocks of differing shapes.
+    ``get`` returns that same contiguous slot.
+    """
+
+    def __init__(
+        self,
+        pool: BufferPool,
+        fetcher: BlockFetcher,
+        block_layouts: dict[int, TensorLayout],
+        blocks_number: int,
+        prefetch_depth: int = 0,
+    ) -> None:
+        if blocks_number <= 0:
+            raise ValueError(f"blocks_number must be > 0, got {blocks_number}")
+        if prefetch_depth < 0:
+            raise ValueError(f"prefetch_depth must be >= 0, got {prefetch_depth}")
+        max_layout_nbytes = max((layout_nbytes(layout) for layout in block_layouts.values()), default=0)
+        if pool.slot_nbytes < max_layout_nbytes:
+            raise ValueError(
+                f"pool slot is too small for the largest block: slot {pool.slot_nbytes} bytes < {max_layout_nbytes}"
+            )
 
-    def __init__(self, pool: WeightPool, reader: DiskBlockReader) -> None:
         self._pool = pool
-        self._cache: OrderedDict[int, dict[str, torch.Tensor]] = OrderedDict()
+        self._blocks_number = blocks_number
-        self._events: dict[int, torch.cuda.Event] = {}
+        self._prefetch_depth = prefetch_depth
-        self._reader = reader
+        self._fetcher = fetcher
+        self._block_layouts = block_layouts
+        self._scheduled: dict[int, _Scheduled] = {}
+        self._in_flight: dict[int, torch.Tensor] = {}
 
-    def get(self, idx: int) -> dict[str, torch.Tensor]:
+    def block_layout(self, idx: int) -> TensorLayout:
-        """Return CPU weights for block *idx*. Reads from disk on miss."""
+        return self._block_layouts[idx]
-        if idx in self._cache:
-            return self._cache[idx]
 
-        if len(self._cache) >= self._pool.capacity:
+    @property
-            evicted_idx, evicted_weights = self._cache.popitem(last=False)
+    def slot_nbytes(self) -> int:
-            self._pool.release(evicted_weights, event=self._events.pop(evicted_idx, None))
+        return self._pool.slot_nbytes
 
-        weights = self._pool.acquire()
+    def get(self, idx: int) -> torch.Tensor:
-        self._reader.read_into(weights, idx)
+        if idx in self._in_flight:
-        self._cache[idx] = weights
+            raise RuntimeError(f"Block {idx} is already in flight; release it before getting it again")
-        return weights
 
-    def release(self, idx: int, event: torch.cuda.Event) -> None:
+        scheduled = self._scheduled.pop(idx, None)
-        """Attach an H2D event -- waited before this buffer is recycled."""
+        if scheduled is None:
-        self._events[idx] = event
+            scheduled = self._schedule(idx)
+        error = scheduled.status.wait()
+        if error is not None:
+            self._pool.release(scheduled.raw)
+            raise error
+
+        self._in_flight[idx] = scheduled.raw
+        for k in range(1, self._prefetch_depth + 1):
+            self._ensure_scheduled((idx + k) % self._blocks_number)
+        return scheduled.raw
+
+    def release(self, idx: int, event: StreamEvent | None) -> None:
+        raw_buffer = self._in_flight.pop(idx)
+        self._pool.release(raw_buffer, event=event)
 
     def cleanup(self) -> None:
-        """Clear cache and close the disk reader."""
+        self._fetcher.cleanup()
-        self._cache.clear()
+        while self._in_flight:
-        self._events.clear()
+            _, raw_buffer = self._in_flight.popitem()
-        self._reader.cleanup()
+            self._pool.release(raw_buffer)
+        while self._scheduled:
+            _, scheduled = self._scheduled.popitem()
+            scheduled.status.wait()
+            self._pool.release(scheduled.raw)
 
-    def __len__(self) -> int:
+    def _ensure_scheduled(self, idx: int) -> None:
-        return len(self._cache)
+        """Schedule a read for *idx* if one is not already pending."""
+        if idx not in self._scheduled:
+            self._scheduled[idx] = self._schedule(idx)
+
+    def _schedule(self, idx: int) -> _Scheduled:
+        """Acquire a raw slot, carve it to block *idx*, enqueue a read, return the handle.
+        The raw slot and its fetch status are returned together so the caller can
+        track both as one unit; the fetcher only receives the carved views to fill.
+        """
+        raw_buffer = self._pool.acquire()
+        carved = carve_buffer(raw_buffer, self._block_layouts[idx])
+        status = self._fetcher.submit(idx, carved)
+        return _Scheduled(raw_buffer, status)
+
+
+class PinnedBlock(NamedTuple):
+    """A pre-loaded pinned block: its single contiguous buffer and the layout to carve it with."""
+
+    buffer: torch.Tensor
+    layout: TensorLayout
 
 
 class PinnedWeightSource(WeightSource):
-    """Pre-loaded pinned CPU weights."""
+    """Pre-loaded pinned CPU weights, one contiguous (possibly heterogeneous) buffer per block."""
 
-    def __init__(self, weights: dict[int, dict[str, torch.Tensor]]) -> None:
+    def __init__(self, blocks: dict[int, PinnedBlock]) -> None:
-        self._weights = weights
+        if not blocks:
+            raise ValueError("PinnedWeightSource requires at least one block")
+        self._blocks = blocks
+        self._slot_nbytes = max(layout_nbytes(block.layout) for block in blocks.values())
 
-    def get(self, idx: int) -> dict[str, torch.Tensor]:
+    def block_layout(self, idx: int) -> TensorLayout:
-        return self._weights[idx]
+        return self._blocks[idx].layout
 
-    def release(self, idx: int, event: torch.cuda.Event) -> None:
+    @property
+    def slot_nbytes(self) -> int:
+        return self._slot_nbytes
+
+    def get(self, idx: int) -> torch.Tensor:
+        return self._blocks[idx].buffer
+
+    def release(self, idx: int, event: StreamEvent | None) -> None:
         pass
 
     def cleanup(self) -> None:
-        self._weights.clear()
+        self._blocks.clear()
 
     def __len__(self) -> int:
-        return len(self._weights)
+        return len(self._blocks)

packages/ltx-core/src/ltx_core/block_streaming/stream_sync.py

@@ -0,0 +1,174 @@
+"""Copy/compute synchronization for block streaming, abstracted across backends.
+Weight streaming overlaps an H2D weight copy with block compute. The two
+operations must be ordered both ways:
+* copy -> compute: a block must not be read before its weights have landed.
+* compute -> reuse: a GPU buffer slot must not be overwritten by the next
+  copy until the compute that read it has finished.
+On CUDA these are expressed with a dedicated copy stream and cross-stream
+events. MPS exposes no user-facing streams (only ``torch.mps.Event`` on a single
+implicit queue), and CPU is fully synchronous. :class:`StreamSync` hides those
+differences behind one protocol; :func:`create_stream_sync` picks the backend
+implementation. The event types (``torch.cuda.Event`` / ``torch.mps.Event``)
+share the small :class:`StreamEvent` surface the pool and source rely on.
+Kept internal to the streaming module -- nothing else needs stream coordination.
+"""
+
+from __future__ import annotations
+
+import contextlib
+from typing import Protocol, runtime_checkable
+
+import torch
+
+from ltx_core.devices import is_mps_available
+
+
+@runtime_checkable
+class StreamEvent(Protocol):
+    """A device synchronization marker (``torch.cuda.Event`` / ``torch.mps.Event``)."""
+
+    def wait(self) -> None:
+        """Device-side: make subsequently queued work wait for this event."""
+        ...
+
+    def synchronize(self) -> None:
+        """Host-side: block the calling thread until this event completes."""
+        ...
+
+
+class StreamSync(Protocol):
+    """Coordinates the H2D copy against block compute for one streaming model."""
+
+    @property
+    def is_async_copy(self) -> bool:
+        """Whether H2D copies may be enqueued asynchronously."""
+        ...
+
+    def copy_scope(self) -> contextlib.AbstractContextManager[None]:
+        """Context to enqueue the H2D copy under (the copy stream on CUDA)."""
+        ...
+
+    def commit_copy(self) -> StreamEvent | None:
+        """Record a copy-done event and make compute wait on it.
+        Returns the event so the source can guard reuse of the CPU buffer (the
+        disk path host-synchronizes on it), or ``None`` when copies are
+        synchronous and no guard is needed.
+        """
+        ...
+
+    def record_compute_done(self) -> StreamEvent | None:
+        """Record an event marking the end of a block's compute, for slot reuse."""
+        ...
+
+    def reuse_barrier(self, event: StreamEvent | None) -> None:
+        """Before a slot is overwritten by a new copy, wait for *event* (prior compute)."""
+        ...
+
+    def synchronize(self) -> None:
+        """Drain all outstanding copy and compute work."""
+        ...
+
+
+class CudaStreamSync:
+    """CUDA: a dedicated copy stream plus cross-stream events.
+    H2D copies run on ``copy_stream`` so they overlap compute on the default
+    stream; events order the two directions explicitly.
+    """
+
+    def __init__(self, device: torch.device) -> None:
+        self._device = device
+        self._copy_stream = torch.cuda.Stream(device=device)
+
+    @property
+    def is_async_copy(self) -> bool:
+        return True
+
+    def copy_scope(self) -> contextlib.AbstractContextManager[None]:
+        return torch.cuda.stream(self._copy_stream)
+
+    def commit_copy(self) -> StreamEvent:
+        event = torch.cuda.Event()
+        event.record(self._copy_stream)
+        torch.cuda.current_stream(self._device).wait_event(event)
+        return event
+
+    def record_compute_done(self) -> StreamEvent:
+        event = torch.cuda.Event()
+        event.record(torch.cuda.current_stream(self._device))
+        return event
+
+    def reuse_barrier(self, event: StreamEvent | None) -> None:
+        if event is not None:
+            self._copy_stream.wait_event(event)
+
+    def synchronize(self) -> None:
+        self._copy_stream.synchronize()
+        torch.cuda.current_stream(self._device).synchronize()
+
+
+class MpsStreamSync:
+    """MPS: one implicit queue with ``torch.mps.Event`` markers.
+    There is no user-facing copy stream, so copy and compute already serialize
+    on the single default queue. The events make that ordering explicit -- and,
+    crucially, let the buffer pool guard slot reuse on the compute-done event
+    rather than relying on the implicit single-queue ordering. ``Event.wait``
+    enqueues a device-side wait on the default queue (it does not block the
+    host); ``Event.synchronize`` is the host-blocking variant.
+    """
+
+    @property
+    def is_async_copy(self) -> bool:
+        return False
+
+    def copy_scope(self) -> contextlib.AbstractContextManager[None]:
+        return contextlib.nullcontext()
+
+    def commit_copy(self) -> StreamEvent:
+        event = torch.mps.Event()
+        event.record()
+        event.wait()
+        return event
+
+    def record_compute_done(self) -> StreamEvent:
+        event = torch.mps.Event()
+        event.record()
+        return event
+
+    def reuse_barrier(self, event: StreamEvent | None) -> None:
+        if event is not None:
+            event.wait()
+
+    def synchronize(self) -> None:
+        torch.mps.synchronize()
+
+
+class SynchronousStreamSync:
+    """CPU (and any non-accelerator backend): copies are synchronous, no events."""
+
+    @property
+    def is_async_copy(self) -> bool:
+        return False
+
+    def copy_scope(self) -> contextlib.AbstractContextManager[None]:
+        return contextlib.nullcontext()
+
+    def commit_copy(self) -> None:
+        return None
+
+    def record_compute_done(self) -> None:
+        return None
+
+    def reuse_barrier(self, event: StreamEvent | None) -> None:  # noqa: ARG002
+        return None
+
+    def synchronize(self) -> None:
+        return None
+
+
+def create_stream_sync(device: torch.device) -> StreamSync:
+    """Return the :class:`StreamSync` implementation for *device*'s backend."""
+    if device.type == "cuda":
+        return CudaStreamSync(device)
+    if device.type == "mps" and is_mps_available():
+        return MpsStreamSync()
+    return SynchronousStreamSync()

packages/ltx-core/src/ltx_core/block_streaming/utils.py

@@ -2,14 +2,24 @@
 
 from __future__ import annotations
 
-import itertools
+import math
-from typing import TYPE_CHECKING, Any
+import weakref
+from dataclasses import dataclass
+from typing import Any, NamedTuple
 
 import torch
 from torch import nn
 
-if TYPE_CHECKING:
+from ltx_core.loader.primitives import TensorLayout
-    from ltx_core.block_streaming.pool import BlockLayout
+
+FP8_DTYPES = frozenset({torch.float8_e4m3fn, torch.float8_e5m2})
+
+_BUFFER_ALIGN = 16
+
+
+def make_block_key(blocks_prefix: str, block_idx: int, param_name: str) -> str:
+    """Return the state-dict key for *param_name* under block *block_idx*."""
+    return f"{blocks_prefix}.{block_idx}.{param_name}"
 
 
 def resolve_attr(module: nn.Module, dotted_path: str) -> nn.ModuleList:
@@ -40,21 +50,120 @@ def assign_tensor_to_module(root: nn.Module, dotted_name: str, tensor: torch.Ten
         raise AttributeError(f"{leaf} is not a parameter or buffer of {type(parent).__name__}")
 
 
-def build_pool_layout(block: nn.Module, dtype: torch.dtype) -> BlockLayout:
+def derive_layout(tensors: dict[str, torch.Tensor], dtype: torch.dtype | None = None) -> TensorLayout:
-    """Derive a buffer layout from a block's parameters and buffers.
+    """Derive a layout from a ``{name: tensor}`` dict.
-    Works on meta-device blocks (shapes are valid regardless of device).
+    If ``dtype`` is given, non-FP8 dtypes are coerced to it (FP8 preserved). If
-    The *dtype* argument overrides each tensor's dtype so the pool matches
+    ``None``, the source dtype is preserved as-is.
-    the target inference precision.
     """
-    layout: BlockLayout = {}
-    for name, tensor in itertools.chain(block.named_parameters(), block.named_buffers()):
-        layout[name] = (tensor.shape, dtype)
-    return layout
-
-
-def allocate_buffer(layout: BlockLayout, device: torch.device, pin_memory: bool = False) -> dict[str, torch.Tensor]:
-    """Allocate a single buffer dict matching *layout*."""
     return {
-        name: torch.empty(shape, dtype=dtype, device=device, pin_memory=pin_memory)
+        name: (t.shape, t.dtype if dtype is None or t.dtype in FP8_DTYPES else dtype) for name, t in tensors.items()
-        for name, (shape, dtype) in layout.items()
     }
+
+
+def _align_up(offset: int, alignment: int) -> int:
+    return (offset + alignment - 1) & ~(alignment - 1)
+
+
+def _alloc_pinned_exact(nbytes: int) -> torch.Tensor | None:
+    """Allocate exactly ``nbytes`` of pinned host memory via ``cudaHostRegister``.
+    Bypasses PyTorch's ``CachingHostAllocator``, which rounds every
+    ``pin_memory=True`` request up to ``PowerOf2Ceil(N)`` (see
+    ``aten/src/ATen/core/CachingHostAllocator.h``). Returns ``None`` if
+    registration fails. The unregister hook is bound to the storage (not the
+    tensor) so views of the buffer keep the registration alive until the
+    memory is actually freed. Caller is responsible for ensuring CUDA is
+    available.
+    """
+    cudart = torch.cuda.cudart()
+    buf = torch.empty(nbytes, dtype=torch.uint8)
+    ptr = buf.data_ptr()
+    err = int(cudart.cudaHostRegister(ptr, nbytes, 0))
+    if err != 0:
+        return None
+    weakref.finalize(buf.untyped_storage(), lambda p=ptr: cudart.cudaHostUnregister(p))
+    return buf
+
+
+def alloc_buffer(nbytes: int, device: torch.device | None, pin_memory: bool) -> torch.Tensor:
+    """Allocate one ``uint8`` buffer for :func:`allocate_layout_views`.
+    For pinned host buffers, prefer ``cudaHostRegister`` to dodge the caching
+    allocator's power-of-2 rounding. Falls back to the caching allocator if
+    registration fails. Pinning is fundamentally a CUDA driver operation; when
+    requested without a CUDA runtime (e.g. on MPS/CPU, where H2D copies are
+    synchronous and pinning is meaningless) it degrades to a normal allocation.
+    """
+    if pin_memory and not torch.cuda.is_available():
+        pin_memory = False  # pinning is CUDA-only; degrade gracefully off-CUDA
+    if pin_memory and (device is None or torch.device(device).type == "cpu"):
+        buf = _alloc_pinned_exact(nbytes)
+        if buf is not None:
+            return buf
+    return torch.empty(nbytes, dtype=torch.uint8, device=device, pin_memory=pin_memory)
+
+
+@dataclass(frozen=True)
+class _TensorSlice:
+    """Location of a single tensor view within the buffer."""
+
+    offset: int
+    shape: torch.Size
+    dtype: torch.dtype
+
+    def size(self) -> int:
+        return math.prod(self.shape) * self.dtype.itemsize
+
+
+class LayoutSlices(NamedTuple):
+    """Per-key tensor slices of a layout plus the total aligned buffer size."""
+
+    slices: dict[str, _TensorSlice]
+    nbytes: int
+
+
+def _layout_slices(layout: TensorLayout) -> LayoutSlices:
+    """Compute the byte offset of each key in *layout* and the total aligned size.
+    The size is at least one byte so empty layouts still produce a valid buffer.
+    """
+    slices: dict[str, _TensorSlice] = {}
+    cursor = 0
+    for key, (shape, dtype) in layout.items():
+        cursor = _align_up(cursor, _BUFFER_ALIGN)
+        slices[key] = _TensorSlice(offset=cursor, shape=shape, dtype=dtype)
+        cursor += slices[key].size()
+    return LayoutSlices(slices, max(_align_up(cursor, _BUFFER_ALIGN), 1))
+
+
+def layout_nbytes(layout: TensorLayout) -> int:
+    """Byte size of one contiguous, 16-byte-aligned buffer holding *layout* (>= 1)."""
+    return _layout_slices(layout).nbytes
+
+
+def carve_buffer(buffer: torch.Tensor, layout: TensorLayout) -> dict[str, torch.Tensor]:
+    """Carve per-key tensor views for *layout* into the front of *buffer*.
+    *buffer* is a 1-D ``uint8`` tensor at least :func:`layout_nbytes` long. Each
+    returned tensor is a non-overlapping slice of its leading bytes reinterpreted
+    at the requested shape and dtype; any trailing bytes are left unused. That
+    slack is what lets one max-sized pool slot hold a smaller (heterogeneous)
+    block. The views keep *buffer*'s storage alive via PyTorch refcounting.
+    """
+    if buffer.dtype != torch.uint8 or buffer.dim() != 1:
+        raise ValueError(f"carve_buffer expects a 1-D uint8 buffer, got {buffer.dim()}-D {buffer.dtype}")
+    slices, nbytes = _layout_slices(layout)
+    if buffer.numel() < nbytes:
+        raise ValueError(f"buffer too small to carve layout: need {nbytes} bytes, got {buffer.numel()}")
+    return {key: buffer[s.offset : s.offset + s.size()].view(s.dtype).view(s.shape) for key, s in slices.items()}
+
+
+def allocate_layout_views(
+    layout: TensorLayout,
+    device: torch.device | None = None,
+    pin_memory: bool = False,
+) -> dict[str, torch.Tensor]:
+    """Allocate a single ``uint8`` buffer and return per-key tensor views into it.
+    All keys in *layout* live in one contiguous allocation; each returned
+    tensor is a non-overlapping slice of that buffer reinterpreted at the
+    requested shape and dtype. The views keep the underlying storage alive
+    via PyTorch refcounting — drop them all to release the memory.
+    """
+    buffer = alloc_buffer(layout_nbytes(layout), device, pin_memory)
+    return carve_buffer(buffer, layout)

packages/ltx-core/src/ltx_core/block_streaming/wrapper.py

@@ -42,6 +42,10 @@ class BlockStreamingWrapper(nn.Module):
         self._hooks: list[torch.utils.hooks.RemovableHandle] = []
         self._register_hooks()
 
+    @property
+    def num_blocks(self) -> int:
+        return self._model.num_blocks
+
     # ------------------------------------------------------------------
     # Hook registration
     # ------------------------------------------------------------------
@@ -55,10 +59,10 @@ class BlockStreamingWrapper(nn.Module):
             assign_tensor_to_module(block, name, gpu_weights[name])
 
     def _post_hook(self, block_idx: int) -> None:
-        """Record a compute-done event and release the block weights."""
+        """Release the block weights once its forward pass has been enqueued.
-        compute_done = torch.cuda.Event()
+        The provider guards the buffer against reuse until this block's compute
-        compute_done.record(torch.cuda.current_stream(self._target_device))
+        completes."""
-        self._provider.release(block_idx, event=compute_done)
+        self._provider.mark_block_done(block_idx)
 
     def _register_hooks(self) -> None:
         for idx, block in enumerate(self._blocks):

packages/ltx-core/src/ltx_core/components/diffusion_steps.py

@@ -4,6 +4,24 @@ from ltx_core.components.protocols import DiffusionStepProtocol
 from ltx_core.utils import to_velocity
 
 
+def _get_ancestral_step(
+    sigma_from: torch.Tensor,
+    sigma_to: torch.Tensor,
+    eta: float = 1.0,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    """Compute ``(sigma_down, sigma_up)`` for one DDIM ancestral sampling step.
+    Both inputs are in the rescaled parameterization ``sigma / alpha``.
+    Returns ``sigma_down`` (deterministic component) and ``sigma_up``
+    (stochastic component) in the same rescaled space.
+    """
+    if not eta:
+        return sigma_to, torch.zeros_like(sigma_to)
+    variance = sigma_to**2 * (sigma_from**2 - sigma_to**2).clamp(min=0) / sigma_from**2
+    sigma_up = (eta * variance**0.5).clamp(max=sigma_to)
+    sigma_down = (sigma_to**2 - sigma_up**2).clamp(min=0) ** 0.5
+    return sigma_down, sigma_up
+
+
 class EulerDiffusionStep(DiffusionStepProtocol):
     """
     First-order Euler method for diffusion sampling.
@@ -104,3 +122,65 @@ class Res2sDiffusionStep(DiffusionStepProtocol):
         # Mix deterministic and stochastic components
         x_noised = alpha_ratio * (denoised_next + sigma_down * eps_next) + sigma_up * noise
         return x_noised.to(output_dtype)
+
+
+class EulerCfgPpDiffusionStep(DiffusionStepProtocol):
+    """Euler step using the CFG++ correction for the ODE derivative.
+    Instead of the standard velocity formula, the ODE derivative is computed
+    from the unconditioned prediction, keeping the conditioned prediction as
+    the target denoised state.  Ancestral (DDIM) noise injection is applied
+    in the rescaled sigma parameterization (sigma / alpha).
+    All diffusion quantities (alpha, ODE derivative, ancestral coefficients)
+    are computed internally from ``sigmas`` and ``uncond_denoised``.
+    Reference: CFG++ (https://arxiv.org/abs/2406.08070).
+    """
+
+    def __init__(self, eta: float = 1.0, s_noise: float = 1.0) -> None:
+        self.eta = eta
+        self.s_noise = s_noise
+
+    def step(
+        self,
+        sample: torch.Tensor,
+        denoised_sample: torch.Tensor,
+        sigmas: torch.Tensor,
+        step_index: int,
+        uncond_denoised: torch.Tensor,
+        noise: torch.Tensor | None = None,
+        **_kwargs,
+    ) -> torch.Tensor:
+        """Advance one CFG++ Euler step.
+        Args:
+            sample: Current noisy latent x_t.
+            denoised_sample: Conditioned denoised prediction x_0^cond.
+            sigmas: Full sigma schedule tensor.
+            step_index: Current step index.
+            uncond_denoised: Unconditioned denoised prediction x_0^uncond,
+                used to compute the ODE derivative direction.
+            noise: Noise tensor for stochastic injection; ignored when
+                ``eta=0`` or ``s_noise=0``.
+        Returns:
+            Updated latent x_{t-1}.
+        """
+        sigma_s = sigmas[step_index].to(torch.float32)
+        sigma_t = sigmas[step_index + 1].to(torch.float32)
+        _eps = torch.finfo(torch.float32).eps
+        # Clamp to avoid division by zero when sigma == 1.0 exactly.
+        alpha_s = (1.0 - sigma_s).clamp(min=_eps)
+        alpha_t = (1.0 - sigma_t).clamp(min=_eps)
+
+        x = sample.to(torch.float32)
+        denoised = denoised_sample.to(torch.float32)
+        uncond = uncond_denoised.to(torch.float32)
+
+        # ODE derivative: direction toward noise using uncond prediction (CFG++ correction)
+        d = (x - alpha_s * uncond) / sigma_s
+
+        # Ancestral step in rescaled sigma space (sigma / alpha)
+        sigma_down, sigma_up = _get_ancestral_step(sigma_s / alpha_s, sigma_t / alpha_t, eta=self.eta)
+        sigma_down = alpha_t * sigma_down
+
+        x_next = alpha_t * denoised + sigma_down * d
+        if noise is not None and self.eta > 0 and self.s_noise > 0:
+            x_next = x_next + alpha_t * noise.to(torch.float32) * self.s_noise * sigma_up
+        return x_next.to(sample.dtype)

packages/ltx-core/src/ltx_core/components/guiders.py

@@ -253,6 +253,11 @@ class MultiModalGuider:
         and as scale * (cond - uncond) for stg, steering the denoising process away from the unconditioned
         prediction.
         """
+        dtype = cond.dtype
+        cond = cond.float()
+        uncond_text = uncond_text.float() if isinstance(uncond_text, torch.Tensor) else uncond_text
+        uncond_perturbed = uncond_perturbed.float() if isinstance(uncond_perturbed, torch.Tensor) else uncond_perturbed
+        uncond_modality = uncond_modality.float() if isinstance(uncond_modality, torch.Tensor) else uncond_modality
         pred = (
             cond
             + (self.params.cfg_scale - 1) * (cond - uncond_text)
@@ -265,7 +270,7 @@ class MultiModalGuider:
             factor = self.params.rescale_scale * factor + (1 - self.params.rescale_scale)
             pred = pred * factor
 
-        return pred
+        return pred.to(dtype)
 
     def do_unconditional_generation(self) -> bool:
         """Returns True if the guider is doing unconditional generation."""

packages/ltx-core/src/ltx_core/components/noisers.py

@@ -17,18 +17,20 @@ class GaussianNoiser(Noiser):
 
     def __init__(self, generator: torch.Generator):
         super().__init__()
-
         self.generator = generator
 
-    def __call__(self, latent_state: LatentState, noise_scale: float = 1.0) -> LatentState:
+    def _sample_noise(self, latent_state: LatentState) -> torch.Tensor:
-        noise = torch.randn(
+        return torch.randn(
             *latent_state.latent.shape,
             device=latent_state.latent.device,
             dtype=latent_state.latent.dtype,
             generator=self.generator,
         )
-        scaled_mask = latent_state.denoise_mask * noise_scale
+
-        latent = noise * scaled_mask + latent_state.latent * (1 - scaled_mask)
+    def __call__(self, latent_state: LatentState, noise_scale: float = 1.0) -> LatentState:
+        noise = self._sample_noise(latent_state)
+        latent = torch.lerp(latent_state.latent.float(), noise.float(), noise_scale)
+        latent = torch.lerp(latent_state.clean_latent.float(), latent, latent_state.denoise_mask)
         return replace(
             latent_state,
             latent=latent.to(latent_state.latent.dtype),

packages/ltx-core/src/ltx_core/components/patchifiers.py

@@ -151,17 +151,22 @@ def get_pixel_coords(
             that treat frame zero differently still yield non-negative timestamps.
     """
     # Broadcast the VAE scale factors so they align with the `(batch, axis, patch, bound)` layout.
+    # Axis 1 of `latent_coords` is ordered (frame/time, height, width) — match that explicitly by
+    # pulling fields from the NamedTuple rather than relying on tuple iteration order.
     broadcast_shape = [1] * latent_coords.ndim
     broadcast_shape[1] = -1  # axis dimension corresponds to (frame/time, height, width)
-    scale_tensor = torch.tensor(scale_factors, device=latent_coords.device).view(*broadcast_shape)
+    scale_tensor = torch.tensor(
+        [scale_factors.time, scale_factors.height, scale_factors.width],
+        device=latent_coords.device,
+    ).view(*broadcast_shape)
 
     # Apply per-axis scaling to convert latent bounds into pixel-space coordinates.
     pixel_coords = latent_coords * scale_tensor
 
     if causal_fix:
-        # VAE temporal stride for the very first frame is 1 instead of `scale_factors[0]`.
+        # VAE temporal stride for the very first frame is 1 instead of `scale_factors.time`.
         # Shift and clamp to keep the first-frame timestamps causal and non-negative.
-        pixel_coords[:, 0, ...] = (pixel_coords[:, 0, ...] + 1 - scale_factors[0]).clamp(min=0)
+        pixel_coords[:, 0, ...] = (pixel_coords[:, 0, ...] + 1 - scale_factors.time).clamp(min=0)
 
     return pixel_coords
 

packages/ltx-core/src/ltx_core/conditioning/__init__.py

@@ -3,17 +3,29 @@
 from ltx_core.conditioning.exceptions import ConditioningError
 from ltx_core.conditioning.item import ConditioningItem
 from ltx_core.conditioning.types import (
+    AudioConditionByReferenceLatent,
     ConditioningItemAttentionStrengthWrapper,
+    DrivingMode,
+    VideoConditionByDrivingLatent,
     VideoConditionByKeyframeIndex,
     VideoConditionByLatentIndex,
+    VideoConditionByMask,
+    VideoConditionByMaskChannels,
     VideoConditionByReferenceLatent,
+    encode_mask_channels,
 )
 
 __all__ = [
+    "AudioConditionByReferenceLatent",
     "ConditioningError",
     "ConditioningItem",
     "ConditioningItemAttentionStrengthWrapper",
+    "DrivingMode",
+    "VideoConditionByDrivingLatent",
     "VideoConditionByKeyframeIndex",
     "VideoConditionByLatentIndex",
+    "VideoConditionByMask",
+    "VideoConditionByMaskChannels",
     "VideoConditionByReferenceLatent",
+    "encode_mask_channels",
 ]

packages/ltx-core/src/ltx_core/conditioning/cond_channels.py

@@ -0,0 +1,28 @@
+"""Helpers for in-context conditioning channels (``LatentState.cond_channels``).
+
+Conditioning channels are extra per-token input features (e.g. SCAIL-2 mask
+channels) that ride alongside the latent in patchified token space ``(B, T, C)``
+and are concatenated onto the latent before the model's first projection. When
+present they must stay length-aligned with the token sequence, so any
+conditioning item that *appends* tokens must also extend the channels.
+"""
+
+from __future__ import annotations
+
+import torch
+
+from ltx_core.types import LatentState
+
+
+def extend_cond_channels(latent_state: LatentState, num_new_tokens: int) -> torch.Tensor | None:
+    """Return ``cond_channels`` extended with ``num_new_tokens`` zero rows, or ``None``.
+
+    Appended tokens (reference / driving / keyframe) carry no in-context signal by default, so they
+    are padded with zeros to preserve the ``T``-alignment invariant. Returns ``None`` unchanged when
+    the state has no conditioning channels (the standard case), keeping non-SCAIL pipelines untouched.
+    """
+    cond_channels = latent_state.cond_channels
+    if cond_channels is None:
+        return None
+    zeros = cond_channels.new_zeros(cond_channels.shape[0], num_new_tokens, cond_channels.shape[2])
+    return torch.cat([cond_channels, zeros], dim=1)

packages/ltx-core/src/ltx_core/conditioning/types/__init__.py

@@ -1,13 +1,23 @@
 """Conditioning type implementations."""
 
 from ltx_core.conditioning.types.attention_strength_wrapper import ConditioningItemAttentionStrengthWrapper
+from ltx_core.conditioning.types.driving_video_cond import DrivingMode, VideoConditionByDrivingLatent
 from ltx_core.conditioning.types.keyframe_cond import VideoConditionByKeyframeIndex
 from ltx_core.conditioning.types.latent_cond import VideoConditionByLatentIndex
+from ltx_core.conditioning.types.mask_channels_cond import VideoConditionByMaskChannels, encode_mask_channels
+from ltx_core.conditioning.types.mask_cond import VideoConditionByMask
+from ltx_core.conditioning.types.reference_audio_cond import AudioConditionByReferenceLatent
 from ltx_core.conditioning.types.reference_video_cond import VideoConditionByReferenceLatent
 
 __all__ = [
+    "AudioConditionByReferenceLatent",
     "ConditioningItemAttentionStrengthWrapper",
+    "DrivingMode",
+    "VideoConditionByDrivingLatent",
     "VideoConditionByKeyframeIndex",
     "VideoConditionByLatentIndex",
+    "VideoConditionByMask",
+    "VideoConditionByMaskChannels",
     "VideoConditionByReferenceLatent",
+    "encode_mask_channels",
 ]

packages/ltx-core/src/ltx_core/conditioning/types/driving_video_cond.py

@@ -0,0 +1,162 @@
+"""Driving-video conditioning for SCAIL-2-style end-to-end character animation.
+
+Ports mechanism 1 + 3 of SCAIL-2 (arXiv:2606.10804) to LTX-2: the *driving*
+video latent is concatenated directly into the DiT token sequence (no skeleton /
+pose intermediate), and carries a fixed spatial offset ``width_offset`` (the
+paper's ΔW) on the RoPE width axis so its coordinates stay detached from the
+main video tokens. This is the inference-only PoC path -- it reuses the existing
+frozen-reference-token machinery and does not touch the transformer or RoPE.
+
+This mirrors :class:`ltx_core.conditioning.types.reference_video_cond.VideoConditionByReferenceLatent`
+(same patchify -> positions -> append -> attention-mask flow); the only new
+behaviour is the mode-aware coordinate assignment.
+
+Scope note (Phase 1): SCAIL-2's in-context mask channels (mechanism 2) and the
+reference-latent height shift ΔH_ref of Replacement Mode require model-weight
+surgery / a separate reference token group and are intentionally out of scope
+here. Because LTX handles the reference image as a frame-0 in-place replacement
+(not a separate token group), the driving-token placement is identical for both
+:class:`DrivingMode` values in Phase 1 -- ``mode`` is stored for forward
+compatibility and to document intent, but does not yet alter the driving
+coordinates. See the plan for the deferred Phase 2 work.
+"""
+
+from __future__ import annotations
+
+from enum import Enum
+
+import torch
+
+from ltx_core.components.patchifiers import get_pixel_coords
+from ltx_core.conditioning.cond_channels import extend_cond_channels
+from ltx_core.conditioning.item import ConditioningItem
+from ltx_core.conditioning.mask_utils import update_attention_mask
+from ltx_core.tools import VideoLatentTools
+from ltx_core.types import LatentState, VideoLatentShape
+
+# Default normalization ceiling for the RoPE width axis. Must stay in sync with
+# the model's ``positional_embedding_max_pos[2]`` (see rope.py:precompute_freqs_cis
+# default ``max_pos=[20, 2048, 2048]`` and model.py `_init_` default). Driving
+# width coordinates that reach or exceed this value would wrap under RoPE.
+DEFAULT_MAX_WIDTH_POSITION = 2048
+
+
+class DrivingMode(Enum):
+    """SCAIL-2 conditioning mode. See module docstring for the Phase 1 caveat."""
+
+    ANIMATION = "animation"
+    REPLACEMENT = "replacement"
+
+
+class VideoConditionByDrivingLatent(ConditioningItem):
+    """Append driving-video tokens with a width-axis RoPE offset (SCAIL-2 ΔW).
+
+    The driving tokens are appended after the target sequence as clean latents
+    (placeholder zeros in the noisy latent), kept frozen (``denoise_mask =
+    1 - strength``), temporally aligned to the target, and shifted along the
+    width axis by ``width_offset`` so they occupy ``[ΔW, ΔW + Wv)`` while the
+    target stays at ``[0, Wv)``.
+
+    Args:
+        latent: Driving video latents ``[B, C, F, H, W]``. Must match the target
+            shape (same F/H/W) so tokens align frame-for-frame with the target.
+        mode: SCAIL-2 mode (reserved for Phase 2; see module docstring).
+        width_offset: ΔW in RoPE pixel-space width units. ``None`` (default) uses
+            the target's pixel width (``target_shape.width * scale_factors.width``),
+            placing the driving tokens immediately to the right of the target.
+        strength: 1.0 keeps the driving latent fully clean (frozen); 0.0 would
+            denoise it. Default 1.0.
+        max_width_position: RoPE width normalization ceiling; validation raises if
+            the shifted driving coordinates would reach it. Keep in sync with the
+            model's ``positional_embedding_max_pos[2]``.
+    """
+
+    def __init__(
+        self,
+        latent: torch.Tensor,
+        mode: DrivingMode = DrivingMode.ANIMATION,
+        width_offset: float | None = None,
+        strength: float = 1.0,
+        max_width_position: int = DEFAULT_MAX_WIDTH_POSITION,
+    ):
+        self.latent = latent
+        self.mode = mode
+        self.width_offset = width_offset
+        self.strength = strength
+        self.max_width_position = max_width_position
+
+    def apply_to(
+        self,
+        latent_state: LatentState,
+        latent_tools: VideoLatentTools,
+    ) -> LatentState:
+        """Append driving tokens with target-aligned time and a ΔW width shift."""
+        tokens = latent_tools.patchifier.patchify(self.latent)
+
+        num_target_tokens = latent_tools.patchifier.get_token_count(latent_tools.target_shape)
+        if tokens.shape[1] != num_target_tokens:
+            raise ValueError(
+                "VideoConditionByDrivingLatent expects the driving latent to match the target shape "
+                f"(same F/H/W): got {tokens.shape[1]} driving tokens vs {num_target_tokens} target tokens. "
+                "Resize/resample the driving video to the target resolution and frame count."
+            )
+
+        # Compute the driving tokens' own pixel-space coordinates (same flow as
+        # the base reference conditioning and create_initial_state).
+        latent_coords = latent_tools.patchifier.get_patch_grid_bounds(
+            output_shape=VideoLatentShape.from_torch_shape(self.latent.shape),
+            device=self.latent.device,
+        )
+        positions = get_pixel_coords(
+            latent_coords=latent_coords,
+            scale_factors=latent_tools.scale_factors,
+            causal_fix=latent_tools.causal_fix,
+        ).to(dtype=torch.float32)
+
+        # Temporal alignment: copy the target's time coordinates so the driving
+        # tokens sit on exactly the same time grid as z_t (robust to causal_fix /
+        # fps nuances). Token order is a flattened (f h w) grid identical to the
+        # target's, so a token-wise copy is frame-aligned.
+        positions[:, 0:1, :] = latent_state.positions[:, 0:1, :num_target_tokens].to(dtype=torch.float32)
+
+        # ΔW: shift the driving tokens along the width axis so they stay spatially
+        # detached from the target tokens. Default offset = target pixel width,
+        # giving target=[0, Wv), driving=[Wv, 2*Wv).
+        width_offset = self.width_offset
+        if width_offset is None:
+            width_offset = float(latent_tools.target_shape.width * latent_tools.scale_factors.width)
+        positions[:, 2, ...] = positions[:, 2, ...] + width_offset
+
+        max_width = positions[:, 2, ...].max().item()
+        if max_width >= self.max_width_position:
+            raise ValueError(
+                f"Driving width coordinate {max_width:.1f} reaches the RoPE ceiling "
+                f"{self.max_width_position} and would wrap. Reduce width_offset "
+                f"(currently {width_offset:.1f}) or lower the output width."
+            )
+
+        denoise_mask = torch.full(
+            size=(*tokens.shape[:2], 1),
+            fill_value=1.0 - self.strength,
+            device=self.latent.device,
+            dtype=self.latent.dtype,
+        )
+
+        new_attention_mask = update_attention_mask(
+            latent_state=latent_state,
+            attention_mask=None,
+            num_noisy_tokens=num_target_tokens,
+            num_new_tokens=tokens.shape[1],
+            batch_size=tokens.shape[0],
+            device=self.latent.device,
+            dtype=self.latent.dtype,
+        )
+
+        return LatentState(
+            latent=torch.cat([latent_state.latent, torch.zeros_like(tokens)], dim=1),
+            denoise_mask=torch.cat([latent_state.denoise_mask, denoise_mask], dim=1),
+            positions=torch.cat([latent_state.positions, positions], dim=2),
+            clean_latent=torch.cat([latent_state.clean_latent, tokens], dim=1),
+            attention_mask=new_attention_mask,
+            cond_channels=extend_cond_channels(latent_state, tokens.shape[1]),
+        )

packages/ltx-core/src/ltx_core/conditioning/types/keyframe_cond.py

@@ -1,6 +1,7 @@
 import torch
 
 from ltx_core.components.patchifiers import get_pixel_coords
+from ltx_core.conditioning.cond_channels import extend_cond_channels
 from ltx_core.conditioning.item import ConditioningItem
 from ltx_core.conditioning.mask_utils import update_attention_mask
 from ltx_core.tools import VideoLatentTools
@@ -10,8 +11,9 @@ from ltx_core.types import LatentState, VideoLatentShape
 class VideoConditionByKeyframeIndex(ConditioningItem):
     """
     Conditions video generation on keyframe latents at a specific frame index.
-    Appends keyframe tokens to the latent state with positions offset by frame_idx,
+    Appends keyframe tokens to the sequence with positions offset by frame_idx: the keyframe
-    and sets denoise strength according to the strength parameter.
+    latents become clean-latent tokens (placeholder zeros in the noisy latent) and the denoise
+    mask is set from the strength parameter.
     To add attention masking, wrap with :class:`ConditioningItemAttentionStrengthWrapper`.
     Args:
         keyframes: Keyframe latents [B, C, F, H, W].
@@ -75,9 +77,10 @@ class VideoConditionByKeyframeIndex(ConditioningItem):
         )
 
         return LatentState(
-            latent=torch.cat([latent_state.latent, tokens], dim=1),
+            latent=torch.cat([latent_state.latent, torch.zeros_like(tokens)], dim=1),
             denoise_mask=torch.cat([latent_state.denoise_mask, denoise_mask], dim=1),
             positions=torch.cat([latent_state.positions, positions], dim=2),
             clean_latent=torch.cat([latent_state.clean_latent, tokens], dim=1),
             attention_mask=new_attention_mask,
+            cond_channels=extend_cond_channels(latent_state, tokens.shape[1]),
         )

packages/ltx-core/src/ltx_core/conditioning/types/latent_cond.py

@@ -9,8 +9,8 @@ from ltx_core.types import LatentState
 class VideoConditionByLatentIndex(ConditioningItem):
     """
     Conditions video generation by injecting latents at a specific latent frame index.
-    Replaces tokens in the latent state at positions corresponding to latent_idx,
+    Sets the clean latents at positions corresponding to latent_idx to the injected latents,
-    and sets denoise strength according to the strength parameter.
+    sets denoise strength according to the strength parameter.
     """
 
     def __init__(self, latent: torch.Tensor, strength: float, latent_idx: int):
@@ -37,7 +37,6 @@ class VideoConditionByLatentIndex(ConditioningItem):
 
         latent_state = latent_state.clone()
 
-        latent_state.latent[:, start_token:stop_token] = tokens
         latent_state.clean_latent[:, start_token:stop_token] = tokens
         latent_state.denoise_mask[:, start_token:stop_token] = 1.0 - self.strength
 

packages/ltx-core/src/ltx_core/conditioning/types/mask_channels_cond.py

@@ -0,0 +1,122 @@
+"""SCAIL-2 in-context mask conditioning (mechanism 2) for LTX-2.
+
+Encodes ``K+1`` semantic pixel-space masks (1 environment switch + ``K`` character
+binding slots) into per-token conditioning channels and writes them onto the
+sequence via :attr:`LatentState.cond_channels`. Following the paper, mask signals
+are carried by the *driving* (and reference) tokens while the noisy target keeps
+an all-zero mask -- so by default the channels are written onto the last ``N``
+tokens of the sequence (the appended driving group), which means **this item must
+be applied after the driving conditioning**.
+
+Channel expansion (LTX-2 faithful port of the paper's ``4(K+1)``): each semantic
+mask is spatially downsampled to the latent grid and temporally stacked along the
+channel dimension by the VAE temporal factor ``t`` (8 for LTX-2, vs 4 for the
+paper's Wan-2.1 backbone), giving ``t*(K+1)`` channels -- ``8*7 = 56`` for the
+default ``K=6``. The model's ``patchify_proj`` must be built with a matching
+``mask_conditioning_channels`` (see :class:`ltx_core.model.transformer.model.LTXModel`).
+"""
+
+from __future__ import annotations
+
+from dataclasses import replace
+
+import torch
+import torch.nn.functional as F
+from einops import rearrange
+
+from ltx_core.conditioning.item import ConditioningItem
+from ltx_core.tools import VideoLatentTools
+from ltx_core.types import LatentState
+
+
+def encode_mask_channels(
+    masks: torch.Tensor,
+    temporal_factor: int,
+    height_lat: int,
+    width_lat: int,
+    frames_lat: int,
+) -> torch.Tensor:
+    """Encode ``[B, K+1, F_pix, H_pix, W_pix]`` masks into ``[B, t*(K+1), F_lat, H_lat, W_lat]``.
+
+    Each semantic mask is area-downsampled to the latent spatial grid, then the pixel frames that
+    map to one latent frame are stacked along the channel dimension (temporal factor ``t``). The
+    causal first latent frame corresponds to a single pixel frame, which is replicated across its
+    ``t`` stacked channels so every latent frame yields a uniform ``t`` channels per semantic class.
+    """
+    b, s, f_pix, _h_pix, _w_pix = masks.shape
+    t = temporal_factor
+    expected_f_pix = (frames_lat - 1) * t + 1
+    if f_pix != expected_f_pix:
+        raise ValueError(
+            f"mask pixel frames ({f_pix}) incompatible with latent frames ({frames_lat}) at temporal "
+            f"factor {t}: expected (F_lat - 1) * {t} + 1 = {expected_f_pix}."
+        )
+
+    # Spatial downsample every (semantic, frame) mask to the latent grid.
+    flat = rearrange(masks.to(dtype=torch.float32), "b s f h w -> (b s f) 1 h w")
+    down = F.interpolate(flat, size=(height_lat, width_lat), mode="area")
+    down = rearrange(down, "(b s f) 1 h w -> b s f h w", b=b, s=s)
+
+    # Temporal stacking. Latent frame 0 = pixel frame 0 (causal), replicated across t channels;
+    # latent frames 1.. group t consecutive pixel frames.
+    first = down[:, :, :1].repeat(1, 1, t, 1, 1).unsqueeze(2)  # [B, S, 1, t, H, W]
+    rest = rearrange(down[:, :, 1:], "b s (fl t) h w -> b s fl t h w", t=t)  # [B, S, F_lat-1, t, H, W]
+    stacked = torch.cat([first, rest], dim=2)  # [B, S, F_lat, t, H, W]
+
+    # Fold (semantic, temporal) into a single channel axis, semantic-major: channel = s * t + ti.
+    return rearrange(stacked, "b s fl t h w -> b (s t) fl h w")  # [B, t*(K+1), F_lat, H_lat, W_lat]
+
+
+class VideoConditionByMaskChannels(ConditioningItem):
+    """Write SCAIL-2 in-context mask channels onto the driving/reference tokens.
+
+    Args:
+        masks: Pixel-space semantic masks ``[B, K+1, F_pix, H_pix, W_pix]``. Channel 0 is the
+            environment switch (whether the environment comes from the reference vs the driving
+            video); channels ``1..K`` are the character binding slots (regions sharing a slot share
+            motion). ``F_pix`` must equal ``(F_lat - 1) * temporal_factor + 1``.
+        applies_to_last_n: Number of trailing tokens to write the mask onto. ``None`` (default) uses
+            the target token count, i.e. the appended driving group when this item runs right after
+            the driving conditioning. The noisy target tokens keep an all-zero mask (paper-faithful).
+    """
+
+    def __init__(self, masks: torch.Tensor, applies_to_last_n: int | None = None):
+        self.masks = masks
+        self.applies_to_last_n = applies_to_last_n
+
+    def apply_to(self, latent_state: LatentState, latent_tools: VideoLatentTools) -> LatentState:
+        shape = latent_tools.target_shape
+        cond = encode_mask_channels(
+            masks=self.masks,
+            temporal_factor=latent_tools.scale_factors.time,
+            height_lat=shape.height,
+            width_lat=shape.width,
+            frames_lat=shape.frames,
+        )
+        tokens = latent_tools.patchifier.patchify(cond)  # [B, T_region, C_mask]
+        b, n_region, c_mask = tokens.shape
+
+        total_tokens = latent_state.latent.shape[1]
+        n = self.applies_to_last_n if self.applies_to_last_n is not None else n_region
+        if n != n_region:
+            raise ValueError(
+                f"applies_to_last_n ({n}) must equal the encoded mask token count ({n_region})."
+            )
+        if n > total_tokens:
+            raise ValueError(
+                f"cannot write {n} mask tokens onto a sequence of only {total_tokens} tokens."
+            )
+
+        cond_channels = latent_state.cond_channels
+        if cond_channels is None:
+            cond_channels = tokens.new_zeros(b, total_tokens, c_mask)
+        else:
+            if cond_channels.shape[2] != c_mask:
+                raise ValueError(
+                    f"existing cond_channels width ({cond_channels.shape[2]}) != mask channels ({c_mask})."
+                )
+            cond_channels = cond_channels.clone()
+
+        start = total_tokens - n
+        cond_channels[:, start : start + n] = tokens.to(dtype=cond_channels.dtype)
+        return replace(latent_state, cond_channels=cond_channels)

packages/ltx-core/src/ltx_core/conditioning/types/mask_cond.py

@@ -0,0 +1,49 @@
+"""Mask-based conditioning for inpainting and spatial conditioning."""
+
+from dataclasses import replace
+
+import torch
+
+from ltx_core.conditioning.item import ConditioningItem
+from ltx_core.tools import LatentTools
+from ltx_core.types import LatentState
+
+
+class VideoConditionByMask(ConditioningItem):
+    """Condition video generation using a binary mask over latent frames.
+    Masked positions (mask=1) receive the provided clean latent values and are
+    excluded from denoising (denoise_mask set to ``1 - strength``). Unmasked
+    positions (mask=0) are left unchanged and denoised normally.
+    The mask operates in **unpatchified latent** space — it should have shape
+    ``[B, F, H, W]`` matching the latent dimensions (after VAE encoding,
+    before patchification). This is consistent with the latent input format
+    used by all other conditioning items.
+    Args:
+        latent: Clean conditioning latents in unpatchified format [B, C, F, H, W].
+            Must match the target shape of the latent tools.
+        mask: Binary mask [B, F, H, W] in unpatchified latent space.
+            1 = conditioning position (clean, excluded from denoising),
+            0 = generated position (noised, denoised normally).
+        strength: Conditioning strength for masked positions. 1.0 = fully clean
+            (no denoising), 0.0 = no conditioning effect. Default 1.0.
+    """
+
+    def __init__(self, latent: torch.Tensor, mask: torch.Tensor, strength: float = 1.0):
+        self.latent = latent
+        self.mask = mask
+        self.strength = strength
+
+    def apply_to(self, latent_state: LatentState, latent_tools: LatentTools) -> LatentState:
+        """Apply mask-based conditioning to the latent state."""
+        tokens = latent_tools.patchifier.patchify(self.latent)
+
+        mask = latent_tools.patchifier.patchify(self.mask.unsqueeze(1))
+
+        m = mask.to(dtype=latent_state.latent.dtype)
+        inv = 1 - m
+
+        return replace(
+            latent_state,
+            clean_latent=latent_state.clean_latent * inv + tokens * m,
+            denoise_mask=latent_state.denoise_mask * inv + (1.0 - self.strength) * m,
+        )

packages/ltx-core/src/ltx_core/conditioning/types/reference_audio_cond.py

@@ -0,0 +1,61 @@
+"""Audio reference conditioning items."""
+
+from __future__ import annotations
+
+import torch
+
+from ltx_core.conditioning.cond_channels import extend_cond_channels
+from ltx_core.conditioning.mask_utils import update_attention_mask
+from ltx_core.tools import LatentTools
+from ltx_core.types import LatentState
+
+
+class AudioConditionByReferenceLatent:
+    """Append patchified reference audio tokens after the target audio sequence.
+    Mirrors :class:`ltx_core.conditioning.types.reference_video_cond.VideoConditionByReferenceLatent`
+    but for audio. The reference tokens are appended so the target audio tokens stay
+    in the first ``num_noisy_tokens`` positions and can be kept by
+    :meth:`ltx_core.tools.LatentTools.clear_conditioning`.
+    Args:
+        patchified: Patchified reference latent ``[B, T_ref, C]``.
+        positions: RoPE positions for reference tokens, ``[B, 1, T_ref, 2]``.
+        strength: 1.0 keeps reference clean; 0.0 would fully denoise it.
+    """
+
+    def __init__(
+        self,
+        patchified: torch.Tensor,
+        positions: torch.Tensor,
+        strength: float = 1.0,
+    ) -> None:
+        self.patchified = patchified
+        self.positions = positions.to(dtype=torch.float32)
+        self.strength = strength
+
+    def apply_to(self, latent_state: LatentState, latent_tools: LatentTools) -> LatentState:
+        tokens = self.patchified
+        denoise_mask = torch.full(
+            size=(*tokens.shape[:2], 1),
+            fill_value=1.0 - self.strength,
+            device=tokens.device,
+            dtype=tokens.dtype,
+        )
+
+        new_attention_mask = update_attention_mask(
+            latent_state=latent_state,
+            attention_mask=None,
+            num_noisy_tokens=latent_tools.patchifier.get_token_count(latent_tools.target_shape),
+            num_new_tokens=tokens.shape[1],
+            batch_size=tokens.shape[0],
+            device=tokens.device,
+            dtype=tokens.dtype,
+        )
+
+        return LatentState(
+            latent=torch.cat([latent_state.latent, torch.zeros_like(tokens)], dim=1),
+            denoise_mask=torch.cat([latent_state.denoise_mask, denoise_mask], dim=1),
+            positions=torch.cat([latent_state.positions, self.positions], dim=2),
+            clean_latent=torch.cat([latent_state.clean_latent, tokens], dim=1),
+            attention_mask=new_attention_mask,
+            cond_channels=extend_cond_channels(latent_state, tokens.shape[1]),
+        )

packages/ltx-core/src/ltx_core/conditioning/types/reference_video_cond.py

@@ -3,6 +3,7 @@
 import torch
 
 from ltx_core.components.patchifiers import get_pixel_coords
+from ltx_core.conditioning.cond_channels import extend_cond_channels
 from ltx_core.conditioning.item import ConditioningItem
 from ltx_core.conditioning.mask_utils import update_attention_mask
 from ltx_core.tools import VideoLatentTools
@@ -14,7 +15,8 @@ class VideoConditionByReferenceLatent(ConditioningItem):
     Conditions video generation on a reference video latent for IC-LoRA inference.
     IC-LoRAs are trained by concatenating reference (control signal) and target tokens,
     learning to attend across both. This class replicates that setup at inference by
-    appending reference tokens to the latent sequence.
+    appending the reference tokens to the sequence as clean latents (with placeholder zeros
+    in the noisy latent).
     IC-LoRAs can be trained with lower-resolution references than the target (e.g., 384px
     reference for 768px output) for efficiency and better generalization. The
     `downscale_factor` scales reference positions to match target coordinates, preserving
@@ -22,9 +24,9 @@ class VideoConditionByReferenceLatent(ConditioningItem):
     (stored in LoRA metadata).
     To add attention masking, wrap with :class:`ConditioningItemAttentionStrengthWrapper`.
     Args:
-        latent: Reference video latents [B, C, F, H, W]
+        latent: Reference video latents [B, C, F, H, W].
-        downscale_factor: Target/reference resolution ratio (e.g., 2 = half-resolution
+        downscale_factor: Target/reference spatial ratio (e.g. 2 = half-res ref).
-            reference). Spatial positions are scaled by this factor.
+        temporal_scale_factor: Target/reference temporal ratio S (e.g. 4 = ref at 1/4 fps).
         strength: Conditioning strength. 1.0 = full (reference kept clean),
             0.0 = none (reference denoised). Default 1.0.
     """
@@ -33,10 +35,12 @@ class VideoConditionByReferenceLatent(ConditioningItem):
         self,
         latent: torch.Tensor,
         downscale_factor: int = 1,
+        temporal_scale_factor: int = 1,
         strength: float = 1.0,
     ):
         self.latent = latent
         self.downscale_factor = downscale_factor
+        self.temporal_scale_factor = temporal_scale_factor
         self.strength = strength
 
     def apply_to(
@@ -44,10 +48,9 @@ class VideoConditionByReferenceLatent(ConditioningItem):
         latent_state: LatentState,
         latent_tools: VideoLatentTools,
     ) -> LatentState:
-        """Append reference video tokens with scaled positions."""
+        """Append reference video tokens with positions translated into the target frame."""
         tokens = latent_tools.patchifier.patchify(self.latent)
 
-        # Compute positions for the reference video's actual dimensions
         latent_coords = latent_tools.patchifier.get_patch_grid_bounds(
             output_shape=VideoLatentShape.from_torch_shape(self.latent.shape),
             device=self.latent.device,
@@ -58,9 +61,18 @@ class VideoConditionByReferenceLatent(ConditioningItem):
             causal_fix=latent_tools.causal_fix,
         )
         positions = positions.to(dtype=torch.float32)
-        positions[:, 0, ...] /= latent_tools.fps
 
-        # Scale spatial positions to match target coordinate space
+        # Place ref tokens on their own time spacing (= target_fps / S).
+        positions[:, 0, ...] /= latent_tools.fps / self.temporal_scale_factor
+
+        # Translate into the target's frame so ref's last patch ends with target's last
+        # patch; clamp the causal patch's negative start back to [0, 1/target_fps).
+        if self.temporal_scale_factor != 1:
+            t_target = latent_state.positions[:, 0, 0:1, 1:2].to(dtype=torch.float32)  # = 1/target_fps
+            positions[:, 0, ...] = torch.clamp(
+                positions[:, 0, ...] - (self.temporal_scale_factor - 1) * t_target,
+                min=0,
+            )
         if self.downscale_factor != 1:
             positions[:, 1, ...] *= self.downscale_factor  # height axis
             positions[:, 2, ...] *= self.downscale_factor  # width axis
@@ -83,9 +95,10 @@ class VideoConditionByReferenceLatent(ConditioningItem):
         )
 
         return LatentState(
-            latent=torch.cat([latent_state.latent, tokens], dim=1),
+            latent=torch.cat([latent_state.latent, torch.zeros_like(tokens)], dim=1),
             denoise_mask=torch.cat([latent_state.denoise_mask, denoise_mask], dim=1),
             positions=torch.cat([latent_state.positions, positions], dim=2),
             clean_latent=torch.cat([latent_state.clean_latent, tokens], dim=1),
             attention_mask=new_attention_mask,
+            cond_channels=extend_cond_channels(latent_state, tokens.shape[1]),
         )

packages/ltx-core/src/ltx_core/devices.py

@@ -0,0 +1,92 @@
+"""Device abstraction for CUDA, Apple Silicon (MPS), and CPU backends.
+Centralizes backend detection and the handful of APIs that genuinely differ
+across accelerators (synchronization, allocator cache, memory queries, RNG
+state). Selection order is CUDA -> MPS -> CPU.
+CUDA-only optimizations (FlashAttention, Triton blockwise FP8/FP6,
+bitsandbytes, NCCL) are gated at their call sites, not here. MPS in particular
+has no ``float64`` support and no fp8 dtype support; use
+:func:`highest_precision_float` to stay within what the backend can represent.
+"""
+
+from __future__ import annotations
+
+import gc
+import logging
+
+import torch
+
+logger = logging.getLogger(__name__)
+
+DeviceSpec = torch.device | None
+
+
+def is_mps_available() -> bool:
+    """Return whether PyTorch can use the Apple Metal/MPS backend."""
+    mps_backend = getattr(torch.backends, "mps", None)
+    return bool(mps_backend is not None and mps_backend.is_available())
+
+
+def get_preferred_device(local_rank: int | None = None) -> torch.device:
+    """Prefer CUDA, then MPS, then CPU.
+    ``local_rank`` is only meaningful for CUDA multi-process launches. MPS exposes
+    a single logical device in PyTorch, so rank-based indexing is not used there.
+    """
+    if torch.cuda.is_available():
+        index = torch.cuda.current_device() if local_rank is None else local_rank
+        return torch.device("cuda", index)
+    if is_mps_available():
+        return torch.device("mps")
+    return torch.device("cpu")
+
+
+def resolve_device(device: DeviceSpec = None, *, local_rank: int | None = None) -> torch.device:
+    """Return *device*, or the best available accelerator when it is ``None``."""
+    if device is None:
+        return get_preferred_device(local_rank=local_rank)
+    return device
+
+
+def supports_float64(device: DeviceSpec) -> bool:
+    """Return whether *device* can represent ``torch.float64``.
+    MPS has no double-precision support; CUDA and CPU do.
+    """
+    return resolve_device(device).type != "mps"
+
+
+def highest_precision_float(device: DeviceSpec) -> torch.dtype:
+    """Return the widest float the backend supports: ``float64`` on CUDA/CPU,
+    ``float32`` on MPS.
+    Use for numerically sensitive accumulators (e.g. sampler ODE math) that
+    request double precision but must degrade gracefully on MPS.
+    """
+    return torch.float64 if supports_float64(device) else torch.float32
+
+
+def synchronize_device(device: DeviceSpec = None) -> None:
+    """Synchronize CUDA or MPS work if the selected backend supports it."""
+    resolved = resolve_device(device)
+    if resolved.type == "cuda" and torch.cuda.is_available():
+        torch.cuda.synchronize(resolved)
+    elif resolved.type == "mps" and is_mps_available():
+        torch.mps.synchronize()
+
+
+def empty_device_cache(device: DeviceSpec = None) -> None:
+    """Release cached allocator memory for CUDA or MPS."""
+    resolved = resolve_device(device)
+    if resolved.type == "cuda" and torch.cuda.is_available():
+        torch.cuda.empty_cache()
+    elif resolved.type == "mps" and is_mps_available():
+        torch.mps.empty_cache()
+
+
+def cleanup_accelerator_memory(device: DeviceSpec = None) -> None:
+    """Run Python GC and release CUDA/MPS allocator caches."""
+    gc.collect()
+    empty_device_cache(device)
+    synchronize_device(device)
+    try:
+        if hasattr(torch._C, "_host_emptyCache"):
+            torch._C._host_emptyCache()
+    except Exception:
+        logger.warning("Host empty cache cleanup failed; ignoring.", exc_info=True)

packages/ltx-core/src/ltx_core/guidance/perturbations.py

@@ -1,17 +1,19 @@
 from dataclasses import dataclass
-from enum import Enum
+from enum import IntEnum
 
 import torch
 from torch._prims_common import DeviceLikeType
 
 
-class PerturbationType(Enum):
+class PerturbationType(IntEnum):
-    """Types of attention perturbations for STG (Spatio-Temporal Guidance)."""
+    """Types of attention perturbations for STG (Spatio-Temporal Guidance).
+    The integer value is the row index into ``BatchedPerturbationConfig._block_masks`` dim 0.
+    """
 
-    SKIP_A2V_CROSS_ATTN = "skip_a2v_cross_attn"
+    SKIP_VIDEO_SELF_ATTN = 0
-    SKIP_V2A_CROSS_ATTN = "skip_v2a_cross_attn"
+    SKIP_AUDIO_SELF_ATTN = 1
-    SKIP_VIDEO_SELF_ATTN = "skip_video_self_attn"
+    SKIP_A2V_CROSS_ATTN = 2
-    SKIP_AUDIO_SELF_ATTN = "skip_audio_self_attn"
+    SKIP_V2A_CROSS_ATTN = 3
 
 
 @dataclass(frozen=True)
@@ -48,32 +50,84 @@ class PerturbationConfig:
         return PerturbationConfig([])
 
 
-@dataclass(frozen=True)
 class BatchedPerturbationConfig:
-    """Perturbation configurations for a batch, with utilities for generating attention masks."""
+    """Per-block attention keep-masks for a batch, built once from a list of per-sample configs.
+    Construction materializes ``_block_masks`` -- a ``(len(PerturbationType), num_blocks, B)`` tensor
+    (1 = keep, 0 = perturbed) whose dim-0 row index is the ``PerturbationType`` value -- from the
+    perturbation structure. The per-sample config list is NOT retained: every consumer reads the
+    tensor (``mask`` indexes it; ``any_in_batch`` / ``all_in_batch`` read the host mirror).
+    The host build (reading the Python structure) happens here, in ``__init__``, so it MUST be run
+    eagerly OUTSIDE any ``torch.compile`` / CUDA-graph-capture region. The compiled block then reads
+    perturbation purely as the runtime ``_block_masks`` tensor and never recompiles per config.
+    """
 
-    perturbations: list[PerturbationConfig]
+    _block_masks: torch.Tensor  # keep-mask on the compute device, indexed [PerturbationType, block, sample]
+    # Host mirror so any_in_batch / all_in_batch stay sync-free and graph-break-free. Present for
+    # configs that may hit the eager skip shortcuts; None for compiled-only configs built via
+    # ``from_masks`` (the compiled processor reads only ``_block_masks``).
+    _block_masks_cpu: torch.Tensor | None
 
-    def mask(
+    def __init__(
-        self, perturbation_type: PerturbationType, block: int, device: DeviceLikeType, dtype: torch.dtype
+        self,
-    ) -> torch.Tensor:
+        perturbations: list[PerturbationConfig],
-        mask = torch.ones((len(self.perturbations),), device=device, dtype=dtype)
+        num_blocks: int,
-        for batch_idx, perturbation in enumerate(self.perturbations):
+        device: DeviceLikeType | None = None,
-            if perturbation.is_perturbed(perturbation_type, block):
+        dtype: torch.dtype | None = None,
-                mask[batch_idx] = 0
+    ) -> None:
+        keep = [
+            [
+                [not pc.is_perturbed(PerturbationType(direction), block) for pc in perturbations]
+                for block in range(num_blocks)
+            ]
+            for direction in range(len(PerturbationType))
+        ]
+        self._block_masks_cpu = torch.tensor(keep, dtype=dtype, device="cpu")
+        self._block_masks = self._block_masks_cpu if device is None else self._block_masks_cpu.to(device)
 
-        return mask
+    @classmethod
+    def from_masks(
+        cls, block_masks: torch.Tensor, block_masks_cpu: torch.Tensor | None = None
+    ) -> "BatchedPerturbationConfig":
+        """Construct from prebuilt mask tensors (e.g. a batch-dim slice), bypassing the host build.
+        ``block_masks_cpu`` is only consumed by ``any_in_batch`` / ``all_in_batch`` (the eager
+        processor's skip shortcuts); pass it when the result may take that path. The compiled
+        processor reads only ``_block_masks``, so callers on that path may omit the mirror.
+        """
+        obj = cls.__new__(cls)
+        obj._block_masks = block_masks
+        obj._block_masks_cpu = block_masks_cpu
+        return obj
 
-    def mask_like(self, perturbation_type: PerturbationType, block: int, values: torch.Tensor) -> torch.Tensor:
+    def batch_slice(self, start: int, end: int) -> "BatchedPerturbationConfig":
-        mask = self.mask(perturbation_type, block, values.device, values.dtype)
+        """A view over samples ``[start:end]`` of the batch, by slicing the mask tensors.
-        return mask.view(mask.numel(), *([1] * len(values.shape[1:])))
+        Slicing (never rebuilding) keeps the host mask build outside any compiled / capture region.
+        """
+        cpu_mask = self._block_masks_cpu[:, :, start:end] if self._block_masks_cpu is not None else None
+        return BatchedPerturbationConfig.from_masks(self._block_masks[:, :, start:end], cpu_mask)
+
+    def mask(self, perturbation_type: PerturbationType, block: int) -> torch.Tensor:
+        """This block's ``(B, 1, 1)`` keep-mask for one perturbation type, as an OWNED tensor.
+        A ``clone`` (not a view into ``_block_masks``) so the masks attached to a block
+        (e.g. self- and cross-attention) don't alias the same storage -- aliased graph inputs are
+        fragile under ``torch.compile``.
+        """
+        return self._block_masks[perturbation_type, block].reshape(-1, 1, 1).clone()
 
     def any_in_batch(self, perturbation_type: PerturbationType, block: int) -> bool:
-        return any(perturbation.is_perturbed(perturbation_type, block) for perturbation in self.perturbations)
+        assert self._block_masks_cpu is not None, "host mirror required by the skip-shortcut processor path"
+        return bool((self._block_masks_cpu[perturbation_type, block] == 0).any())
 
     def all_in_batch(self, perturbation_type: PerturbationType, block: int) -> bool:
-        return all(perturbation.is_perturbed(perturbation_type, block) for perturbation in self.perturbations)
+        assert self._block_masks_cpu is not None, "host mirror required by the skip-shortcut processor path"
+        return bool((self._block_masks_cpu[perturbation_type, block] == 0).all())
 
     @staticmethod
-    def empty(batch_size: int) -> "BatchedPerturbationConfig":
+    def empty(
-        return BatchedPerturbationConfig([PerturbationConfig.empty() for _ in range(batch_size)])
+        batch_size: int,
+        num_blocks: int,
+        device: DeviceLikeType | None = None,
+        dtype: torch.dtype | None = None,
+    ) -> "BatchedPerturbationConfig":
+        return BatchedPerturbationConfig(
+            [PerturbationConfig.empty() for _ in range(batch_size)], num_blocks, device, dtype
+        )

packages/ltx-core/src/ltx_core/loader/attention_ops.py

@@ -0,0 +1,42 @@
+"""Builder ops for swapping attention backends on a meta model before load."""
+
+import torch
+
+from ltx_core.loader.module_ops import ModuleOps
+from ltx_core.model.transformer.attention import (
+    Attention,
+    AttentionCallable,
+    AttentionFunction,
+    MaskedAttentionCallable,
+    MaskedAttentionFunction,
+)
+
+
+def set_attention_module_op(
+    attention: AttentionFunction | AttentionCallable | None = None,
+    masked_attention: MaskedAttentionFunction | MaskedAttentionCallable | None = None,
+) -> ModuleOps:
+    """Build a ``ModuleOps`` that overrides the attention callables on every
+    ``Attention`` submodule of a model. Applied via ``create_meta_model`` so
+    the meta model is mutated before weight loading. Matcher returns False
+    for models with no ``Attention`` submodules, so the op is a no-op there.
+    Either or both slots may be supplied; *None* leaves that slot untouched.
+    """
+    fn = attention.to_callable() if isinstance(attention, AttentionFunction) else attention
+    masked_fn = (
+        masked_attention.to_callable() if isinstance(masked_attention, MaskedAttentionFunction) else masked_attention
+    )
+
+    def matcher(model: torch.nn.Module) -> bool:
+        return any(isinstance(m, Attention) for m in model.modules())
+
+    def mutator(model: torch.nn.Module) -> torch.nn.Module:
+        for module in model.modules():
+            if isinstance(module, Attention):
+                if fn is not None:
+                    module.attention_function = fn
+                if masked_fn is not None:
+                    module.masked_attention_function = masked_fn
+        return model
+
+    return ModuleOps(name="set_attention_backend", matcher=matcher, mutator=mutator)

packages/ltx-core/src/ltx_core/loader/fuse_loras.py

@@ -1,133 +1,198 @@
-from collections.abc import Iterator
+from collections.abc import Callable, Iterable, Iterator
+from dataclasses import dataclass, replace
+from typing import NamedTuple
 
 import torch
 
 from ltx_core.loader.primitives import LoraStateDictWithStrength, StateDict
-from ltx_core.quantization.fp8_cast import _fused_add_round_launch
-from ltx_core.quantization.fp8_scaled_mm import quantize_weight_to_fp8_per_tensor
 
 
-def _get_device() -> torch.device:
+class LoraProduct(NamedTuple):
+    """A LoRA's ``A``, ``B`` factors and its strength scalar."""
+
+    a: torch.Tensor
+    b: torch.Tensor
+    strength: float
+
+
+#: Signature for a fuse callable used by :class:`FuseRule`.
+#:
+#: Args:
+#:     key: The state-dict key being fused (e.g. ``"...layers.0.attn.q.weight"``).
+#:     weight: The current value at ``key`` from ``model_sd``, on the fusion device.
+#:     deltas: The pre-aggregated LoRA delta for ``key``, in ``aggregation_dtype``.
+#:     model_sd: The full state dict, for rules that need companion keys
+#:         (e.g. an existing ``.weight_scale``).
+#:
+#: Returns a dict of state-dict keys to overwrite -- at minimum ``{key: new_weight}``,
+#: plus any companion keys (e.g. an updated ``.weight_scale``) the policy needs to
+#: keep in sync.
+FuseFn = Callable[[str, torch.Tensor, torch.Tensor, StateDict], dict[str, torch.Tensor]]
+
+
+@dataclass(frozen=True)
+class FuseRule:
+    """Fuse an aggregated LoRA delta into one weight key.
+    Each policy supplies its own rule (see ``QuantizationPolicy.fuse_rule``);
+    ``fuse_lora_weights`` is policy-agnostic boilerplate around it.
+    Attributes:
+        aggregation_dtype: Dtype callers must pre-aggregate LoRA deltas in
+            before invoking the rule.
+        fuse_fn: Callable that applies the pre-aggregated deltas to the weight
+            (and any companion keys) and returns a dict of keys to overwrite —
+            at minimum ``{key: new_weight}``, plus any companion keys (e.g. an
+            updated ``.weight_scale`` for scaled-FP8 layouts) the policy needs
+            to keep in sync.
+    """
+
+    aggregation_dtype: torch.dtype
+    fuse_fn: FuseFn
+
+    def __call__(
+        self,
+        key: str,
+        weight: torch.Tensor,
+        deltas: torch.Tensor,
+        model_sd: StateDict,
+    ) -> dict[str, torch.Tensor]:
+        return self.fuse_fn(key, weight, deltas, model_sd)
+
+
+def _bf16_fuse(
+    key: str,
+    weight: torch.Tensor,
+    deltas: torch.Tensor,
+    model_sd: StateDict,  # noqa: ARG001
+) -> dict[str, torch.Tensor]:
+    deltas.add_(weight)
+    return {key: deltas.to(dtype=weight.dtype)}
+
+
+bf16_fuse_rule = FuseRule(aggregation_dtype=torch.bfloat16, fuse_fn=_bf16_fuse)
+
+
+def device_fuse_rule(target_device: torch.device, base_rule: FuseRule) -> FuseRule:
+    """Return the fuse rule to use when fusing onto *target_device*.
+    On MPS, swap the rule's aggregation dtype to fp32: the LoRA ``B@A`` then runs
+    on the GPU (far faster than fusing on CPU) and fp32 sidesteps the bf16-on-MPS
+    numerical unreliability that would otherwise force the slow CPU path. The
+    rule's ``fuse_fn`` still casts the fused result back to the weight dtype.
+    CUDA/CPU keep *base_rule* unchanged.
+    """
+    if target_device.type == "mps":
+        return replace(base_rule, aggregation_dtype=torch.float32)
+    return base_rule
+
+
+def _fusion_device(target_device: torch.device) -> torch.device:
+    """Device to run the fusion on: the target's own accelerator (CUDA/MPS), else
+    CUDA when present (accelerating a CPU-resident fuse), else CPU. The caller
+    moves the fused result back to the weight's device afterwards.
+    """
+    if target_device.type in ("cuda", "mps"):
+        return target_device
     if torch.cuda.is_available():
         return torch.device("cuda", torch.cuda.current_device())
     return torch.device("cpu")
 
 
+def aggregate_lora_products(
+    products: Iterable[LoraProduct],
+    dtype: torch.dtype,
+) -> torch.Tensor | None:
+    """Accumulate ``sum((B * strength) @ A)`` across :class:`LoraProduct` items.
+    The first product materializes a freshly-allocated aggregator via
+    ``torch.matmul(B * strength, A).to(dtype)`` -- preserving the
+    ``(B * strength) @ A`` rounding pattern. Subsequent products use
+    ``addmm_`` to avoid allocating the full intermediate delta.
+    Returns the aggregator, or ``None`` if ``products`` was empty.
+    """
+    aggregated: torch.Tensor | None = None
+    for product in products:
+        if aggregated is None:
+            aggregated = torch.matmul(product.b * product.strength, product.a).to(dtype=dtype)
+        else:
+            aggregated.addmm_(product.b, product.a, alpha=product.strength)
+    return aggregated
+
+
 def fuse_lora_weights(
     model_sd: StateDict,
     lora_sd_and_strengths: list[LoraStateDictWithStrength],
-    dtype: torch.dtype | None = None,
+    fuse_rule: FuseRule = bf16_fuse_rule,
+    preserve_input_device: bool = True,
 ) -> Iterator[tuple[str, torch.Tensor]]:
     """Yield ``(key, fused_tensor)`` for each weight modified by at least one LoRA.
-    For scaled-FP8 weights, this includes both the updated ``.weight`` tensor
+    The fusion math is delegated to ``fuse_rule``.
-    and its corresponding ``.weight_scale`` tensor.
+    Output dtypes are the rule's responsibility.
+    When ``preserve_input_device`` is False, fused tensors are yielded on the device
+    used for fusion; caller is responsible for moving them to their final
+    destination.
     """
-    for key, original_weight in model_sd.sd.items():
+    rule = device_fuse_rule(model_sd.device, fuse_rule)
-        if original_weight is None or key.endswith(".weight_scale"):
+    fusion_device = _fusion_device(model_sd.device)
+    for key in _affected_weight_keys(lora_sd_and_strengths):
+        original_weight = model_sd.sd.get(key)
+        if original_weight is None:
             continue
-        original_device = original_weight.device
-        weight = original_weight.to(device=_get_device())
-        target_dtype = dtype if dtype is not None else weight.dtype
-        deltas_dtype = target_dtype if target_dtype not in [torch.float8_e4m3fn, torch.float8_e5m2] else torch.bfloat16
 
-        deltas = _prepare_deltas(lora_sd_and_strengths, key, deltas_dtype, weight.device)
+        products = _products_for_sd_key(lora_sd_and_strengths, key, rule.aggregation_dtype, fusion_device)
+        deltas = aggregate_lora_products(products, rule.aggregation_dtype)
         if deltas is None:
             continue
 
-        scale_key = key.replace(".weight", ".weight_scale") if key.endswith(".weight") else None
+        original_device = original_weight.device
-        is_scaled_fp8 = scale_key is not None and scale_key in model_sd.sd
+        weight = original_weight.to(device=fusion_device)
 
-        if weight.dtype == torch.float8_e4m3fn:
+        fused = rule(key, weight, deltas, model_sd)
-            if is_scaled_fp8:
-                fused = _fuse_delta_with_scaled_fp8(deltas, weight, key, scale_key, model_sd)
-            else:
-                fused = _fuse_delta_with_cast_fp8(deltas, weight, key, target_dtype)
-        elif weight.dtype == torch.bfloat16:
-            fused = _fuse_delta_with_bfloat16(deltas, weight, key, target_dtype)
-        else:
-            raise ValueError(f"Unsupported dtype: {weight.dtype}")
 
         for k, v in fused.items():
-            yield k, v.to(device=original_device)
+            yield k, v.to(device=original_device) if preserve_input_device else v
 
 
 def apply_loras(
     model_sd: StateDict,
     lora_sd_and_strengths: list[LoraStateDictWithStrength],
-    dtype: torch.dtype | None = None,
+    fuse_rule: FuseRule = bf16_fuse_rule,
     destination_sd: StateDict | None = None,
 ) -> StateDict:
+    """Fuse LoRAs into ``model_sd`` and place the results in ``destination_sd``.
+    When ``destination_sd`` is provided, the fused tensors are placed directly into it.
+    """
+    fused_iter = fuse_lora_weights(
+        model_sd,
+        lora_sd_and_strengths,
+        fuse_rule=fuse_rule,
+    )
     if destination_sd is not None:
-        sd = destination_sd.sd
+        for key, fused in fused_iter:
-        for key, tensor in fuse_lora_weights(model_sd, lora_sd_and_strengths, dtype):
+            destination_sd.sd[key] = fused
-            sd[key] = tensor
         return destination_sd
 
-    fused = dict(fuse_lora_weights(model_sd, lora_sd_and_strengths, dtype))
+    fused = dict(fused_iter)
     sd = {k: (fused[k] if k in fused else v.clone()) for k, v in model_sd.sd.items()}
     return StateDict(sd, model_sd.device, model_sd.size, model_sd.dtype)
 
 
-def _prepare_deltas(
+def _affected_weight_keys(lora_sd_and_strengths: list[LoraStateDictWithStrength]) -> set[str]:
-    lora_sd_and_strengths: list[LoraStateDictWithStrength], key: str, dtype: torch.dtype, device: torch.device
+    """Return the set of ``.weight`` keys touched by at least one LoRA in the list."""
-) -> torch.Tensor | None:
+    suffix = ".lora_A.weight"
-    deltas = []
+    return {k[: -len(suffix)] + ".weight" for lsd, _ in lora_sd_and_strengths for k in lsd.sd if k.endswith(suffix)}
+
+
+def _products_for_sd_key(
+    lora_sd_and_strengths: list[LoraStateDictWithStrength],
+    key: str,
+    dtype: torch.dtype,
+    device: torch.device,
+) -> Iterator[LoraProduct]:
+    """Yield :class:`LoraProduct` items matching *key* across state-dict-backed LoRAs."""
     prefix = key[: -len(".weight")]
     key_a = f"{prefix}.lora_A.weight"
     key_b = f"{prefix}.lora_B.weight"
     for lsd, coef in lora_sd_and_strengths:
         if key_a not in lsd.sd or key_b not in lsd.sd:
             continue
-        a = lsd.sd[key_a].to(device=device)
+        a = lsd.sd[key_a].to(device=device, dtype=dtype, non_blocking=True)
-        b = lsd.sd[key_b].to(device=device)
+        b = lsd.sd[key_b].to(device=device, dtype=dtype, non_blocking=True)
-        product = torch.matmul(b * coef, a)
+        yield LoraProduct(a, b, coef)
-        del a, b
-        deltas.append(product.to(dtype=dtype))
-    if len(deltas) == 0:
-        return None
-    elif len(deltas) == 1:
-        return deltas[0]
-    return torch.sum(torch.stack(deltas, dim=0), dim=0)
-
-
-def _fuse_delta_with_scaled_fp8(
-    deltas: torch.Tensor,
-    weight: torch.Tensor,
-    key: str,
-    scale_key: str,
-    model_sd: StateDict,
-) -> dict[str, torch.Tensor]:
-    """Dequantize scaled FP8 weight, add LoRA delta, and re-quantize."""
-    weight_scale = model_sd.sd[scale_key]
-
-    original_weight = weight.t().to(torch.float32) * weight_scale
-
-    new_weight = original_weight + deltas.to(torch.float32)
-
-    new_fp8_weight, new_weight_scale = quantize_weight_to_fp8_per_tensor(new_weight)
-    return {key: new_fp8_weight, scale_key: new_weight_scale}
-
-
-def _fuse_delta_with_cast_fp8(
-    deltas: torch.Tensor,
-    weight: torch.Tensor,
-    key: str,
-    target_dtype: torch.dtype,
-) -> dict[str, torch.Tensor]:
-    """Fuse LoRA delta with cast-only FP8 weight (no scale factor)."""
-    if str(weight.device).startswith("cuda"):
-        _fused_add_round_launch(deltas, weight, seed=0)
-    else:
-        deltas.add_(weight.to(dtype=deltas.dtype))
-    return {key: deltas.to(dtype=target_dtype)}
-
-
-def _fuse_delta_with_bfloat16(
-    deltas: torch.Tensor,
-    weight: torch.Tensor,
-    key: str,
-    target_dtype: torch.dtype,
-) -> dict[str, torch.Tensor]:
-    """Fuse LoRA delta with bfloat16 weight."""
-    deltas.add_(weight)
-    return {key: deltas.to(dtype=target_dtype)}

packages/ltx-core/src/ltx_core/loader/kernels.py

@@ -1,10 +1,17 @@
 # ruff: noqa: ANN001, ANN201, ERA001, N803, N806
-import triton
+try:
-import triton.language as tl
+    import triton
+    import triton.language as tl
+
+    TRITON_AVAILABLE = True
+except (ImportError, OSError):
+    TRITON_AVAILABLE = False
 
 
-@triton.jit
+if TRITON_AVAILABLE:
-def fused_add_round_kernel(
+
+    @triton.jit
+    def fused_add_round_kernel(
         x_ptr,
         output_ptr,  # contents will be added to the output
         seed,
@@ -12,7 +19,7 @@ def fused_add_round_kernel(
         EXPONENT_BIAS,
         MANTISSA_BITS,
         BLOCK_SIZE: tl.constexpr,
-):
+    ):
         """
         A kernel to upcast 8bit quantized weights to bfloat16 with stochastic rounding
         and add them to bfloat16 output weights. Might be used to upcast original model weights

packages/ltx-core/src/ltx_core/loader/primitives.py

@@ -1,17 +1,25 @@
 from __future__ import annotations
 
 from dataclasses import dataclass
-from typing import TYPE_CHECKING, NamedTuple, Protocol
+from typing import TYPE_CHECKING, Any, NamedTuple, Protocol, TypeVar
 
 import torch
 
 from ltx_core.loader.module_ops import ModuleOps
 from ltx_core.loader.sd_ops import SDOps
-from ltx_core.model.model_protocol import ModelType
 
 if TYPE_CHECKING:
+    from typing_extensions import Self
+
+    from ltx_core.loader.fuse_loras import FuseRule
     from ltx_core.loader.registry import Registry
 
+BuiltType = TypeVar("BuiltType", covariant=True)  # noqa: PLC0105
+
+
+# Per-key shape and dtype description for a flat collection of tensors.
+TensorLayout = dict[str, tuple[torch.Size, torch.dtype]]
+
 
 @dataclass(frozen=True)
 class StateDict:
@@ -45,62 +53,80 @@ class StateDictLoader(Protocol):
         """
         Load metadata from path
         """
+        ...
 
     def load(self, path: str | list[str], sd_ops: SDOps | None = None, device: torch.device | None = None) -> StateDict:
         """
         Load state dict from path or paths (for sharded model storage) and apply sd_ops
         """
-
-
-class ModelBuilderProtocol(Protocol[ModelType]):
-    """
-    Protocol for building PyTorch models from configuration dictionaries.
-    Implementations must provide:
-    - meta_model: Create a model from configuration dictionary and apply module operations
-    - build: Create and initialize a model from state dictionary and apply dtype transformations
-    """
-
-    model_sd_ops: SDOps | None
-    module_ops: tuple[ModuleOps, ...]
-    loras: tuple["LoraPathStrengthAndSDOps", ...]
-    registry: "Registry"
-
-    def meta_model(self, config: dict, module_ops: list[ModuleOps] | None = None) -> ModelType:
-        """
-        Create a model on the meta device from a configuration dictionary.
-        This decouples model creation from weight loading, allowing the model
-        architecture to be instantiated without allocating memory for parameters.
-        Args:
-            config: Model configuration dictionary.
-            module_ops: Optional list of module operations to apply (e.g., quantization).
-        Returns:
-            Model instance on meta device (no actual memory allocated for parameters).
-        """
         ...
 
-    def with_sd_ops(self, sd_ops: SDOps | None) -> "ModelBuilderProtocol[ModelType]":
-        """Return a copy of this builder with the given state-dict key remapping ops."""
-        ...
 
-    def with_module_ops(self, module_ops: tuple[ModuleOps, ...]) -> "ModelBuilderProtocol[ModelType]":
+class BuilderProtocol(Protocol[BuiltType]):
-        """Return a copy of this builder with the given module operations (e.g. quantization)."""
+    """Protocol for model builders that produce a model via ``build()``."""
-        ...
 
-    def with_loras(self, loras: tuple["LoraPathStrengthAndSDOps", ...]) -> "ModelBuilderProtocol[ModelType]":
+    def build(
-        """Return a copy of this builder with the given LoRAs to fuse at build time."""
+        self,
-        ...
+        device: torch.device | None = None,
+        dtype: torch.dtype | None = None,
+        **kwargs: Any,  # noqa: ANN401
+    ) -> BuiltType: ...
 
-    def with_registry(self, registry: "Registry") -> "ModelBuilderProtocol[ModelType]":
+    @property
+    def registry(self) -> "Registry": ...
+
+    def with_registry(self, registry: "Registry") -> "Self":
         """Return a copy of this builder using the given weight registry for allocation."""
         ...
 
-    def with_lora_load_device(self, device: torch.device) -> "ModelBuilderProtocol[ModelType]":
+
+class ModelBuilderProtocol(BuilderProtocol[BuiltType], Protocol[BuiltType]):
+    """
+    Protocol for building PyTorch models from configuration dictionaries.
+    Implementations must provide:
+    - build: Create and initialize a model from state dictionary and apply dtype transformations
+    """
+
+    @property
+    def checkpoint(self) -> str | tuple[str, ...]:
+        """Path(s) to the checkpoint this builder loads from (for logging/diagnostics)."""
+        ...
+
+    @property
+    def model_sd_ops(self) -> SDOps | None: ...
+
+    @property
+    def module_ops(self) -> tuple[ModuleOps, ...]: ...
+
+    @property
+    def loras(self) -> tuple["LoraPathStrengthAndSDOps", ...]: ...
+
+    def with_sd_ops(self, sd_ops: SDOps | None) -> "Self":
+        """Return a copy of this builder with the given state-dict key remapping ops."""
+        ...
+
+    def with_module_ops(self, module_ops: tuple[ModuleOps, ...]) -> "Self":
+        """Return a copy of this builder with the given module operations (e.g. quantization)."""
+        ...
+
+    def with_loras(self, loras: tuple["LoraPathStrengthAndSDOps", ...]) -> "Self":
+        """Return a copy of this builder with the given LoRAs to fuse at build time."""
+        ...
+
+    def with_lora_load_device(self, device: torch.device) -> "Self":
         """Return a copy of this builder that loads LoRA weights onto the given device."""
         ...
 
+    def with_fuse_rule(self, fuse_rule: "FuseRule") -> "Self":
+        """Return a copy of this builder with the given LoRA fuse rule (e.g. from a quantization policy)."""
+        ...
+
     def build(
-        self, device: torch.device | None = None, dtype: torch.dtype | None = None, **kwargs: object
+        self,
-    ) -> ModelType:
+        device: torch.device | None = None,
+        dtype: torch.dtype | None = None,
+        **kwargs: Any,  # noqa: ANN401
+    ) -> BuiltType:
         """
         Build the model
         Args:
@@ -123,8 +149,7 @@ class LoRAAdaptableProtocol(Protocol):
     - lora: Add a LoRA to the model
     """
 
-    def lora(self, lora_path: str, strength: float) -> "LoRAAdaptableProtocol":
+    def lora(self, lora_path: str, strength: float, sd_ops: SDOps) -> "LoRAAdaptableProtocol": ...
-        pass
 
 
 class LoraPathStrengthAndSDOps(NamedTuple):

packages/ltx-core/src/ltx_core/loader/sd_ops.py

@@ -24,6 +24,7 @@ class ContentMatching:
 
     prefix: str = ""
     suffix: str = ""
+    contains: str = ""
 
 
 class KeyValueOperationResult(NamedTuple):
@@ -72,10 +73,10 @@ class SDOps:
         new_mapping = (*self.mapping, ContentReplacement(content, replacement))
         return replace(self, mapping=new_mapping)
 
-    def with_matching(self, prefix: str = "", suffix: str = "") -> "SDOps":
+    def with_matching(self, prefix: str = "", suffix: str = "", contains: str = "") -> "SDOps":
-        """Create a new SDOps instance with the specified prefix and suffix matching added to the mapping."""
+        """Create a new SDOps instance with the specified prefix, suffix and contains matching added to the mapping."""
 
-        new_mapping = (*self.mapping, ContentMatching(prefix, suffix))
+        new_mapping = (*self.mapping, ContentMatching(prefix, suffix, contains))
         return replace(self, mapping=new_mapping)
 
     def with_additional_allowed_keys(self, keys: frozenset[str]) -> "SDOps":
@@ -100,7 +101,12 @@ class SDOps:
     def apply_to_key(self, key: str) -> str | None:
         """Apply the mapping to the given name."""
         matchers = [content for content in self.mapping if isinstance(content, ContentMatching)]
-        valid = any(key.startswith(f.prefix) and key.endswith(f.suffix) for f in matchers)
+        valid = any(
+            key.startswith(matcher.prefix)
+            and key.endswith(matcher.suffix)
+            and (not matcher.contains or matcher.contains in key)
+            for matcher in matchers
+        )
         if not valid:
             return None
 

packages/ltx-core/src/ltx_core/loader/single_gpu_model_builder.py

@@ -1,11 +1,13 @@
+from __future__ import annotations
+
+import copy
 import logging
-from dataclasses import dataclass, field, replace
+from typing import TYPE_CHECKING, Final, Generic
-from typing import Generic
 
 import torch
 from torch import nn
 
-from ltx_core.loader.fuse_loras import apply_loras
+from ltx_core.loader.fuse_loras import FuseRule, apply_loras, bf16_fuse_rule
 from ltx_core.loader.helpers import create_meta_model, load_state_dict, read_model_config
 from ltx_core.loader.module_ops import ModuleOps
 from ltx_core.loader.primitives import (
@@ -21,6 +23,9 @@ from ltx_core.loader.sd_ops import SDOps
 from ltx_core.loader.sft_loader import SafetensorsModelStateDictLoader
 from ltx_core.model.model_protocol import ModelConfigurator, ModelType
 
+if TYPE_CHECKING:
+    from typing_extensions import Self
+
 logger: logging.Logger = logging.getLogger(__name__)
 
 
@@ -46,6 +51,7 @@ def _load_model_weights(
     dtype: torch.dtype | None,
     model_sd_ops: SDOps | None = None,
     lora_load_device: torch.device | None = None,
+    fuse_rule: FuseRule = bf16_fuse_rule,
 ) -> None:
     """Load base weights and fuse LoRAs into *meta_model* in-place."""
     if lora_load_device is None:
@@ -57,7 +63,7 @@ def _load_model_weights(
     if not lora_strengths or (min(lora_strengths) == 0 and max(lora_strengths) == 0):
         sd = model_sd.sd
         if dtype is not None:
-            sd = {key: value.to(dtype=dtype) for key, value in model_sd.sd.items()}
+            sd = {key: value.to(dtype=dtype) for key, value in sd.items()}
         meta_model.load_state_dict(sd, strict=False, assign=True)
         return
 
@@ -68,16 +74,21 @@ def _load_model_weights(
     final_sd = apply_loras(
         model_sd=model_sd,
         lora_sd_and_strengths=lora_sd_and_strengths,
-        dtype=dtype,
+        fuse_rule=fuse_rule,
         destination_sd=model_sd if isinstance(registry, DummyRegistry) else None,
     )
-    meta_model.load_state_dict(final_sd.sd, strict=False, assign=True)
+    fused_sd = final_sd.sd
+    if dtype is not None:
+        fused_sd = {key: value.to(dtype=dtype) for key, value in fused_sd.items()}
+    meta_model.load_state_dict(fused_sd, strict=False, assign=True)
 
 
-@dataclass(frozen=True)
 class SingleGPUModelBuilder(Generic[ModelType], ModelBuilderProtocol[ModelType], LoRAAdaptableProtocol):
     """
     Builder for PyTorch models residing on a single GPU.
+    The builder is immutable: ``with_*``/``lora`` return modified copies. The
+    ``ModelBuilderProtocol`` surface is exposed via read-only properties backed
+    by private attributes.
     Attributes:
         model_class_configurator: Class responsible for constructing the model from a config dict.
         model_path: Path (or tuple of shard paths) to the model's `.safetensors` checkpoint(s).
@@ -91,52 +102,113 @@ class SingleGPUModelBuilder(Generic[ModelType], ModelBuilderProtocol[ModelType],
             ``torch.device("cpu")``, which keeps LoRA weights in CPU memory and transfers them to
             the target GPU sequentially during fusion, reducing peak GPU memory usage compared to
             loading all LoRA weights directly onto the GPU at once.
+        fuse_rule: Per-policy LoRA merge rule. Defaults to ``bf16_fuse_rule``;
     """
 
-    model_class_configurator: type[ModelConfigurator[ModelType]]
+    def __init__(
-    model_path: str | tuple[str, ...]
+        self,
-    model_sd_ops: SDOps | None = None
+        model_class_configurator: type[ModelConfigurator[ModelType]],
-    module_ops: tuple[ModuleOps, ...] = field(default_factory=tuple)
+        model_path: str | tuple[str, ...],
-    loras: tuple[LoraPathStrengthAndSDOps, ...] = field(default_factory=tuple)
+        model_sd_ops: SDOps | None = None,
-    model_loader: StateDictLoader = field(default_factory=SafetensorsModelStateDictLoader)
+        module_ops: tuple[ModuleOps, ...] = (),
-    registry: Registry = field(default_factory=DummyRegistry)
+        loras: tuple[LoraPathStrengthAndSDOps, ...] = (),
-    lora_load_device: torch.device = field(default_factory=lambda: torch.device("cpu"))
+        model_loader: StateDictLoader | None = None,
+        registry: Registry | None = None,
+        lora_load_device: torch.device | None = None,
+        fuse_rule: FuseRule = bf16_fuse_rule,
+    ) -> None:
+        # Read-only: typed with the covariant ModelType, so it must not be a mutable attribute.
+        self._model_class_configurator: Final = model_class_configurator
+        self._model_path = model_path
+        self._model_sd_ops = model_sd_ops
+        self._module_ops = module_ops
+        self._loras = loras
+        self._model_loader = model_loader if model_loader is not None else SafetensorsModelStateDictLoader()
+        self._registry = registry if registry is not None else DummyRegistry()
+        self._lora_load_device = lora_load_device if lora_load_device is not None else torch.device("cpu")
+        self._fuse_rule = fuse_rule
 
-    def lora(self, lora_path: str, strength: float = 1.0, sd_ops: SDOps | None = None) -> "SingleGPUModelBuilder":
+    @property
-        return replace(self, loras=(*self.loras, LoraPathStrengthAndSDOps(lora_path, strength, sd_ops)))
+    def model_sd_ops(self) -> SDOps | None:
+        return self._model_sd_ops
 
-    def with_sd_ops(self, sd_ops: SDOps | None) -> "SingleGPUModelBuilder":
+    @property
-        return replace(self, model_sd_ops=sd_ops)
+    def module_ops(self) -> tuple[ModuleOps, ...]:
+        return self._module_ops
 
-    def with_module_ops(self, module_ops: tuple[ModuleOps, ...]) -> "SingleGPUModelBuilder":
+    @property
-        return replace(self, module_ops=module_ops)
+    def loras(self) -> tuple[LoraPathStrengthAndSDOps, ...]:
+        return self._loras
 
-    def with_loras(self, loras: tuple[LoraPathStrengthAndSDOps, ...]) -> "SingleGPUModelBuilder":
+    @property
-        return replace(self, loras=loras)
+    def registry(self) -> Registry:
+        return self._registry
 
-    def with_registry(self, registry: Registry) -> "SingleGPUModelBuilder":
+    @property
-        return replace(self, registry=registry)
+    def model_path(self) -> str | tuple[str, ...]:
+        return self._model_path
 
-    def with_lora_load_device(self, device: torch.device) -> "SingleGPUModelBuilder":
+    @property
-        return replace(self, lora_load_device=device)
+    def checkpoint(self) -> str | tuple[str, ...]:
+        return self._model_path
+
+    @property
+    def model_loader(self) -> StateDictLoader:
+        return self._model_loader
+
+    @property
+    def lora_load_device(self) -> torch.device:
+        return self._lora_load_device
+
+    @property
+    def fuse_rule(self) -> FuseRule:
+        return self._fuse_rule
+
+    def lora(self, lora_path: str, strength: float, sd_ops: SDOps) -> Self:
+        clone = copy.copy(self)
+        clone._loras = (*self._loras, LoraPathStrengthAndSDOps(lora_path, strength, sd_ops))
+        return clone
+
+    def with_sd_ops(self, sd_ops: SDOps | None) -> Self:
+        clone = copy.copy(self)
+        clone._model_sd_ops = sd_ops
+        return clone
+
+    def with_module_ops(self, module_ops: tuple[ModuleOps, ...]) -> Self:
+        clone = copy.copy(self)
+        clone._module_ops = module_ops
+        return clone
+
+    def with_loras(self, loras: tuple[LoraPathStrengthAndSDOps, ...]) -> Self:
+        clone = copy.copy(self)
+        clone._loras = loras
+        return clone
+
+    def with_registry(self, registry: Registry) -> Self:
+        clone = copy.copy(self)
+        clone._registry = registry
+        return clone
+
+    def with_lora_load_device(self, device: torch.device) -> Self:
+        clone = copy.copy(self)
+        clone._lora_load_device = device
+        return clone
+
+    def with_fuse_rule(self, fuse_rule: FuseRule) -> Self:
+        clone = copy.copy(self)
+        clone._fuse_rule = fuse_rule
+        return clone
 
     def model_config(self) -> dict:
-        return read_model_config(self.model_path, self.model_loader)
+        return read_model_config(self._model_path, self._model_loader)
 
     def meta_model(self, config: dict, module_ops: tuple[ModuleOps, ...]) -> ModelType:
-        return create_meta_model(self.model_class_configurator, config, module_ops)
+        return create_meta_model(self._model_class_configurator, config, module_ops)
 
     def load_sd(
         self, paths: list[str], registry: Registry, device: torch.device | None, sd_ops: SDOps | None = None
     ) -> StateDict:
-        return load_state_dict(paths, self.model_loader, registry, device, sd_ops)
+        return load_state_dict(paths, self._model_loader, registry, device, sd_ops)
-
-    def _return_model(self, meta_model: ModelType, device: torch.device) -> ModelType:
-        uninitialized = _check_uninitialized(meta_model)
-        if uninitialized:
-            logger.warning(f"Uninitialized parameters or buffers: {uninitialized}")
-            return meta_model
-        return meta_model.to(device)
 
     def build(
         self,
@@ -146,17 +218,23 @@ class SingleGPUModelBuilder(Generic[ModelType], ModelBuilderProtocol[ModelType],
     ) -> ModelType:
         device = torch.device("cuda") if device is None else device
         config = self.model_config()
-        meta_model = self.meta_model(config, self.module_ops)
+        meta_model = self.meta_model(config, self._module_ops)
 
         _load_model_weights(
             meta_model=meta_model,
-            model_path=self.model_path,
+            model_path=self._model_path,
-            loras=self.loras,
+            loras=self._loras,
-            loader=self.model_loader,
+            loader=self._model_loader,
-            registry=self.registry,
+            registry=self._registry,
             device=device,
             dtype=dtype,
-            model_sd_ops=self.model_sd_ops,
+            model_sd_ops=self._model_sd_ops,
-            lora_load_device=self.lora_load_device,
+            lora_load_device=self._lora_load_device,
+            fuse_rule=self._fuse_rule,
         )
-        return self._return_model(meta_model, device)
+
+        uninitialized = _check_uninitialized(meta_model)
+        if uninitialized:
+            logger.warning(f"Uninitialized parameters or buffers: {uninitialized}")
+            return meta_model
+        return meta_model.to(device)

packages/ltx-core/src/ltx_core/modality_tiling.py

@@ -21,11 +21,14 @@ from ltx_core.types import VideoLatentShape
 @dataclass(frozen=True)
 class TilingContext:
     """Opaque context produced by :meth:`VideoModalityTilingHelper.tile_modality`.
-    Carries the token-level keep mask and per-conditioning-token blend
+    Carries the token-level keep indices and per-conditioning-token blend
     weights needed by :meth:`~VideoModalityTilingHelper.blend`.
     """
 
-    keep_mask: torch.Tensor
+    keep_indices: torch.Tensor
+    """``(num_kept,)`` int64 — sorted indices of tokens the tile processes."""
+    num_total_tokens: int
+    """Total number of tokens in the full (untiled) sequence."""
     cond_blend_weights: torch.Tensor | None
     """``(num_kept_cond,)`` — weight for each kept conditioning token,
     equal to ``1 / num_tiles_that_keep_this_token``.  ``None`` when
@@ -81,14 +84,32 @@ class VideoModalityTilingHelper:
             A ``(tiled_modality, context)`` tuple.  Pass *context* to
             :meth:`blend` together with the model output.
         """
-        keep_mask = self._keep_mask(modality, tile)
+        device = modality.positions.device
+        gen_indices = self._generated_token_indices(tile, device=device)
+        num_total = modality.latent.shape[1]
+
+        cond_blend_weights: torch.Tensor | None = None
+        if num_total > self._num_generated_tokens:
+            keep_per_tile_cond = self._all_tiles_cond_keep(modality)  # (num_tiles, num_cond) bool
+            tile_idx = next((i for i, t in enumerate(self._tiles) if t.in_coords == tile.in_coords), None)
+            if tile_idx is None:
+                raise RuntimeError(
+                    f"Tile with in_coords={tile.in_coords} is not in this helper's tile set; "
+                    f"pass a tile obtained from `helper.tiles`."
+                )
+            my_cond_keep = keep_per_tile_cond[tile_idx]
+            cond_indices = self._num_generated_tokens + my_cond_keep.nonzero(as_tuple=False).squeeze(1)
+            keep_indices = torch.cat([gen_indices, cond_indices])
+            total_keepers = keep_per_tile_cond.sum(dim=0).float()  # (num_cond,)
+            cond_blend_weights = 1.0 / total_keepers[my_cond_keep]
+        else:
+            keep_indices = gen_indices
 
         tile_attention_mask = None
         if modality.attention_mask is not None:
-            keep_indices = keep_mask.nonzero(as_tuple=False).squeeze(1)
             tile_attention_mask = modality.attention_mask[:, keep_indices, :][:, :, keep_indices]
 
-        positions = modality.positions[:, :, keep_mask, :]
+        positions = modality.positions[:, :, keep_indices, :]
         if normalize_positions:
             num_tile_gen = self._tile_generated_token_count(tile)
             gen_pos = positions[:, :, :num_tile_gen, :]  # (B, 3, num_tile_gen, 2)
@@ -97,26 +118,15 @@ class VideoModalityTilingHelper:
 
         tiled = replace(
             modality,
-            latent=modality.latent[:, keep_mask, :],
+            latent=modality.latent[:, keep_indices, :],
-            timesteps=modality.timesteps[:, keep_mask],
+            timesteps=modality.timesteps[:, keep_indices],
             positions=positions,
             attention_mask=tile_attention_mask,
         )
 
-        cond_blend_weights = None
+        return tiled, TilingContext(
-        num_total = modality.latent.shape[1]
+            keep_indices=keep_indices, num_total_tokens=num_total, cond_blend_weights=cond_blend_weights
-        if num_total > self._num_generated_tokens:
+        )
-            cond_keep = keep_mask[self._num_generated_tokens :]
-            # Count how many tiles keep each conditioning token.
-            cond_counts = torch.zeros(cond_keep.sum(), dtype=torch.float32)
-            for t in self._tiles:
-                other_mask = self._keep_mask(modality, t)
-                other_cond = other_mask[self._num_generated_tokens :]
-                # Map other tile's kept cond tokens into this tile's kept subset.
-                cond_counts += other_cond[cond_keep].float()
-            cond_blend_weights = 1.0 / cond_counts
-
-        return tiled, TilingContext(keep_mask=keep_mask, cond_blend_weights=cond_blend_weights)
 
     # -- blend -------------------------------------------------------------
 
@@ -147,14 +157,14 @@ class VideoModalityTilingHelper:
         """
         batch, _, dim = tile_to_blend.shape
         num_tile_gen = self._tile_generated_token_count(tile)
-        gen_indices = self._generated_token_indices(tile)
+        gen_indices = self._generated_token_indices(tile, device=tile_to_blend.device)
 
-        num_total_tokens = context.keep_mask.shape[0]
+        num_total_tokens = context.num_total_tokens
         expected_shape = (batch, num_total_tokens, dim)
 
         if output is not None:
             if output.shape != expected_shape:
-                raise ValueError(f"Expected output shape {expected_shape}, got {output.shape}")
+                raise RuntimeError(f"Expected output shape {expected_shape}, got {output.shape}")
             result = output
         else:
             result = torch.zeros(*expected_shape, device=tile_to_blend.device, dtype=tile_to_blend.dtype)
@@ -168,8 +178,7 @@ class VideoModalityTilingHelper:
         # Scatter kept conditioning tokens, weighted by 1/N where N is
         # the number of tiles that keep each token (so they sum to 1).
         if num_total_tokens > self._num_generated_tokens and context.cond_blend_weights is not None:
-            cond_keep = context.keep_mask[self._num_generated_tokens :]
+            cond_indices = context.keep_indices[context.keep_indices >= self._num_generated_tokens]
-            cond_indices = self._num_generated_tokens + cond_keep.nonzero(as_tuple=False).squeeze(1)
             weights = context.cond_blend_weights.to(device=tile_to_blend.device, dtype=tile_to_blend.dtype)
             result[:, cond_indices, :] += tile_to_blend[:, num_tile_gen:, :] * weights[None, :, None]
 
@@ -189,46 +198,42 @@ class VideoModalityTilingHelper:
         )
         return self._patchifier.get_token_count(tile_shape)
 
-    def _generated_token_indices(self, tile: Tile) -> torch.Tensor:
+    def _generated_token_indices(self, tile: Tile, device: torch.device | None = None) -> torch.Tensor:
         """Flat token indices of *tile*'s generated tokens in the full sequence."""
         frame_slice, height_slice, width_slice = tile.in_coords
-        f = torch.arange(frame_slice.start, frame_slice.stop)
+        f = torch.arange(frame_slice.start, frame_slice.stop, device=device)
-        h = torch.arange(height_slice.start, height_slice.stop)
+        h = torch.arange(height_slice.start, height_slice.stop, device=device)
-        w = torch.arange(width_slice.start, width_slice.stop)
+        w = torch.arange(width_slice.start, width_slice.stop, device=device)
         return (
             f[:, None, None] * self._latent_shape.height * self._latent_shape.width
             + h[None, :, None] * self._latent_shape.width
             + w[None, None, :]
         ).reshape(-1)
 
-    def _keep_mask(self, modality: Modality, tile: Tile) -> torch.Tensor:
+    def _all_tiles_cond_keep(self, modality: Modality) -> torch.Tensor:
-        """Boolean mask ``(num_total_tokens,)`` — True for tokens the tile processes.
+        """Vectorized (num_tiles, num_cond) bool: which tiles keep each conditioning token.
-        Generated tokens are selected by grid position.  Conditioning
+        A conditioning token is kept by a tile when its ``[start, end)`` interval
-        tokens are kept when their ``[start, end)`` intervals overlap
+        overlaps the tile in all three dimensions, or when it has a negative time
-        the tile in all three dimensions, or when they have a negative
+        coordinate (reference token).
-        time coordinate (reference tokens).
         """
-        num_total = modality.latent.shape[1]
-        mask = torch.zeros(num_total, dtype=torch.bool)
-
-        gen_indices = self._generated_token_indices(tile)
-        mask[gen_indices] = True
-
-        if num_total > self._num_generated_tokens:
-            gen_positions = modality.positions[:, :, gen_indices, :]  # (B, 3, num_tile_gen, 2)
-            tile_start = gen_positions[..., 0].amin(dim=2)  # (B, 3)
-            tile_end = gen_positions[..., 1].amax(dim=2)  # (B, 3)
-
         cond_positions = modality.positions[:, :, self._num_generated_tokens :, :]  # (B, 3, num_cond, 2)
+        device = cond_positions.device
 
-            overlaps = (cond_positions[..., 0] < tile_end.unsqueeze(2)) & (
+        # Per-tile (start, end) bounds along each axis; small Python loop (num_tiles <= ~16).
-                cond_positions[..., 1] > tile_start.unsqueeze(2)
+        starts_list: list[torch.Tensor] = []
-            )  # (B, 3, num_cond)
+        ends_list: list[torch.Tensor] = []
-            overlaps_all_dims = overlaps.all(dim=1)  # (B, num_cond)
+        for t in self._tiles:
+            gen_idx = self._generated_token_indices(t, device=device)
+            gen_positions = modality.positions[:, :, gen_idx, :]  # (B, 3, num_tile_gen, 2)
+            starts_list.append(gen_positions[..., 0].amin(dim=2))  # (B, 3)
+            ends_list.append(gen_positions[..., 1].amax(dim=2))  # (B, 3)
+        tile_starts = torch.stack(starts_list, dim=0)  # (num_tiles, B, 3)
+        tile_ends = torch.stack(ends_list, dim=0)  # (num_tiles, B, 3)
 
-            has_negative_time = cond_positions[:, 0, :, 0] < 0  # (B, num_cond)
+        cond_starts = cond_positions[..., 0]  # (B, 3, num_cond)
-
+        cond_ends = cond_positions[..., 1]  # (B, 3, num_cond)
-            keep_cond = (overlaps_all_dims | has_negative_time).any(dim=0)  # (num_cond,)
+        # Broadcast: (1, B, 3, num_cond) vs (num_tiles, B, 3, 1) -> (num_tiles, B, 3, num_cond).
-            mask[self._num_generated_tokens :] = keep_cond
+        overlaps = (cond_starts[None] < tile_ends[..., None]) & (cond_ends[None] > tile_starts[..., None])
-
+        overlaps_all_dims = overlaps.all(dim=2)  # (num_tiles, B, num_cond)
-        return mask
+        has_negative_time = (cond_positions[:, 0, :, 0] < 0)[None]  # (1, B, num_cond)
+        return (overlaps_all_dims | has_negative_time).any(dim=1)  # (num_tiles, num_cond)

packages/ltx-core/src/ltx_core/model/audio_vae/vocoder.py

@@ -1,4 +1,6 @@
+import contextlib
 import math
+from collections.abc import Iterator
 from typing import List
 
 import einops
@@ -13,6 +15,25 @@ def get_padding(kernel_size: int, dilation: int = 1) -> int:
     return int((kernel_size * dilation - dilation) / 2)
 
 
+@contextlib.contextmanager
+def _module_in_fp32(module: nn.Module, *, enabled: bool) -> Iterator[None]:
+    """Temporarily cast *module* to float32, restoring its original dtype on exit.
+    Used for the MPS vocoder path where fp32 autocast is unavailable, so the
+    weights must be materialized in float32 for the forward pass. Restores to the
+    module's original weight dtype (captured here), not the input dtype. When
+    *enabled* is False this is a no-op, so callers can wrap unconditionally.
+    """
+    if not enabled:
+        yield
+        return
+    module_dtype = next(module.parameters()).dtype
+    module.float()
+    try:
+        yield
+    finally:
+        module.to(module_dtype)
+
+
 # ---------------------------------------------------------------------------
 # Anti-aliased resampling helpers (kaiser-sinc filters) for BigVGAN v2
 # Adopted from https://github.com/NVIDIA/BigVGAN
@@ -564,15 +585,29 @@ class VocoderWithBWE(nn.Module):
         # compound through 108 sequential convolutions and degrade spectral
         # metrics (mel_l1, MRSTFT) by 40-90% while perceptual quality (CDPAM)
         # is unaffected.  fp32 eliminates this degradation.
-        # We use autocast(dtype=float32) rather than self.float() because it
+        # On CUDA/CPU we use autocast(dtype=float32) rather than self.float()
-        # upcasts bf16 weights per-op at kernel level, avoiding the temporary
+        # because it upcasts bf16 weights per-op at kernel level, avoiding the
-        # memory spike of self.float() / self.to(original_dtype).
+        # temporary memory spike of self.float() / self.to(original_dtype).
         # Benchmarked on H100 (128.5M-param model):
         #   autocast fp32: +70 MB peak VRAM, 123 ms  (vs 482 MB / 95 ms for bf16)
         #   model.float(): +324 MB peak VRAM, 149 ms
         # Tested: both approaches produce bit-identical output.
+        # MPS autocast does not upcast conv weights to fp32 (it only supports
+        # lower-precision autocast dtypes), which would leave the float32 input
+        # running against bf16 conv weights and raise a dtype mismatch. There we
+        # fall back to materializing the weights in fp32 for the pass (bit-identical
+        # per the note above; the memory spike is negligible for this small model).
+        # The vocoder is normally built in fp32 on MPS, so this fallback is then a
+        # no-op -- it only triggers if a bf16 module is run on MPS directly.
+        device_type = mel_spec.device.type
+        module_dtype = next(self.parameters()).dtype
+        fp32_ctx = (
+            _module_in_fp32(self, enabled=module_dtype != torch.float32)
+            if device_type == "mps"
+            else torch.autocast(device_type=device_type, dtype=torch.float32)
+        )
 
-        with torch.autocast(device_type=mel_spec.device.type, dtype=torch.float32):
+        with fp32_ctx:
             x = self.vocoder(mel_spec.float())
             _, _, length_low_rate = x.shape
             output_length = length_low_rate * self.output_sampling_rate // self.input_sampling_rate

packages/ltx-core/src/ltx_core/model/model_protocol.py

@@ -1,6 +1,14 @@
-from typing import Protocol, TypeVar
+from __future__ import annotations
 
-ModelType = TypeVar("ModelType")
+from typing import TYPE_CHECKING, Protocol, TypeVar
+
+import torch
+
+if TYPE_CHECKING:
+    from ltx_core.guidance.perturbations import BatchedPerturbationConfig
+    from ltx_core.model.transformer.modality import Modality
+
+ModelType = TypeVar("ModelType", covariant=True, bound=torch.nn.Module)  # noqa: PLC0105
 
 
 class ModelConfigurator(Protocol[ModelType]):
@@ -8,3 +16,29 @@ class ModelConfigurator(Protocol[ModelType]):
 
     @classmethod
     def from_config(cls, config: dict) -> ModelType: ...
+
+
+class LTXModelProtocol(Protocol):
+    """Velocity-model forward interface shared by ``LTXModel`` and its multi-GPU wrappers.
+    ``forward`` pins the real signature (enforced structurally); ``__call__`` mirrors it
+    so protocol-typed values stay callable via ``model(...)``.
+    """
+
+    @property
+    def num_blocks(self) -> int:
+        """Number of transformer blocks, delegated through any wrappers to the ``LTXModel``."""
+        ...
+
+    def forward(
+        self,
+        video: Modality | None,
+        audio: Modality | None,
+        perturbations: BatchedPerturbationConfig | None,
+    ) -> tuple[torch.Tensor | None, torch.Tensor | None]: ...
+
+    def __call__(
+        self,
+        video: Modality | None,
+        audio: Modality | None,
+        perturbations: BatchedPerturbationConfig | None,
+    ) -> tuple[torch.Tensor | None, torch.Tensor | None]: ...

packages/ltx-core/src/ltx_core/model/transformer/__init__.py

@@ -3,13 +3,17 @@
 from ltx_core.model.transformer.modality import Modality
 from ltx_core.model.transformer.model import LTXModel, X0Model
 from ltx_core.model.transformer.model_configurator import (
+    LTXV_AUDIO_ONLY_MODEL_COMFY_RENAMING_MAP,
     LTXV_MODEL_COMFY_RENAMING_MAP,
+    LTXAudioOnlyModelConfigurator,
     LTXModelConfigurator,
     LTXVideoOnlyModelConfigurator,
 )
 
 __all__ = [
+    "LTXV_AUDIO_ONLY_MODEL_COMFY_RENAMING_MAP",
     "LTXV_MODEL_COMFY_RENAMING_MAP",
+    "LTXAudioOnlyModelConfigurator",
     "LTXModel",
     "LTXModelConfigurator",
     "LTXVideoOnlyModelConfigurator",

packages/ltx-core/src/ltx_core/model/transformer/attention.py

@@ -1,31 +1,81 @@
+import functools
+import sys
+from dataclasses import dataclass, field
 from enum import Enum
 from typing import Protocol
 
 import torch
+from torch.nn.attention import SDPBackend, sdpa_kernel
+
+from ltx_core.model.transformer.ops import (
+    GatedAttentionCallable,
+    PreAttentionCallable,
+    PytorchGatedAttention,
+    PytorchPreAttention,
+)
+from ltx_core.model.transformer.rope import LTXRopeType
+
+
+def _torch_default_sdpa_priority() -> list[SDPBackend]:
+    """Fetch torch's current default SDPA priority order at runtime.
+    Used as the default for ``PytorchAttention`` so the wrapper-always
+    code path matches torch's native dispatch order without hard-coding it
+    (which would drift if torch updates the default).
+    ``torch._C._get_sdp_priority_order`` is a private API; we accept that
+    risk because the project pins ``torch`` in the lockfile, so any
+    rename/removal surfaces on a controlled torch bump rather than silently.
+    """
+    return [SDPBackend(p) for p in torch._C._get_sdp_priority_order()]
 
-from ltx_core.model.transformer.rope import LTXRopeType, apply_rotary_emb
 
-memory_efficient_attention = None
 flash_attn_interface = None
+flash_attn_4_func = None
 try:
-    from xformers.ops import memory_efficient_attention
-except ImportError:
-    memory_efficient_attention = None
-try:
-    # FlashAttention3 and XFormersAttention cannot be used together
-    if memory_efficient_attention is None:
     import flash_attn_interface
 except ImportError:
     flash_attn_interface = None
+try:
+    from flash_attn.cute import flash_attn_func as flash_attn_4_func
+except ImportError:
+    flash_attn_4_func = None
+try:
+    # macOS only: routes SDPA to Apple's prebuilt MPSGraph attention kernel.
+    from mps_sdpa import sdpa_opt as _mps_sdpa_opt
+except ImportError:
+    _mps_sdpa_opt = None
 
 
 class AttentionCallable(Protocol):
+    """Unmasked attention. Backends without a mask kernel (FA3/FA4) implement only
+    this protocol; backends that support masks too (Pytorch/SDPA) are
+    structurally usable here and as :class:`MaskedAttentionCallable`."""
+
+    def __call__(self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, heads: int) -> torch.Tensor: ...
+
+
+class MaskedAttentionCallable(Protocol):
+    """Masked attention. Mask is required (not optional) -- the caller has already
+    decided this is the masked path and chosen a backend that can serve it. Used
+    by :class:`Attention` when its forward receives a non-None ``mask``."""
+
     def __call__(
-        self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, heads: int, mask: torch.Tensor | None = None
+        self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, heads: int, mask: torch.Tensor
     ) -> torch.Tensor: ...
 
 
 class PytorchAttention(AttentionCallable):
+    def __init__(self, priority: list[SDPBackend] | None = None) -> None:
+        # priority=None -> snapshot torch's default SDPA priority at construction.
+        # Always passed through ``sdpa_kernel(..., set_priority=True)`` so the
+        # call site is uniform regardless of how the priority was chosen.
+        self._priority = priority if priority is not None else _torch_default_sdpa_priority()
+
+    @property
+    def label(self) -> str:
+        """Human-readable identifier. Encodes the SDPA priority list so a
+        single-backend pin reads differently from the full-priority dispatcher walk."""
+        return f"SDPA[{'>'.join(b.name for b in self._priority)}]"
+
     def __call__(
         self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, heads: int, mask: torch.Tensor | None = None
     ) -> torch.Tensor:
@@ -41,99 +91,385 @@ class PytorchAttention(AttentionCallable):
             if mask.ndim == 3:
                 mask = mask.unsqueeze(1)
 
-        out = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=mask, dropout_p=0.0, is_causal=False)
+        with sdpa_kernel(self._priority, set_priority=True):
+            out = torch.nn.functional.scaled_dot_product_attention(
+                q, k, v, attn_mask=mask, dropout_p=0.0, is_causal=False
+            )
         out = out.transpose(1, 2).reshape(b, -1, heads * dim_head)
         return out
 
 
-class XFormersAttention(AttentionCallable):
+class MPSSdpaAttention(AttentionCallable):
+    """Apple-fused scaled-dot-product attention on MPS.
+    Routes to ``mps_sdpa.sdpa_opt``, which calls Apple's prebuilt
+    ``MPSGraph.scaledDotProductAttention`` kernel (via a zero-copy bridge)
+    instead of torch's ``sdpa_general_mps`` graph. The Apple kernel does not
+    materialize the ``[B, H, Nq, Nk]`` score matrix, so it avoids the
+    long-sequence memory wall that makes torch's materializing MPS SDPA
+    unusable on video latents (~32x faster at a 14k-token latent on an M4 Pro).
+    It is a hard dependency on Apple Silicon (the ``mps-sdpa`` platform-marked
+    requirement), so AUTOMATIC always has it on MPS. Unlike a JIT-compiled Metal
+    flash kernel it needs no runtime shader compilation, so it is robust
+    across macOS / Metal revisions.
+    Accepts an optional additive-float or boolean ``mask`` broadcastable to
+    ``[B, H, Nq, Nk]``, so it serves both the unmasked and masked protocols.
+    """
+
+    @property
+    def label(self) -> str:
+        return "MPS-SDPA"
+
     def __call__(
-        self,
+        self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, heads: int, mask: torch.Tensor | None = None
-        q: torch.Tensor,
-        k: torch.Tensor,
-        v: torch.Tensor,
-        heads: int,
-        mask: torch.Tensor | None = None,
     ) -> torch.Tensor:
-        if memory_efficient_attention is None:
+        if _mps_sdpa_opt is None:
-            raise RuntimeError("XFormersAttention was selected but `xformers` is not installed.")
+            raise RuntimeError("MPSSdpaAttention was selected but `mps-sdpa` is not installed.")
+        if q.device.type != "mps":
+            raise RuntimeError("MPSSdpaAttention requires MPS. Use PyTorch SDPA on CPU or CUDA.")
 
         b, _, dim_head = q.shape
         dim_head //= heads
-
+        q, k, v = (t.view(b, -1, heads, dim_head).transpose(1, 2) for t in (q, k, v))
-        # xformers expects [B, M, H, K]
-        q, k, v = (t.view(b, -1, heads, dim_head) for t in (q, k, v))
 
         if mask is not None:
-            # add a singleton batch dimension
+            # add a batch dimension if there isn't already one
             if mask.ndim == 2:
                 mask = mask.unsqueeze(0)
-            # add a singleton heads dimension
+            # add a heads dimension if there isn't already one
             if mask.ndim == 3:
                 mask = mask.unsqueeze(1)
-            # pad to a multiple of 8
-            pad = 8 - mask.shape[-1] % 8
-            # the xformers docs says that it's allowed to have a mask of shape (1, Nq, Nk)
-            # but when using separated heads, the shape has to be (B, H, Nq, Nk)
-            # in flux, this matrix ends up being over 1GB
-            # here, we create a mask with the same batch/head size as the input mask (potentially singleton or full)
-            mask_out = torch.empty(
-                [mask.shape[0], mask.shape[1], q.shape[1], mask.shape[-1] + pad], dtype=q.dtype, device=q.device
-            )
 
-            mask_out[..., : mask.shape[-1]] = mask
+        out = _mps_sdpa_opt(q, k, v, attn_mask=mask)
-            # doesn't this remove the padding again??
+        out = out.transpose(1, 2).reshape(b, -1, heads * dim_head)
-            mask = mask_out[..., : mask.shape[-1]]
-            mask = mask.expand(b, heads, -1, -1)
-
-        out = memory_efficient_attention(q.to(v.dtype), k.to(v.dtype), v, attn_bias=mask, p=0.0)
-        out = out.reshape(b, -1, heads * dim_head)
         return out
 
 
 class FlashAttention3(AttentionCallable):
+    label = "FlashAttention3"
+
     def __call__(
         self,
         q: torch.Tensor,
         k: torch.Tensor,
         v: torch.Tensor,
         heads: int,
-        mask: torch.Tensor | None = None,
     ) -> torch.Tensor:
         if flash_attn_interface is None:
             raise RuntimeError("FlashAttention3 was selected but `FlashAttention3` is not installed.")
+        if q.device.type != "cuda":
+            raise RuntimeError("FlashAttention3 requires CUDA. Use PyTorch SDPA on CPU or MPS.")
 
         b, _, dim_head = q.shape
         dim_head //= heads
 
         q, k, v = (t.view(b, -1, heads, dim_head) for t in (q, k, v))
 
-        if mask is not None:
-            raise NotImplementedError("Mask is not supported for FlashAttention3")
-
         out = flash_attn_interface.flash_attn_func(q.to(v.dtype), k.to(v.dtype), v)
         out = out.reshape(b, -1, heads * dim_head)
         return out
 
 
+class FlashAttention4(AttentionCallable):
+    label = "FlashAttention4"
+
+    def __call__(
+        self,
+        q: torch.Tensor,
+        k: torch.Tensor,
+        v: torch.Tensor,
+        heads: int,
+    ) -> torch.Tensor:
+        if flash_attn_4_func is None:
+            raise RuntimeError("FlashAttention4 was selected but `flash-attn-4` is not installed.")
+        if q.device.type != "cuda":
+            raise RuntimeError("FlashAttention4 requires CUDA. Use PyTorch SDPA on CPU or MPS.")
+
+        b, _, dim_head = q.shape
+        dim_head //= heads
+
+        q, k, v = (t.view(b, -1, heads, dim_head) for t in (q, k, v))
+
+        out, _ = flash_attn_4_func(q.to(v.dtype), k.to(v.dtype), v)
+        out = out.reshape(b, -1, heads * dim_head)
+        return out
+
+
+# --- Automatic selection -----------------------------------------------------
+# AUTOMATIC inspects installed extras and the GPU arch and returns the fastest
+# usable callable for each path. The selection runs once per process (cached).
+# The unmasked and masked picks are independent: each calls its own helper and
+# may end up on different backends (e.g. FA3 unmasked + SDPA masked on H100).
+
+
+def _sdpa_can_use(backend: SDPBackend, *, with_mask: bool) -> bool:
+    """Ask torch whether *backend* can run with the given mask shape.
+    ``MATH`` is the universal SDPA fallback (pure PyTorch ops, no kernel
+    requirements) so it returns True everywhere, CPU included. The other
+    backends use ``torch.backends.cuda.can_use_*`` capability checks (no GPU
+    compute, no synchronization) and are False without CUDA. The probe shapes
+    are small but realistic enough to surface constraints (head dim, dtype)
+    that the per-backend rules care about.
+    """
+    if backend is SDPBackend.MATH:
+        return True
+    if not torch.cuda.is_available():
+        return False
+    q = torch.empty(1, 4, 128, 64, device="cuda", dtype=torch.bfloat16)
+    k = torch.empty(1, 4, 128, 64, device="cuda", dtype=torch.bfloat16)
+    v = torch.empty(1, 4, 128, 64, device="cuda", dtype=torch.bfloat16)
+    mask = torch.zeros(1, 4, 128, 128, device="cuda", dtype=torch.bfloat16) if with_mask else None
+    params = torch.backends.cuda.SDPAParams(q, k, v, mask, 0.0, False, False)
+    if backend is SDPBackend.CUDNN_ATTENTION:
+        return torch.backends.cuda.can_use_cudnn_attention(params, debug=False)
+    if backend is SDPBackend.FLASH_ATTENTION:
+        return torch.backends.cuda.can_use_flash_attention(params, debug=False)
+    if backend is SDPBackend.EFFICIENT_ATTENTION:
+        return torch.backends.cuda.can_use_efficient_attention(params, debug=False)
+    return False
+
+
+_SDPA_FULL_PRIORITY: tuple[SDPBackend, ...] = (
+    SDPBackend.CUDNN_ATTENTION,
+    SDPBackend.FLASH_ATTENTION,
+    SDPBackend.EFFICIENT_ATTENTION,
+    SDPBackend.MATH,
+)
+
+
+def _on_macos() -> bool:
+    """True on macOS, where torch's native SDPA materializes the score matrix and
+    AUTOMATIC routes to Apple's fused ``mps-sdpa`` kernel instead."""
+    return sys.platform == "darwin"
+
+
+def _mps_sdpa_available() -> bool:
+    """True when the ``mps-sdpa`` package is importable. It is a platform-marked
+    hard dependency on Apple Silicon, so this is always True there; it is only
+    False on non-Apple-Silicon macs (e.g. Intel/CPU), where AUTOMATIC falls back
+    to torch's SDPA (acceptable on CPU, which has no MPS memory wall)."""
+    return _mps_sdpa_opt is not None
+
+
+def _sdpa_full_priority() -> PytorchAttention:
+    """Hand SDPA the full backend priority order; let torch's dispatcher pick at call time.
+    ``sdpa_kernel(_SDPA_FULL_PRIORITY, set_priority=True)`` enables all four
+    backends and orders them; torch then walks the order at call time and picks
+    the first backend whose ``can_use_*`` check passes for the actual
+    shapes/dtype/mask. FLASH is rejected automatically when a mask is present;
+    CUDNN may be rejected under deterministic mode; MATH is the universal
+    fallback. Probing per-backend usability up front from generic probe shapes
+    cannot anticipate the variety of real call sites (e.g. broadcast key-only
+    masks, large head dim), so we defer the choice to the dispatcher.
+    """
+    return PytorchAttention(priority=list(_SDPA_FULL_PRIORITY))
+
+
+def _select_primary_attention() -> AttentionCallable:
+    """Pick the fastest unmasked attention based on installed extras and GPU arch.
+    Priority by arch:
+    - Hopper (sm_90, H100): FA3 > FA4 > SDPA.
+    - Datacenter Blackwell (sm_100, B200): FA4 > SDPA. FA4 is intentionally *not*
+      picked on consumer Blackwell (sm_120) -- known regressions in newer
+      FA4 betas; users who want it on sm_120 must opt in explicitly.
+    - macOS (Apple Silicon / MPS): Apple's fused MPSGraph kernel via ``mps-sdpa``
+      (a platform-marked hard dependency on Apple Silicon) -- it avoids the
+      full-score-matrix memory wall on long video sequences. On a non-Apple-Silicon
+      mac (Intel/CPU) it falls back to torch's SDPA.
+    - Everywhere else (Ada, Ampere, CPU): SDPA with the full backend priority
+      list -- torch's runtime dispatcher picks the best fit at call time.
+    """
+    if torch.cuda.is_available():
+        major, _ = torch.cuda.get_device_capability(0)
+        if major == 9:
+            if flash_attn_interface is not None:
+                return FlashAttention3()
+            if flash_attn_4_func is not None:
+                return FlashAttention4()
+        if major == 10 and flash_attn_4_func is not None:
+            return FlashAttention4()
+    if _on_macos():
+        return MPSSdpaAttention() if _mps_sdpa_available() else _sdpa_full_priority()
+    return _sdpa_full_priority()
+
+
+def _select_masked_attention() -> MaskedAttentionCallable:
+    """Pick a mask-aware attention. On macOS, Apple's fused MPSGraph kernel via
+    ``mps-sdpa`` (a hard dependency on Apple Silicon, else torch's SDPA on
+    Intel/CPU macs); else SDPA with the full priority list (the dispatcher
+    rejects FLASH automatically when a mask is present and walks past it --
+    torch SDPA handles the additive mask directly)."""
+    if _on_macos():
+        return MPSSdpaAttention() if _mps_sdpa_available() else _sdpa_full_priority()
+    return _sdpa_full_priority()
+
+
+@functools.cache
+def automatic_attention() -> AttentionCallable:
+    """Cached AUTOMATIC pick for the unmasked path.
+    Cached so every ``AttentionOps`` in the process shares one instance."""
+    return _select_primary_attention()
+
+
+@functools.cache
+def automatic_masked_attention() -> MaskedAttentionCallable:
+    """Cached AUTOMATIC pick for the masked path. See :func:`automatic_attention`."""
+    return _select_masked_attention()
+
+
+def attention_label(fn: AttentionCallable | MaskedAttentionCallable) -> str:
+    """Best-effort human-readable backend name.
+    Built-in callables expose ``.label`` (encoding the SDPA priority list for the
+    Pytorch backends); fall back to the class name for custom or wrapped callables
+    (e.g. the multi-GPU All2All wrappers) that don't define one."""
+    return getattr(fn, "label", type(fn).__name__)
+
+
+def _resolve_sdpa_variant(backend: SDPBackend, name: str, *, with_mask: bool) -> PytorchAttention:
+    """Build a single-backend ``PytorchAttention`` pin, raising if the backend
+    can't actually serve the call on this machine. Used by both
+    :meth:`AttentionFunction.to_callable` and :meth:`MaskedAttentionFunction.to_callable`;
+    ``with_mask`` differs between the two so the capability check considers
+    the protocol the caller intends to use. Not used for ``MATH`` -- MATH is
+    the universal fallback and would falsely fail the CUDA-only probe on CPU.
+    """
+    if not _sdpa_can_use(backend, with_mask=with_mask):
+        raise RuntimeError(
+            f"{name} selected but the SDPA {backend.name} backend is not usable on this machine "
+            "(either no CUDA, the backend rejected the probe shapes, or "
+            "torch.use_deterministic_algorithms(True) excluded it)."
+        )
+    return PytorchAttention(priority=[backend])
+
+
 class AttentionFunction(Enum):
     PYTORCH = "pytorch"
-    XFORMERS = "xformers"
     FLASH_ATTENTION_3 = "flash_attention_3"
-    DEFAULT = "default"
+    FLASH_ATTENTION_4 = "flash_attention_4"
+    SDPA_CUDNN = "sdpa_cudnn"
+    SDPA_FLASH = "sdpa_flash"
+    SDPA_EFFICIENT = "sdpa_efficient"
+    SDPA_MATH = "sdpa_math"
+    # Apple's fused MPSGraph SDPA via the `mps-sdpa` package (macOS/MPS only, a
+    # platform-marked hard dependency on Apple Silicon). The AUTOMATIC default on
+    # MPS; never materializes the score matrix.
+    MPS_SDPA = "mps_sdpa"
+    # Pick the fastest unmasked backend for the current GPU/extras combo; see
+    # :func:`automatic_attention`. Default for :class:`AttentionOps`.
+    AUTOMATIC = "automatic"
 
-    def to_callable(self) -> AttentionCallable:
+    def to_callable(self) -> AttentionCallable:  # noqa: PLR0911, PLR0912
         """Resolve to a concrete callable. Use this at module init time so that
-        torch.compile can trace through the attention call without graph breaks."""
+        torch.compile can trace through the attention call without graph breaks.
-        if self is AttentionFunction.PYTORCH:
+        Every non-AUTOMATIC variant raises :class:`RuntimeError` when the backend
+        isn't usable on this machine -- missing package or SDPA backend rejected
+        on this hardware (e.g. cuDNN under ``torch.use_deterministic_algorithms``).
+        Opting in means "this kernel or fail loudly". ``AUTOMATIC`` returns the
+        cached :func:`automatic_attention` instance so every build shares one callable.
+        """
+        match self:
+            case AttentionFunction.AUTOMATIC:
+                return automatic_attention()
+            case AttentionFunction.PYTORCH:
                 return PytorchAttention()
-        elif self is AttentionFunction.XFORMERS:
+            case AttentionFunction.FLASH_ATTENTION_3:
-            return XFormersAttention()
+                if flash_attn_interface is None:
-        elif self is AttentionFunction.FLASH_ATTENTION_3:
+                    raise RuntimeError(
+                        "AttentionFunction.FLASH_ATTENTION_3 selected but `flash-attn-3` is not installed."
+                    )
+                if not torch.cuda.is_available():
+                    raise RuntimeError(
+                        "AttentionFunction.FLASH_ATTENTION_3 requires CUDA. Use PyTorch SDPA on CPU or MPS."
+                    )
                 return FlashAttention3()
-        else:
+            case AttentionFunction.FLASH_ATTENTION_4:
-            # Default behavior: XFormers if installed else - PyTorch
+                if flash_attn_4_func is None:
-            return XFormersAttention() if memory_efficient_attention is not None else PytorchAttention()
+                    raise RuntimeError(
+                        "AttentionFunction.FLASH_ATTENTION_4 selected but `flash-attn-4` is not installed."
+                    )
+                if not torch.cuda.is_available():
+                    raise RuntimeError(
+                        "AttentionFunction.FLASH_ATTENTION_4 requires CUDA. Use PyTorch SDPA on CPU or MPS."
+                    )
+                return FlashAttention4()
+            case AttentionFunction.SDPA_MATH:
+                return PytorchAttention(priority=[SDPBackend.MATH])
+            case AttentionFunction.MPS_SDPA:
+                if _mps_sdpa_opt is None:
+                    raise RuntimeError("AttentionFunction.MPS_SDPA selected but `mps-sdpa` is not installed.")
+                return MPSSdpaAttention()
+            case AttentionFunction.SDPA_CUDNN:
+                return _resolve_sdpa_variant(
+                    SDPBackend.CUDNN_ATTENTION, "AttentionFunction.SDPA_CUDNN", with_mask=False
+                )
+            case AttentionFunction.SDPA_FLASH:
+                return _resolve_sdpa_variant(
+                    SDPBackend.FLASH_ATTENTION, "AttentionFunction.SDPA_FLASH", with_mask=False
+                )
+            case AttentionFunction.SDPA_EFFICIENT:
+                return _resolve_sdpa_variant(
+                    SDPBackend.EFFICIENT_ATTENTION, "AttentionFunction.SDPA_EFFICIENT", with_mask=False
+                )
+
+
+class MaskedAttentionFunction(Enum):
+    """Backends usable on the masked path. Mirrors :class:`AttentionFunction` minus
+    the variants the torch SDPA dispatcher (or the wrapped kernel) rejects with a
+    mask: ``SDPA_FLASH`` -- FLASH kernel cannot serve an additive ``attn_mask``;
+    ``FLASH_ATTENTION_3``/``FLASH_ATTENTION_4`` -- neither has a mask kernel at all.
+    Keeping them out makes "this backend cannot mask" a type error, not a runtime one."""
+
+    PYTORCH = "pytorch"
+    SDPA_CUDNN = "sdpa_cudnn"
+    SDPA_EFFICIENT = "sdpa_efficient"
+    SDPA_MATH = "sdpa_math"
+    # Apple's fused MPSGraph SDPA via the `mps-sdpa` package (macOS/MPS only, a
+    # platform-marked hard dependency on Apple Silicon); the AUTOMATIC default on
+    # MPS. Mask-aware.
+    MPS_SDPA = "mps_sdpa"
+    # Pick the fastest mask-capable backend for the current extras combo; see
+    # :func:`automatic_masked_attention`. Default for the masked slot of
+    # :class:`AttentionOps`.
+    AUTOMATIC = "automatic"
+
+    def to_callable(self) -> MaskedAttentionCallable:
+        """Resolve to a concrete masked callable. Same backend classes as
+        :meth:`AttentionFunction.to_callable`; the protocol returned just exposes
+        the masked call signature.
+        Non-AUTOMATIC variants raise :class:`RuntimeError` when the backend isn't
+        usable for the masked path on this machine. SDPA probes run with
+        ``with_mask=True`` so the capability check considers the protocol the
+        caller will actually use."""
+        match self:
+            case MaskedAttentionFunction.AUTOMATIC:
+                return automatic_masked_attention()
+            case MaskedAttentionFunction.PYTORCH:
+                return PytorchAttention()
+            case MaskedAttentionFunction.SDPA_MATH:
+                return PytorchAttention(priority=[SDPBackend.MATH])
+            case MaskedAttentionFunction.MPS_SDPA:
+                if _mps_sdpa_opt is None:
+                    raise RuntimeError("MaskedAttentionFunction.MPS_SDPA selected but `mps-sdpa` is not installed.")
+                return MPSSdpaAttention()
+            case MaskedAttentionFunction.SDPA_CUDNN:
+                return _resolve_sdpa_variant(
+                    SDPBackend.CUDNN_ATTENTION, "MaskedAttentionFunction.SDPA_CUDNN", with_mask=True
+                )
+            case MaskedAttentionFunction.SDPA_EFFICIENT:
+                return _resolve_sdpa_variant(
+                    SDPBackend.EFFICIENT_ATTENTION, "MaskedAttentionFunction.SDPA_EFFICIENT", with_mask=True
+                )
+
+
+@dataclass(frozen=True)
+class AttentionOps:
+    """Pluggable callables consumed by :class:`Attention`."""
+
+    attention_function: AttentionCallable = field(default_factory=lambda: AttentionFunction.AUTOMATIC.to_callable())
+    masked_attention_function: MaskedAttentionCallable = field(
+        default_factory=lambda: MaskedAttentionFunction.AUTOMATIC.to_callable()
+    )
+    preattention_function: PreAttentionCallable = field(default_factory=PytorchPreAttention)
+    gated_attention_function: GatedAttentionCallable = field(default_factory=PytorchGatedAttention)
 
 
 class Attention(torch.nn.Module):
@@ -144,17 +480,18 @@ class Attention(torch.nn.Module):
         heads: int = 8,
         dim_head: int = 64,
         norm_eps: float = 1e-6,
-        rope_type: LTXRopeType = LTXRopeType.INTERLEAVED,
+        rope_type: LTXRopeType = LTXRopeType.SPLIT,
-        attention_function: AttentionCallable | AttentionFunction = AttentionFunction.DEFAULT,
+        ops: AttentionOps | None = None,
         apply_gated_attention: bool = False,
     ) -> None:
         super().__init__()
+        if ops is None:
+            ops = AttentionOps()
         self.rope_type = rope_type
-        self.attention_function = (
+        self.attention_function = ops.attention_function
-            attention_function.to_callable()
+        self.masked_attention_function = ops.masked_attention_function
-            if isinstance(attention_function, AttentionFunction)
+        self.preattention_function = ops.preattention_function
-            else attention_function
+        self.gated_attention_function = ops.gated_attention_function
-        )
 
         inner_dim = dim_head * heads
         context_dim = query_dim if context_dim is None else context_dim
@@ -196,7 +533,9 @@ class Attention(torch.nn.Module):
             context: Key/value context tensor of shape ``(B, S, context_dim)``.
                 Falls back to ``x`` (self-attention) when *None*.
             mask: Optional attention mask. Interpretation depends on the attention
-                backend (additive bias for xformers/PyTorch SDPA).
+                backend (additive bias for PyTorch SDPA). A non-None
+                ``mask`` routes to ``masked_attention_function``; ``None`` keeps
+                the unmasked path.
             pe: Rotary positional embeddings applied to both ``q`` and ``k``.
             k_pe: Separate rotary positional embeddings for ``k`` only. When
                 *None*, ``pe`` is reused for keys.
@@ -221,29 +560,17 @@ class Attention(torch.nn.Module):
         else:
             q = self.to_q(x)
             k = self.to_k(context)
-
+            q, k = self.preattention_function(q, k, self, mask, pe, k_pe)
-            q = self.q_norm(q)
+            if mask is None:
-            k = self.k_norm(k)
+                out = self.attention_function(q, k, v, self.heads)  # (B, T, H*D)
-
+            else:
-            if pe is not None:
+                out = self.masked_attention_function(q, k, v, self.heads, mask)
-                q = apply_rotary_emb(q, pe, self.rope_type)
-                k = apply_rotary_emb(k, pe if k_pe is None else k_pe, self.rope_type)
-
-            out = self.attention_function(q, k, v, self.heads, mask)  # (B, T, H*D)
 
             if perturbation_mask is not None:
                 out = out * perturbation_mask + v * (1 - perturbation_mask)
 
         # Apply per-head gating if enabled
         if self.to_gate_logits is not None:
-            gate_logits = self.to_gate_logits(x)  # (B, T, H)
+            out = self.gated_attention_function(x, out, self)
-            b, t, _ = out.shape
-            # Reshape to (B, T, H, D) for per-head gating
-            out = out.view(b, t, self.heads, self.dim_head)
-            # Apply gating: 2 * sigmoid(x) so that zero-init gives identity (2 * 0.5 = 1.0)
-            gates = 2.0 * torch.sigmoid(gate_logits)  # (B, T, H)
-            out = out * gates.unsqueeze(-1)  # (B, T, H, D) * (B, T, H, 1)
-            # Reshape back to (B, T, H*D)
-            out = out.view(b, t, self.heads * self.dim_head)
 
         return self.to_out(out)

packages/ltx-core/src/ltx_core/model/transformer/compiling.py

@@ -1,19 +1,119 @@
+from dataclasses import dataclass, field, replace
+from typing import Any
+
 import torch
 
+from ltx_core.guidance.perturbations import BatchedPerturbationConfig, PerturbationType
 from ltx_core.loader.module_ops import ModuleOps
 from ltx_core.loader.sd_ops import SDOps
 from ltx_core.model.transformer.model import LTXModel
+from ltx_core.model.transformer.transformer_args import BlockPerturbationsProcessor, TransformerArgs
+
+# Defaults applied inside the patched forward. Overriding via CompilationConfig
+# replaces these wholesale; it does not merge.
+_DEFAULT_INDUCTOR_CONFIG: dict[str, Any] = {}
+_DEFAULT_DYNAMO_CONFIG: dict[str, Any] = {"inline_inbuilt_nn_modules": True, "cache_size_limit": 256}
 
 
-def compile_transformer(model: LTXModel) -> LTXModel:
+@dataclass(frozen=True)
-    model.transformer_blocks = torch.nn.ModuleList(torch.compile(m) for m in model.transformer_blocks)
+class CompilationConfig:
+    """``torch.compile`` configuration for transformer blocks. ``None`` keeps eager."""
+
+    mode: str | None = None
+    backend: str = "inductor"
+    fullgraph: bool = False
+    dynamic: bool | None = None
+    inductor_config: dict[str, Any] = field(default_factory=lambda: dict(_DEFAULT_INDUCTOR_CONFIG))
+    dynamo_config: dict[str, Any] = field(default_factory=lambda: dict(_DEFAULT_DYNAMO_CONFIG))
+
+
+class CompiledBlockPerturbationsProcessor(BlockPerturbationsProcessor):
+    """Per-block input prep for compiled blocks: mark the seq dim dynamic, then attach perturbation
+    as config-independent runtime masks so the block traces ONCE.
+    The ``mark_dynamic`` calls keep the per-block compile artifact shape-polymorphic; they run in
+    eager mode (this processor lives outside the compiled region) on the tensors about to cross into
+    the trace. Both keep-masks are then attached UNCONDITIONALLY and the skip flags pinned False, so
+    the trace is identical for every pass (cond / uncond / STG): the block never sees a flipped
+    Python bool (``self_attn_all_perturbed``) or a None-vs-tensor mask that Dynamo would specialise
+    on, so the STG pass no longer triggers a recompile. An all-keep mask blends to a no-op
+    (``out*1 + v*0``); an all-zero mask reproduces the skip. Reads only ``mask`` (runtime tensor
+    indexing), never the host-side ``all_in_batch`` / ``any_in_batch``.
+    """
+
+    def __call__(
+        self,
+        args: TransformerArgs,
+        perturbations: BatchedPerturbationConfig,
+        block_idx: int,
+        self_attn_type: PerturbationType,
+        cross_attn_type: PerturbationType,
+    ) -> TransformerArgs:
+        # Positional embeddings are second-from-last regardless of rope type:
+        # split rope is (B, H, T, D//2) -- dim -2 == 2; interleaved rope is (B, T, D)
+        # -- dim -2 == 1. Both work via the negative index.
+        torch._dynamo.mark_dynamic(args.x, 1)
+        cos, sin = args.positional_embeddings
+        torch._dynamo.mark_dynamic(cos, cos.ndim - 2)
+        torch._dynamo.mark_dynamic(sin, sin.ndim - 2)
+        if args.cross_positional_embeddings is not None:
+            cross_cos, cross_sin = args.cross_positional_embeddings
+            torch._dynamo.mark_dynamic(cross_cos, cross_cos.ndim - 2)
+            torch._dynamo.mark_dynamic(cross_sin, cross_sin.ndim - 2)
+        if args.self_attention_mask is not None:
+            # Dense form is (B, 1, T, T); key-padding form (from the SP wrapper)
+            # is (B, 1, 1, T) -- leave the size-1 query dim static so Dynamo
+            # keeps the broadcast.
+            if args.self_attention_mask.shape[2] > 1:
+                torch._dynamo.mark_dynamic(args.self_attention_mask, 2)
+            torch._dynamo.mark_dynamic(args.self_attention_mask, 3)
+        if args.context_mask is not None:
+            torch._dynamo.mark_dynamic(args.context_mask, 2)
+        # `timesteps` / `embedded_timestep` are per-token when conditioning sets a
+        # per-position denoise mask, in which case their dim 1 equals the seq length
+        # and must vary with it. When they're a single timestep broadcast across the
+        # sequence (dim 1 == 1), leaving them static lets Dynamo keep the size-1
+        # broadcast.
+        if args.timesteps.shape[1] > 1:
+            torch._dynamo.mark_dynamic(args.timesteps, 1)
+        if args.embedded_timestep.shape[1] > 1:
+            torch._dynamo.mark_dynamic(args.embedded_timestep, 1)
+        # `cross_scale_shift_timestep` is the cross-attn AdaLN scale/shift input
+        # derived from the own-modality per-token timesteps (denoise_mask * sigma),
+        # so its dim 1 equals the seq length when conditioning is per-token.
+        # `cross_gate_timestep` is the cross-modality sigma scalar -- dim 1 is 1
+        # and broadcasts, leave it static. Same guard pattern as `timesteps`.
+        if args.cross_scale_shift_timestep is not None and args.cross_scale_shift_timestep.shape[1] > 1:
+            torch._dynamo.mark_dynamic(args.cross_scale_shift_timestep, 1)
+        # Perturbation as config-independent runtime masks (skip flags pinned False -> no recompile).
+        return replace(
+            args,
+            self_attn_perturbation_mask=perturbations.mask(self_attn_type, block_idx),
+            self_attn_all_perturbed=False,
+            cross_attn_perturbation_mask=perturbations.mask(cross_attn_type, block_idx),
+            cross_attn_skip_all=False,
+        )
+
+
+def compile_transformer(model: LTXModel, config: CompilationConfig) -> LTXModel:
+    """Compile each transformer block via ``torch.compile`` with the given settings.
+    The patched forward emits ``torch.compiler.cudagraph_mark_step_begin()`` once
+    per step. Under CUDA-graph-enabling modes (``"reduce-overhead"`` /
+    ``"max-autotune"``) this overrides Dynamo's per-invocation auto-mark
+    heuristic, which would otherwise fire once per compiled block call (48 per
+    forward) and treat each block call as a fresh iteration. Under other modes
+    the mark is a no-op (decrements an unread counter).
+    """
+    model.transformer_blocks = torch.nn.ModuleList(
+        torch.compile(m, mode=config.mode, backend=config.backend, fullgraph=config.fullgraph, dynamic=config.dynamic)
+        for m in model.transformer_blocks
+    )
+    model.block_input_processor = CompiledBlockPerturbationsProcessor()
 
     def patched_dynamo_forward(*args, **kwargs) -> tuple[torch.Tensor, torch.Tensor]:
+        torch.compiler.cudagraph_mark_step_begin()
         with (
-            torch._inductor.config.patch(unsafe_skip_cache_dynamic_shape_guards=True),
+            torch._inductor.config.patch(**config.inductor_config),
-            torch._dynamo.config.patch(  # type: ignore[attr-defined]
+            torch._dynamo.config.patch(**config.dynamo_config),  # type: ignore[attr-defined]
-                inline_inbuilt_nn_modules=True, cache_size_limit=256, allow_unspec_int_on_nn_module=True
-            ),
         ):
             return model.forward_without_compilation(*args, **kwargs)
 
@@ -22,11 +122,13 @@ def compile_transformer(model: LTXModel) -> LTXModel:
     return model
 
 
-COMPILE_TRANSFORMER = ModuleOps(
+def build_compile_transformer_op(config: CompilationConfig) -> ModuleOps:
+    """Build a ``ModuleOps`` that compiles transformer blocks with the given settings."""
+    return ModuleOps(
         name="compile_transformer",
         matcher=lambda model: isinstance(model, LTXModel),
-    mutator=lambda model: compile_transformer(model),
+        mutator=lambda model: compile_transformer(model, config),
-)
+    )
 
 
 def modify_sd_ops_for_compilation(original_sd_ops: SDOps, number_of_blocks: int = 48) -> SDOps:

packages/ltx-core/src/ltx_core/model/transformer/mask_channels_checkpoint.py

@@ -0,0 +1,92 @@
+"""Checkpoint surgery to enable SCAIL-2 in-context mask channels on a trained LTX-2 model.
+
+Widening ``LTXModel.mask_conditioning_channels`` from 0 to ``C`` grows the video
+``patchify_proj`` weight from ``[inner, in]`` to ``[inner, in + C]``. This helper
+appends ``C`` **zero** input columns to that weight so a converted checkpoint is
+numerically identical to the original until the new columns are finetuned -- the
+extra channels contribute nothing at load time.
+
+Usage::
+
+    sd = load_state_dict(path)
+    widen_patchify_proj_for_mask_channels(sd, mask_channels=56)
+    model = LTXModel(..., mask_conditioning_channels=56)
+    model.load_state_dict(sd)
+"""
+
+from __future__ import annotations
+
+import torch
+
+_VIDEO_PROJ_SUFFIX = "patchify_proj.weight"
+_AUDIO_PROJ_SUFFIX = "audio_patchify_proj.weight"
+
+
+def widen_module_patchify_proj_for_mask_channels(model: torch.nn.Module, mask_channels: int) -> torch.nn.Module:
+    """In place, widen a live ``LTXModel``'s video ``patchify_proj`` to accept ``mask_channels`` extra inputs.
+
+    Replaces ``model.patchify_proj`` with a wider ``Linear`` whose original input columns are copied and
+    whose ``mask_channels`` new columns are zero (so the model reproduces its pre-widening output until
+    those columns are trained). Also sets ``model.mask_conditioning_channels`` for metadata/consistency.
+    Idempotent guard: raises if the model is already widened to a different value.
+    """
+    if mask_channels < 0:
+        raise ValueError(f"mask_channels must be non-negative, got {mask_channels}")
+
+    proj = getattr(model, "patchify_proj", None)
+    if not isinstance(proj, torch.nn.Linear):
+        raise AttributeError("model has no linear 'patchify_proj' to widen")
+
+    already = int(getattr(model, "mask_conditioning_channels", 0))
+    if mask_channels in (0, already):
+        model.mask_conditioning_channels = mask_channels
+        return model
+    if already != 0:
+        raise ValueError(f"patchify_proj already widened for {already} mask channels, refusing to re-widen")
+
+    new_in = proj.in_features + mask_channels
+    wider = torch.nn.Linear(new_in, proj.out_features, bias=proj.bias is not None)
+    wider = wider.to(device=proj.weight.device, dtype=proj.weight.dtype)
+    with torch.no_grad():
+        wider.weight.zero_()
+        wider.weight[:, : proj.in_features].copy_(proj.weight)
+        if proj.bias is not None:
+            wider.bias.copy_(proj.bias)
+    model.patchify_proj = wider
+    model.mask_conditioning_channels = mask_channels
+    return model
+
+
+def widen_patchify_proj_for_mask_channels(
+    state_dict: dict[str, torch.Tensor],
+    mask_channels: int,
+) -> dict[str, torch.Tensor]:
+    """In place, append ``mask_channels`` zero input columns to the video ``patchify_proj`` weight.
+
+    Only the video projection is touched (audio ``audio_patchify_proj`` is left alone). The bias,
+    if present, is unchanged. Returns the same dict for convenience. Idempotency is the caller's
+    responsibility -- calling twice widens twice.
+    """
+    if mask_channels < 0:
+        raise ValueError(f"mask_channels must be non-negative, got {mask_channels}")
+    if mask_channels == 0:
+        return state_dict
+
+    target_keys = [
+        key
+        for key in state_dict
+        if key.endswith(_VIDEO_PROJ_SUFFIX) and not key.endswith(_AUDIO_PROJ_SUFFIX)
+    ]
+    if not target_keys:
+        raise KeyError(
+            f"No '{_VIDEO_PROJ_SUFFIX}' weight found in the state dict; cannot widen for mask channels."
+        )
+
+    for key in target_keys:
+        weight = state_dict[key]  # [inner_dim, in_features]
+        if weight.ndim != 2:
+            raise ValueError(f"Expected 2-D weight for '{key}', got shape {tuple(weight.shape)}")
+        pad = weight.new_zeros(weight.shape[0], mask_channels)
+        state_dict[key] = torch.cat([weight, pad], dim=1)
+
+    return state_dict

packages/ltx-core/src/ltx_core/model/transformer/modality.py

@@ -17,8 +17,19 @@ class Modality:
             the batch size, *T* is the total number of tokens (noisy +
             conditioning), and *D* is the input dimension.
         timesteps: Per-token timestep embeddings, shape ``(B, T)``.
-        positions: Positional coordinates, shape ``(B, 3, T)`` for video
+        positions: Per-token patch coordinates used to build the RoPE
-            (time, height, width) or ``(B, 1, T)`` for audio.
+            frequencies. With the default ``use_middle_indices_grid=True``,
+            shape is ``(B, n_pos_dims, T, 2)`` where ``n_pos_dims=3`` for
+            video (time, height, width) and ``n_pos_dims=1`` for audio
+            (time); the last dim of size 2 holds the ``[start, end)``
+            index bounds of each patch, and RoPE is evaluated at the
+            *middle* of that range -- hence the flag name. Taking the
+            patch midpoint produces a smoother and more accurate
+            positional signal than indexing by the patch's start when
+            patches span more than one spatial / temporal unit.
+            When ``use_middle_indices_grid=False``, the legacy 3-D form
+            ``(B, n_pos_dims, T)`` of integer positional indices is
+            accepted instead and used as-is (no midpoint derivation).
         context: Text conditioning embeddings from the prompt encoder.
         enabled: Whether this modality is active in the current forward pass.
         context_mask: Optional mask for the text context tokens.
@@ -27,6 +38,9 @@ class Modality:
             attention. ``None`` means unrestricted (full) attention between
             all tokens. Built incrementally by conditioning items; see
             :class:`~ltx_core.conditioning.types.attention_strength_wrapper.ConditioningItemAttentionStrengthWrapper`.
+        cond_channels: Optional in-context conditioning channels, shape ``(B, T, C_cond)``, aligned token-for-token
+            with ``latent``. Concatenated onto ``latent`` before the first projection (see
+            :meth:`TransformerArgsPreprocessor.prepare`); never noised. ``None`` for standard models.
     """
 
     latent: (
@@ -34,13 +48,15 @@ class Modality:
     )  # Shape: (B, T, D) where B is the batch size, T is the number of tokens, and D is input dimension
     sigma: torch.Tensor  # Shape: (B,). Current sigma value, used for cross-attention timestep calculation.
     timesteps: torch.Tensor  # Shape: (B, T) where T is the number of timesteps
-    positions: (
+    # Shape: (B, n_pos_dims, T, 2) by default (use_middle_indices_grid=True);
-        torch.Tensor
+    # n_pos_dims=3 for video, 1 for audio; last dim holds [start, end) patch bounds.
-    )  # Shape: (B, 3, T) for video, where 3 is the number of dimensions and T is the number of tokens
+    # Legacy form (B, n_pos_dims, T) when use_middle_indices_grid=False.
+    positions: torch.Tensor
     context: torch.Tensor
     enabled: bool = True
     context_mask: torch.Tensor | None = None
     attention_mask: torch.Tensor | None = None
+    cond_channels: torch.Tensor | None = None
 
     def split(self, sizes: list[int]) -> list[Modality]:
         """Split along the batch dimension into chunks of the given sizes."""

packages/ltx-core/src/ltx_core/model/transformer/model.py

@@ -1,20 +1,30 @@
+import logging
 from enum import Enum
 
 import torch
 
-from ltx_core.guidance.perturbations import BatchedPerturbationConfig
+from ltx_core.guidance.perturbations import BatchedPerturbationConfig, PerturbationType
+from ltx_core.model.model_protocol import LTXModelProtocol
 from ltx_core.model.transformer.adaln import AdaLayerNormSingle, adaln_embedding_coefficient
-from ltx_core.model.transformer.attention import AttentionCallable, AttentionFunction
+from ltx_core.model.transformer.attention import attention_label
 from ltx_core.model.transformer.modality import Modality
 from ltx_core.model.transformer.rope import LTXRopeType
-from ltx_core.model.transformer.transformer import BasicAVTransformerBlock, TransformerConfig
+from ltx_core.model.transformer.transformer import (
+    DEFAULT_TRANSFORMER_OPS,
+    BasicAVTransformerBlock,
+    TransformerConfig,
+    TransformerOpsConfig,
+)
 from ltx_core.model.transformer.transformer_args import (
+    BlockPerturbationsProcessor,
     MultiModalTransformerArgsPreprocessor,
     TransformerArgs,
     TransformerArgsPreprocessor,
 )
 from ltx_core.utils import to_denoised
 
+logger = logging.getLogger(__name__)
+
 
 class LTXModelType(Enum):
     AudioVideo = "ltx av model"
@@ -45,7 +55,7 @@ class LTXModel(torch.nn.Module):
         num_layers: int = 48,
         cross_attention_dim: int = 4096,
         norm_eps: float = 1e-06,
-        attention_type: AttentionFunction | AttentionCallable = AttentionFunction.DEFAULT,
+        ops: TransformerOpsConfig = DEFAULT_TRANSFORMER_OPS,
         positional_embedding_theta: float = 10000.0,
         positional_embedding_max_pos: list[int] | None = None,
         timestep_scale_multiplier: int = 1000,
@@ -57,16 +67,30 @@ class LTXModel(torch.nn.Module):
         audio_cross_attention_dim: int = 2048,
         audio_positional_embedding_max_pos: list[int] | None = None,
         av_ca_timestep_scale_multiplier: int = 1,
-        rope_type: LTXRopeType = LTXRopeType.INTERLEAVED,
+        rope_type: LTXRopeType = LTXRopeType.SPLIT,
         double_precision_rope: bool = False,
         apply_gated_attention: bool = False,
         caption_projection: torch.nn.Module | None = None,
         audio_caption_projection: torch.nn.Module | None = None,
         cross_attention_adaln: bool = False,
+        mask_conditioning_channels: int = 0,
     ):
         super().__init__()
+        # Log the attention backends this transformer is built with. Reading the resolved
+        # ``label`` off the ops reports whatever was selected -- AUTOMATIC, an explicit pin
+        # (PYTORCH/FA3/FA4/SDPA_*), or a directly supplied callable -- so this is the
+        # single source of truth for which kernel a build uses. Fires once per build.
+        logger.info(
+            "Building transformer with attention backends -- self: %s, masked: %s",
+            attention_label(ops.attention_ops.attention_function),
+            attention_label(ops.attention_ops.masked_attention_function),
+        )
         self._enable_gradient_checkpointing = False
         self.cross_attention_adaln = cross_attention_adaln
+        # Extra per-token input channels (e.g. SCAIL-2 in-context mask channels) concatenated onto the
+        # video latent before ``patchify_proj``. 0 keeps the standard input width; >0 widens the first
+        # projection. Fed via ``Modality.cond_channels`` (see TransformerArgsPreprocessor.prepare).
+        self.mask_conditioning_channels = mask_conditioning_channels
         self.use_middle_indices_grid = use_middle_indices_grid
         self.rope_type = rope_type
         self.double_precision_rope = double_precision_rope
@@ -115,9 +139,13 @@ class LTXModel(torch.nn.Module):
             audio_attention_head_dim=audio_attention_head_dim if model_type.is_audio_enabled() else 0,
             audio_cross_attention_dim=audio_cross_attention_dim,
             norm_eps=norm_eps,
-            attention_type=attention_type,
+            ops=ops,
             apply_gated_attention=apply_gated_attention,
         )
+        # Hook for per-block input prep. Compile transforms in `compiling.py`
+        # wrap (not replace) this with a processor that also marks the seq dim
+        # dynamic, so any caller customisation here is preserved as the inner.
+        self.block_input_processor = BlockPerturbationsProcessor()
 
     @property
     def _adaln_embedding_coefficient(self) -> int:
@@ -131,8 +159,11 @@ class LTXModel(torch.nn.Module):
         caption_projection: torch.nn.Module | None = None,
     ) -> None:
         """Initialize video-specific components."""
-        # Video input components
+        # Video input components. When ``mask_conditioning_channels > 0`` the first projection is
-        self.patchify_proj = torch.nn.Linear(in_channels, self.inner_dim, bias=True)
+        # widened to also accept the in-context conditioning channels concatenated onto the latent.
+        self.patchify_proj = torch.nn.Linear(
+            in_channels + self.mask_conditioning_channels, self.inner_dim, bias=True
+        )
         if caption_projection is not None:
             self.caption_projection = caption_projection
 
@@ -284,7 +315,7 @@ class LTXModel(torch.nn.Module):
         audio_attention_head_dim: int,
         audio_cross_attention_dim: int,
         norm_eps: float,
-        attention_type: AttentionFunction | AttentionCallable,
+        ops: TransformerOpsConfig,
         apply_gated_attention: bool,
     ) -> None:
         """Initialize transformer blocks for LTX."""
@@ -315,14 +346,13 @@ class LTXModel(torch.nn.Module):
         self.transformer_blocks = torch.nn.ModuleList(
             [
                 BasicAVTransformerBlock(
-                    idx=idx,
                     video=video_config,
                     audio=audio_config,
                     rope_type=self.rope_type,
                     norm_eps=norm_eps,
-                    attention_function=attention_type,
+                    ops=ops,
                 )
-                for idx in range(num_layers)
+                for _ in range(num_layers)
             ]
         )
 
@@ -336,33 +366,45 @@ class LTXModel(torch.nn.Module):
         """
         self._enable_gradient_checkpointing = enable
 
+    @property
+    def num_blocks(self) -> int:
+        """Number of transformer blocks."""
+        return len(self.transformer_blocks)
+
     def _process_transformer_blocks(
         self,
         video: TransformerArgs | None,
         audio: TransformerArgs | None,
         perturbations: BatchedPerturbationConfig,
-    ) -> tuple[TransformerArgs, TransformerArgs]:
+    ) -> tuple[TransformerArgs | None, TransformerArgs | None]:
-        """Process transformer blocks for LTXAV."""
+        """Process transformer blocks for LTX."""
+        for block_idx, block in enumerate(self.transformer_blocks):
+            if video is not None:
+                video = self.block_input_processor(
+                    video,
+                    perturbations,
+                    block_idx,
+                    self_attn_type=PerturbationType.SKIP_VIDEO_SELF_ATTN,
+                    cross_attn_type=PerturbationType.SKIP_A2V_CROSS_ATTN,
+                )
+            if audio is not None:
+                audio = self.block_input_processor(
+                    audio,
+                    perturbations,
+                    block_idx,
+                    self_attn_type=PerturbationType.SKIP_AUDIO_SELF_ATTN,
+                    cross_attn_type=PerturbationType.SKIP_V2A_CROSS_ATTN,
+                )
 
-        # Process transformer blocks
-        for block in self.transformer_blocks:
             if self._enable_gradient_checkpointing and self.training:
-                # Use gradient checkpointing to save memory during training.
-                # With use_reentrant=False, we can pass dataclasses directly -
-                # PyTorch will track all tensor leaves in the computation graph.
                 video, audio = torch.utils.checkpoint.checkpoint(
                     block,
                     video,
                     audio,
-                    perturbations,
                     use_reentrant=False,
                 )
             else:
-                video, audio = block(
+                video, audio = block(video=video, audio=audio)
-                    video=video,
-                    audio=audio,
-                    perturbations=perturbations,
-                )
 
         return video, audio
 
@@ -387,8 +429,8 @@ class LTXModel(torch.nn.Module):
         return x
 
     def forward(
-        self, video: Modality | None, audio: Modality | None, perturbations: BatchedPerturbationConfig
+        self, video: Modality | None, audio: Modality | None, perturbations: BatchedPerturbationConfig | None
-    ) -> tuple[torch.Tensor, torch.Tensor]:
+    ) -> tuple[torch.Tensor | None, torch.Tensor | None]:
         """
         Forward pass for LTX models.
         Returns:
@@ -401,6 +443,11 @@ class LTXModel(torch.nn.Module):
 
         video_args = self.video_args_preprocessor.prepare(video, audio) if video is not None else None
         audio_args = self.audio_args_preprocessor.prepare(audio, video) if audio is not None else None
+        # Materialize the no-perturbation mask here (eager); a None config means "perturb nothing"
+        # -> all-keep masks. The block loop never builds masks.
+        if perturbations is None:
+            ref = (video_args or audio_args).x
+            perturbations = BatchedPerturbationConfig.empty(ref.shape[0], self.num_blocks, ref.device, ref.dtype)
         # Process transformer blocks
         video_out, audio_out = self._process_transformer_blocks(
             video=video_args,
@@ -436,10 +483,15 @@ class LegacyX0Model(torch.nn.Module):
     Returns fully denoised output based on the velocities produced by the base model.
     """
 
-    def __init__(self, velocity_model: LTXModel):
+    def __init__(self, velocity_model: LTXModelProtocol):
         super().__init__()
         self.velocity_model = velocity_model
 
+    @property
+    def num_blocks(self) -> int:
+        """Number of transformer blocks."""
+        return self.velocity_model.num_blocks
+
     def forward(
         self,
         video: Modality | None,
@@ -465,15 +517,20 @@ class X0Model(torch.nn.Module):
     Applies scaled denoising to the video and audio according to the timesteps = sigma * denoising_mask.
     """
 
-    def __init__(self, velocity_model: LTXModel):
+    def __init__(self, velocity_model: LTXModelProtocol):
         super().__init__()
         self.velocity_model = velocity_model
 
+    @property
+    def num_blocks(self) -> int:
+        """Number of transformer blocks."""
+        return self.velocity_model.num_blocks
+
     def forward(
         self,
         video: Modality | None,
         audio: Modality | None,
-        perturbations: BatchedPerturbationConfig,
+        perturbations: BatchedPerturbationConfig | None,
     ) -> tuple[torch.Tensor | None, torch.Tensor | None]:
         """
         Denoise the video and audio according to the sigma.

packages/ltx-core/src/ltx_core/model/transformer/model_configurator.py

@@ -2,10 +2,10 @@ import torch
 
 from ltx_core.loader.sd_ops import SDOps
 from ltx_core.model.model_protocol import ModelConfigurator
-from ltx_core.model.transformer.attention import AttentionFunction
 from ltx_core.model.transformer.model import LTXModel, LTXModelType
 from ltx_core.model.transformer.rope import LTXRopeType
 from ltx_core.model.transformer.text_projection import create_caption_projection
+from ltx_core.model.transformer.transformer import DEFAULT_TRANSFORMER_OPS, TransformerOpsConfig
 from ltx_core.utils import check_config_value
 
 
@@ -16,7 +16,7 @@ class LTXModelConfigurator(ModelConfigurator[LTXModel]):
     """
 
     @classmethod
-    def from_config(cls: type[LTXModel], config: dict) -> LTXModel:
+    def from_config(cls, config: dict, ops: TransformerOpsConfig = DEFAULT_TRANSFORMER_OPS) -> LTXModel:
         # Build caption projections for 19B models (projection handled in transformer).
         caption_projection, audio_caption_projection = _build_caption_projections(config, is_av=True)
 
@@ -40,6 +40,7 @@ class LTXModelConfigurator(ModelConfigurator[LTXModel]):
         check_config_value(config, "share_ff", False)
         check_config_value(config, "av_cross_ada_norm", True)
         check_config_value(config, "use_middle_indices_grid", True)
+        check_config_value(config, "num_attention_heads", config.get("audio_num_attention_heads", float("nan")))
 
         return LTXModel(
             model_type=LTXModelType.AudioVideo,
@@ -50,7 +51,7 @@ class LTXModelConfigurator(ModelConfigurator[LTXModel]):
             num_layers=config.get("num_layers", 48),
             cross_attention_dim=config.get("cross_attention_dim", 4096),
             norm_eps=config.get("norm_eps", 1e-06),
-            attention_type=AttentionFunction(config.get("attention_type", "default")),
+            ops=ops,
             positional_embedding_theta=config.get("positional_embedding_theta", 10000.0),
             positional_embedding_max_pos=config.get("positional_embedding_max_pos", [20, 2048, 2048]),
             timestep_scale_multiplier=config.get("timestep_scale_multiplier", 1000),
@@ -62,12 +63,13 @@ class LTXModelConfigurator(ModelConfigurator[LTXModel]):
             audio_cross_attention_dim=config.get("audio_cross_attention_dim", 2048),
             audio_positional_embedding_max_pos=config.get("audio_positional_embedding_max_pos", [20]),
             av_ca_timestep_scale_multiplier=config.get("av_ca_timestep_scale_multiplier", 1),
-            rope_type=LTXRopeType(config.get("rope_type", "interleaved")),
+            rope_type=LTXRopeType(config.get("rope_type", "split")),
             double_precision_rope=config.get("frequencies_precision", False) == "float64",
             apply_gated_attention=config.get("apply_gated_attention", False),
             caption_projection=caption_projection,
             audio_caption_projection=audio_caption_projection,
             cross_attention_adaln=config.get("cross_attention_adaln", False),
+            mask_conditioning_channels=config.get("mask_conditioning_channels", 0),
         )
 
 
@@ -78,7 +80,7 @@ class LTXVideoOnlyModelConfigurator(ModelConfigurator[LTXModel]):
     """
 
     @classmethod
-    def from_config(cls: type[LTXModel], config: dict) -> LTXModel:
+    def from_config(cls, config: dict, ops: TransformerOpsConfig = DEFAULT_TRANSFORMER_OPS) -> LTXModel:
         # Build caption projection for 19B model (projection handled in transformer).
         caption_projection, _ = _build_caption_projections(config, is_av=False)
 
@@ -109,16 +111,69 @@ class LTXVideoOnlyModelConfigurator(ModelConfigurator[LTXModel]):
             num_layers=config.get("num_layers", 48),
             cross_attention_dim=config.get("cross_attention_dim", 4096),
             norm_eps=config.get("norm_eps", 1e-06),
-            attention_type=AttentionFunction(config.get("attention_type", "default")),
+            ops=ops,
             positional_embedding_theta=config.get("positional_embedding_theta", 10000.0),
             positional_embedding_max_pos=config.get("positional_embedding_max_pos", [20, 2048, 2048]),
             timestep_scale_multiplier=config.get("timestep_scale_multiplier", 1000),
             use_middle_indices_grid=config.get("use_middle_indices_grid", True),
-            rope_type=LTXRopeType(config.get("rope_type", "interleaved")),
+            rope_type=LTXRopeType(config.get("rope_type", "split")),
             double_precision_rope=config.get("frequencies_precision", False) == "float64",
             apply_gated_attention=config.get("apply_gated_attention", False),
             caption_projection=caption_projection,
             cross_attention_adaln=config.get("cross_attention_adaln", False),
+            mask_conditioning_channels=config.get("mask_conditioning_channels", 0),
+        )
+
+
+class LTXAudioOnlyModelConfigurator(ModelConfigurator[LTXModel]):
+    """
+    Configurator for LTX audio only model.
+    Builds an audio-only LTX model (``model_type=AudioOnly``) so the video
+    transformer weights are never instantiated or loaded. Useful for
+    text-to-audio inference where the video branch is unused.
+    """
+
+    @classmethod
+    def from_config(cls, config: dict, ops: TransformerOpsConfig = DEFAULT_TRANSFORMER_OPS) -> LTXModel:
+        # Build audio caption projection for 19B models (projection handled in transformer).
+        _, audio_caption_projection = _build_caption_projections(config, is_av=True)
+
+        config = config.get("transformer", {})
+
+        check_config_value(config, "dropout", 0.0)
+        check_config_value(config, "attention_bias", True)
+        check_config_value(config, "num_vector_embeds", None)
+        check_config_value(config, "activation_fn", "gelu-approximate")
+        check_config_value(config, "num_embeds_ada_norm", 1000)
+        check_config_value(config, "use_linear_projection", False)
+        check_config_value(config, "only_cross_attention", False)
+        check_config_value(config, "cross_attention_norm", True)
+        check_config_value(config, "double_self_attention", False)
+        check_config_value(config, "upcast_attention", False)
+        check_config_value(config, "standardization_norm", "rms_norm")
+        check_config_value(config, "norm_elementwise_affine", False)
+        check_config_value(config, "qk_norm", "rms_norm")
+        check_config_value(config, "positional_embedding_type", "rope")
+        check_config_value(config, "use_middle_indices_grid", True)
+
+        return LTXModel(
+            model_type=LTXModelType.AudioOnly,
+            num_layers=config.get("num_layers", 48),
+            norm_eps=config.get("norm_eps", 1e-06),
+            ops=ops,
+            timestep_scale_multiplier=config.get("timestep_scale_multiplier", 1000),
+            use_middle_indices_grid=config.get("use_middle_indices_grid", True),
+            audio_num_attention_heads=config.get("audio_num_attention_heads", 32),
+            audio_attention_head_dim=config.get("audio_attention_head_dim", 64),
+            audio_in_channels=config.get("audio_in_channels", 128),
+            audio_out_channels=config.get("audio_out_channels", 128),
+            audio_cross_attention_dim=config.get("audio_cross_attention_dim", 2048),
+            audio_positional_embedding_max_pos=config.get("audio_positional_embedding_max_pos", [20]),
+            rope_type=LTXRopeType(config.get("rope_type", "split")),
+            double_precision_rope=config.get("frequencies_precision", False) == "float64",
+            apply_gated_attention=config.get("apply_gated_attention", False),
+            audio_caption_projection=audio_caption_projection,
+            cross_attention_adaln=config.get("cross_attention_adaln", False),
         )
 
 
@@ -150,3 +205,16 @@ LTXV_MODEL_COMFY_RENAMING_MAP = (
     .with_matching(prefix="model.diffusion_model.")
     .with_replacement("model.diffusion_model.", "")
 )
+
+LTXV_AUDIO_ONLY_MODEL_COMFY_RENAMING_MAP = (
+    SDOps("LTXV_AUDIO_ONLY_MODEL_COMFY_MAP")
+    .with_matching(prefix="model.diffusion_model.", contains="audio_attn1")
+    .with_matching(prefix="model.diffusion_model.", contains="audio_attn2")
+    .with_matching(prefix="model.diffusion_model.", contains="audio_ff")
+    .with_matching(prefix="model.diffusion_model.", contains="audio_patchify")
+    .with_matching(prefix="model.diffusion_model.", contains="audio_proj_out")
+    .with_matching(prefix="model.diffusion_model.", contains="audio_adaln_single")
+    .with_matching(prefix="model.diffusion_model.", contains="audio_prompt")
+    .with_matching(prefix="model.diffusion_model.", contains="audio_scale_shift_table")
+    .with_replacement("model.diffusion_model.", "")
+)

packages/ltx-core/src/ltx_core/model/transformer/ops.py

@@ -0,0 +1,106 @@
+from typing import List, Protocol
+
+import torch
+from torch import nn
+
+from ltx_core.model.transformer.rope import apply_rotary_emb
+from ltx_core.utils import rms_norm
+
+
+class PreAttentionCallable(Protocol):
+    def __call__(
+        self,
+        q: torch.Tensor,
+        k: torch.Tensor,
+        attn_module: nn.Module,
+        mask: torch.Tensor | None,
+        pe: torch.Tensor | None,
+        k_pe: torch.Tensor | None,
+    ) -> tuple[torch.Tensor, torch.Tensor]: ...
+
+
+class PytorchPreAttention(PreAttentionCallable):
+    def __call__(
+        self,
+        q: torch.Tensor,
+        k: torch.Tensor,
+        attn_module: nn.Module,
+        mask: torch.Tensor | None,  # noqa: ARG002
+        pe: torch.Tensor | None,
+        k_pe: torch.Tensor | None,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        q = attn_module.q_norm(q)
+        k = attn_module.k_norm(k)
+        if pe is not None:
+            q = apply_rotary_emb(q, pe, attn_module.rope_type)
+            k = apply_rotary_emb(k, pe if k_pe is None else k_pe, attn_module.rope_type)
+        return q, k
+
+
+class AdaZeroCallable(Protocol):
+    def __call__(
+        self,
+        x: torch.Tensor,
+        eps: float,
+        scale: torch.Tensor,
+        shift: torch.Tensor,
+    ) -> torch.Tensor: ...
+
+
+class PytorchAdaZeroFunction(AdaZeroCallable):
+    def __call__(
+        self,
+        x: torch.Tensor,
+        eps: float,
+        scale: torch.Tensor,
+        shift: torch.Tensor,
+    ) -> torch.Tensor:
+        return rms_norm(x, eps=eps) * (1 + scale) + shift
+
+
+class PostSACallable(Protocol):
+    def __call__(
+        self,
+        x: torch.Tensor,
+        y: torch.Tensor,
+        norm_weights: torch.Tensor | None,
+        eps: float,
+        gate: torch.Tensor,
+    ) -> List[torch.Tensor]: ...
+
+
+class PytorchPostSAFunction(PostSACallable):
+    def __call__(
+        self,
+        x: torch.Tensor,
+        y: torch.Tensor,
+        norm_weights: torch.Tensor | None,
+        eps: float,
+        gate: torch.Tensor,
+    ) -> List[torch.Tensor]:
+        x_fma = x + y * gate
+        return x_fma, rms_norm(x_fma, norm_weights, eps=eps)
+
+
+class GatedAttentionCallable(Protocol):
+    def __call__(
+        self,
+        x: torch.Tensor,
+        attn_out: torch.Tensor,
+        attn_module: nn.Module,
+    ) -> torch.Tensor: ...
+
+
+class PytorchGatedAttention(GatedAttentionCallable):
+    def __call__(
+        self,
+        x: torch.Tensor,
+        attn_out: torch.Tensor,
+        attn_module: nn.Module,
+    ) -> torch.Tensor:
+        gate_logits = attn_module.to_gate_logits(x)  # (B, T, H)
+        b, t, _ = attn_out.shape
+        out = attn_out.view(b, t, attn_module.heads, attn_module.dim_head)
+        gates = 2.0 * torch.sigmoid(gate_logits)  # (B, T, H)
+        out = out * gates.unsqueeze(-1)  # (B, T, H, D) * (B, T, H, 1)
+        return out.view(b, t, attn_module.heads * attn_module.dim_head)

packages/ltx-core/src/ltx_core/model/transformer/rope.py

@@ -16,9 +16,10 @@ class LTXRopeType(Enum):
 def apply_rotary_emb(
     input_tensor: torch.Tensor,
     freqs_cis: Tuple[torch.Tensor, torch.Tensor],
-    rope_type: LTXRopeType = LTXRopeType.INTERLEAVED,
+    rope_type: LTXRopeType = LTXRopeType.SPLIT,
 ) -> torch.Tensor:
     if rope_type == LTXRopeType.INTERLEAVED:
+        # Note: INTERLEAVED rope is a legacy mode. Prefer SPLIT instead.
         return apply_interleaved_rotary_emb(input_tensor, *freqs_cis)
     elif rope_type == LTXRopeType.SPLIT:
         return apply_split_rotary_emb(input_tensor, *freqs_cis)
@@ -42,11 +43,26 @@ def apply_interleaved_rotary_emb(
 def apply_split_rotary_emb(
     input_tensor: torch.Tensor, cos_freqs: torch.Tensor, sin_freqs: torch.Tensor
 ) -> torch.Tensor:
-    needs_reshape = False
+    if sin_freqs.shape != cos_freqs.shape:
-    if input_tensor.ndim != 4 and cos_freqs.ndim == 4:
+        raise ValueError(
-        b, h, t, _ = cos_freqs.shape
+            f"apply_split_rotary_emb: sin_freqs.shape {tuple(sin_freqs.shape)} must equal "
-        input_tensor = input_tensor.reshape(b, t, h, -1).swapaxes(1, 2)
+            f"cos_freqs.shape {tuple(cos_freqs.shape)}."
-        needs_reshape = True
+        )
+    needs_reshape = input_tensor.ndim != 4 and cos_freqs.ndim == 4
+    if needs_reshape:
+        b_freq = cos_freqs.shape[0]
+        h = cos_freqs.shape[1]
+        b_in = input_tensor.shape[0]
+        if b_freq not in (1, b_in):
+            raise ValueError(
+                f"apply_split_rotary_emb: cos_freqs batch ({b_freq}) must be 1 "
+                f"(broadcast) or equal input_tensor batch ({b_in})."
+            )
+        # `unflatten` only touches the last dim, keeping the batch and seq dims as
+        # the input tensor's own symbolic ints under torch.compile. `reshape(b_in,
+        # t, h, -1)` would have forced Dynamo to specialise those dims because it
+        # cannot prove `b_in == cos_freqs.shape[0]` and `seq == t` across tensors.
+        input_tensor = input_tensor.unflatten(-1, (h, -1)).transpose(1, 2)
 
     split_input = rearrange(input_tensor, "... (d r) -> ... d r", d=2)
     first_half_input = split_input[..., :1, :]
@@ -61,7 +77,9 @@ def apply_split_rotary_emb(
 
     output = rearrange(output, "... d r -> ... (d r)")
     if needs_reshape:
-        output = output.swapaxes(1, 2).reshape(b, t, -1)
+        # `transpose(1, 2).flatten(-2)` keeps the batch and seq dims symbolic; using
+        # `reshape(b_in, t, -1)` would force Dynamo to specialise both axes.
+        output = output.transpose(1, 2).flatten(-2)
 
     return output
 
@@ -183,7 +201,7 @@ def precompute_freqs_cis(
     max_pos: list[int] | None = None,
     use_middle_indices_grid: bool = False,
     num_attention_heads: int = 32,
-    rope_type: LTXRopeType = LTXRopeType.INTERLEAVED,
+    rope_type: LTXRopeType = LTXRopeType.SPLIT,
     freq_grid_generator: Callable[[float, int, int, torch.device], torch.Tensor] = generate_freq_grid_pytorch,
 ) -> tuple[torch.Tensor, torch.Tensor]:
     if max_pos is None:

packages/ltx-core/src/ltx_core/model/transformer/transformer.py

@@ -1,14 +1,29 @@
-from dataclasses import dataclass, replace
+from dataclasses import dataclass, field, replace
 
 import torch
 
-from ltx_core.guidance.perturbations import BatchedPerturbationConfig, PerturbationType
 from ltx_core.model.transformer.adaln import adaln_embedding_coefficient
-from ltx_core.model.transformer.attention import Attention, AttentionCallable, AttentionFunction
+from ltx_core.model.transformer.attention import (
+    Attention,
+    AttentionCallable,
+    AttentionFunction,
+    AttentionOps,
+    MaskedAttentionCallable,
+    MaskedAttentionFunction,
+)
 from ltx_core.model.transformer.feed_forward import FeedForward
+from ltx_core.model.transformer.ops import (
+    AdaZeroCallable,
+    GatedAttentionCallable,
+    PostSACallable,
+    PreAttentionCallable,
+    PytorchAdaZeroFunction,
+    PytorchGatedAttention,
+    PytorchPostSAFunction,
+    PytorchPreAttention,
+)
 from ltx_core.model.transformer.rope import LTXRopeType
 from ltx_core.model.transformer.transformer_args import TransformerArgs
-from ltx_core.utils import rms_norm
 
 
 @dataclass
@@ -21,19 +36,68 @@ class TransformerConfig:
     cross_attention_adaln: bool = False
 
 
+@dataclass(frozen=True)
+class TransformerOpsConfig:
+    """Pluggable ops for :class:`BasicAVTransformerBlock`.
+    Use :meth:`from_functions` to construct from enum values or partial overrides
+    without spelling out a full :class:`AttentionOps`.
+    """
+
+    attention_ops: AttentionOps = field(default_factory=AttentionOps)
+    ada_zero_function: AdaZeroCallable = field(default_factory=PytorchAdaZeroFunction)
+    post_sa_function: PostSACallable = field(default_factory=PytorchPostSAFunction)
+
+    @classmethod
+    def from_functions(
+        cls,
+        attention: AttentionFunction | AttentionCallable = AttentionFunction.AUTOMATIC,
+        masked_attention: MaskedAttentionFunction | MaskedAttentionCallable = MaskedAttentionFunction.AUTOMATIC,
+        preattention: PreAttentionCallable | None = None,
+        gated_attention: GatedAttentionCallable | None = None,
+        ada_zero: AdaZeroCallable | None = None,
+        post_sa: PostSACallable | None = None,
+    ) -> "TransformerOpsConfig":
+        """Build a config from individual functions or enums. Each *None* slot
+        falls back to the standard PyTorch implementation."""
+        attention_callable = attention.to_callable() if isinstance(attention, AttentionFunction) else attention
+        masked_callable = (
+            masked_attention.to_callable()
+            if isinstance(masked_attention, MaskedAttentionFunction)
+            else masked_attention
+        )
+        attention_ops = AttentionOps(
+            attention_function=attention_callable,
+            masked_attention_function=masked_callable,
+            preattention_function=preattention if preattention is not None else PytorchPreAttention(),
+            gated_attention_function=(gated_attention if gated_attention is not None else PytorchGatedAttention()),
+        )
+        return cls(
+            attention_ops=attention_ops,
+            ada_zero_function=ada_zero if ada_zero is not None else PytorchAdaZeroFunction(),
+            post_sa_function=post_sa if post_sa is not None else PytorchPostSAFunction(),
+        )
+
+
+# Frozen, so safe to share as a default argument across callers that want the
+# stock PyTorch ops without explicit construction.
+DEFAULT_TRANSFORMER_OPS = TransformerOpsConfig()
+
+
 class BasicAVTransformerBlock(torch.nn.Module):
     def __init__(
         self,
-        idx: int,
         video: TransformerConfig | None = None,
         audio: TransformerConfig | None = None,
-        rope_type: LTXRopeType = LTXRopeType.INTERLEAVED,
+        rope_type: LTXRopeType = LTXRopeType.SPLIT,
         norm_eps: float = 1e-6,
-        attention_function: AttentionFunction | AttentionCallable = AttentionFunction.DEFAULT,
+        ops: TransformerOpsConfig | None = None,
     ):
         super().__init__()
 
-        self.idx = idx
+        if ops is None:
+            ops = TransformerOpsConfig()
+        self.ada_zero_function = ops.ada_zero_function
+        self.post_sa_function = ops.post_sa_function
         if video is not None:
             self.attn1 = Attention(
                 query_dim=video.dim,
@@ -42,7 +106,7 @@ class BasicAVTransformerBlock(torch.nn.Module):
                 context_dim=None,
                 rope_type=rope_type,
                 norm_eps=norm_eps,
-                attention_function=attention_function,
+                ops=ops.attention_ops,
                 apply_gated_attention=video.apply_gated_attention,
             )
             self.attn2 = Attention(
@@ -52,7 +116,7 @@ class BasicAVTransformerBlock(torch.nn.Module):
                 dim_head=video.d_head,
                 rope_type=rope_type,
                 norm_eps=norm_eps,
-                attention_function=attention_function,
+                ops=ops.attention_ops,
                 apply_gated_attention=video.apply_gated_attention,
             )
             self.ff = FeedForward(video.dim, dim_out=video.dim)
@@ -67,7 +131,7 @@ class BasicAVTransformerBlock(torch.nn.Module):
                 context_dim=None,
                 rope_type=rope_type,
                 norm_eps=norm_eps,
-                attention_function=attention_function,
+                ops=ops.attention_ops,
                 apply_gated_attention=audio.apply_gated_attention,
             )
             self.audio_attn2 = Attention(
@@ -77,7 +141,7 @@ class BasicAVTransformerBlock(torch.nn.Module):
                 dim_head=audio.d_head,
                 rope_type=rope_type,
                 norm_eps=norm_eps,
-                attention_function=attention_function,
+                ops=ops.attention_ops,
                 apply_gated_attention=audio.apply_gated_attention,
             )
             self.audio_ff = FeedForward(audio.dim, dim_out=audio.dim)
@@ -93,7 +157,7 @@ class BasicAVTransformerBlock(torch.nn.Module):
                 dim_head=audio.d_head,
                 rope_type=rope_type,
                 norm_eps=norm_eps,
-                attention_function=attention_function,
+                ops=ops.attention_ops,
                 apply_gated_attention=video.apply_gated_attention,
             )
 
@@ -105,7 +169,7 @@ class BasicAVTransformerBlock(torch.nn.Module):
                 dim_head=audio.d_head,
                 rope_type=rope_type,
                 norm_eps=norm_eps,
-                attention_function=attention_function,
+                ops=ops.attention_ops,
                 apply_gated_attention=audio.apply_gated_attention,
             )
 
@@ -157,7 +221,7 @@ class BasicAVTransformerBlock(torch.nn.Module):
 
     def _apply_text_cross_attention(
         self,
-        x: torch.Tensor,
+        x_normed: torch.Tensor,
         context: torch.Tensor,
         attn: AttentionCallable,
         scale_shift_table: torch.Tensor,
@@ -167,11 +231,14 @@ class BasicAVTransformerBlock(torch.nn.Module):
         context_mask: torch.Tensor | None,
         cross_attention_adaln: bool = False,
     ) -> torch.Tensor:
-        """Apply text cross-attention, with optional AdaLN modulation."""
+        """Apply text cross-attention, with optional AdaLN modulation.
+        ``x_normed`` is the RMS-normalized self-attention output produced by
+        ``post_sa_function`` -- this method does not normalize again.
+        """
         if cross_attention_adaln:
-            shift_q, scale_q, gate = self.get_ada_values(scale_shift_table, x.shape[0], timestep, slice(6, 9))
+            shift_q, scale_q, gate = self.get_ada_values(scale_shift_table, x_normed.shape[0], timestep, slice(6, 9))
             return apply_cross_attention_adaln(
-                x,
+                x_normed,
                 context,
                 attn,
                 shift_q,
@@ -180,24 +247,17 @@ class BasicAVTransformerBlock(torch.nn.Module):
                 prompt_scale_shift_table,
                 prompt_timestep,
                 context_mask,
-                self.norm_eps,
             )
-        return attn(rms_norm(x, eps=self.norm_eps), context=context, mask=context_mask)
+        return attn(x_normed, context=context, mask=context_mask)
 
     def forward(  # noqa: PLR0915
         self,
         video: TransformerArgs | None,
         audio: TransformerArgs | None,
-        perturbations: BatchedPerturbationConfig | None = None,
     ) -> tuple[TransformerArgs | None, TransformerArgs | None]:
         if video is None and audio is None:
             raise ValueError("At least one of video or audio must be provided")
 
-        batch_size = (video or audio).x.shape[0]
-
-        if perturbations is None:
-            perturbations = BatchedPerturbationConfig.empty(batch_size)
-
         vx = video.x if video is not None else None
         ax = audio.x if audio is not None else None
 
@@ -211,30 +271,20 @@ class BasicAVTransformerBlock(torch.nn.Module):
             vshift_msa, vscale_msa, vgate_msa = self.get_ada_values(
                 self.scale_shift_table, vx.shape[0], video.timesteps, slice(0, 3)
             )
-            norm_vx = rms_norm(vx, eps=self.norm_eps) * (1 + vscale_msa) + vshift_msa
+            norm_vx = self.ada_zero_function(vx, self.norm_eps, vscale_msa, vshift_msa)
             del vshift_msa, vscale_msa
 
-            all_perturbed = perturbations.all_in_batch(PerturbationType.SKIP_VIDEO_SELF_ATTN, self.idx)
+            vx_msa_out = self.attn1(
-            none_perturbed = not perturbations.any_in_batch(PerturbationType.SKIP_VIDEO_SELF_ATTN, self.idx)
-            v_mask = (
-                perturbations.mask_like(PerturbationType.SKIP_VIDEO_SELF_ATTN, self.idx, vx)
-                if not all_perturbed and not none_perturbed
-                else None
-            )
-            vx = (
-                vx
-                + self.attn1(
                 norm_vx,
                 pe=video.positional_embeddings,
                 mask=video.self_attention_mask,
-                    perturbation_mask=v_mask,
+                perturbation_mask=video.self_attn_perturbation_mask,
-                    all_perturbed=all_perturbed,
+                all_perturbed=video.self_attn_all_perturbed,
             )
-                * vgate_msa
+            vx, vx_normed = self.post_sa_function(vx, vx_msa_out, None, self.norm_eps, vgate_msa)
-            )
+            del vgate_msa, norm_vx, vx_msa_out
-            del vgate_msa, norm_vx, v_mask
             vx = vx + self._apply_text_cross_attention(
-                vx,
+                vx_normed,
                 video.context,
                 self.attn2,
                 self.scale_shift_table,
@@ -244,35 +294,26 @@ class BasicAVTransformerBlock(torch.nn.Module):
                 video.context_mask,
                 cross_attention_adaln=self.cross_attention_adaln,
             )
+            del vx_normed
 
         if run_ax:
             ashift_msa, ascale_msa, agate_msa = self.get_ada_values(
                 self.audio_scale_shift_table, ax.shape[0], audio.timesteps, slice(0, 3)
             )
 
-            norm_ax = rms_norm(ax, eps=self.norm_eps) * (1 + ascale_msa) + ashift_msa
+            norm_ax = self.ada_zero_function(ax, self.norm_eps, ascale_msa, ashift_msa)
             del ashift_msa, ascale_msa
-            all_perturbed = perturbations.all_in_batch(PerturbationType.SKIP_AUDIO_SELF_ATTN, self.idx)
+            ax_msa_out = self.audio_attn1(
-            none_perturbed = not perturbations.any_in_batch(PerturbationType.SKIP_AUDIO_SELF_ATTN, self.idx)
-            a_mask = (
-                perturbations.mask_like(PerturbationType.SKIP_AUDIO_SELF_ATTN, self.idx, ax)
-                if not all_perturbed and not none_perturbed
-                else None
-            )
-            ax = (
-                ax
-                + self.audio_attn1(
                 norm_ax,
                 pe=audio.positional_embeddings,
                 mask=audio.self_attention_mask,
-                    perturbation_mask=a_mask,
+                perturbation_mask=audio.self_attn_perturbation_mask,
-                    all_perturbed=all_perturbed,
+                all_perturbed=audio.self_attn_all_perturbed,
             )
-                * agate_msa
+            ax, ax_normed = self.post_sa_function(ax, ax_msa_out, None, self.norm_eps, agate_msa)
-            )
+            del agate_msa, norm_ax, ax_msa_out
-            del agate_msa, norm_ax, a_mask
             ax = ax + self._apply_text_cross_attention(
-                ax,
+                ax_normed,
                 audio.context,
                 self.audio_attn2,
                 self.audio_scale_shift_table,
@@ -282,13 +323,15 @@ class BasicAVTransformerBlock(torch.nn.Module):
                 audio.context_mask,
                 cross_attention_adaln=self.cross_attention_adaln,
             )
+            del ax_normed
 
         # Audio - Video cross attention.
         if run_a2v or run_v2a:
-            vx_norm3 = rms_norm(vx, eps=self.norm_eps)
+            # Snapshot vx/ax before A2V mutates vx; V2A's video keys/values must
-            ax_norm3 = rms_norm(ax, eps=self.norm_eps)
+            # use the pre-A2V state so direction order doesn't bias the result.
-
+            vx_pre_av = vx
-            if run_a2v and not perturbations.all_in_batch(PerturbationType.SKIP_A2V_CROSS_ATTN, self.idx):
+            ax_pre_av = ax
+            if run_a2v and not video.cross_attn_skip_all:
                 scale_ca_video_a2v, shift_ca_video_a2v, gate_out_a2v = self.get_av_ca_ada_values(
                     self.scale_shift_table_a2v_ca_video,
                     vx.shape[0],
@@ -296,7 +339,7 @@ class BasicAVTransformerBlock(torch.nn.Module):
                     video.cross_gate_timestep,
                     slice(0, 2),
                 )
-                vx_scaled = vx_norm3 * (1 + scale_ca_video_a2v) + shift_ca_video_a2v
+                a2v_vx_scaled = self.ada_zero_function(vx_pre_av, self.norm_eps, scale_ca_video_a2v, shift_ca_video_a2v)
                 del scale_ca_video_a2v, shift_ca_video_a2v
 
                 scale_ca_audio_a2v, shift_ca_audio_a2v, _ = self.get_av_ca_ada_values(
@@ -306,22 +349,21 @@ class BasicAVTransformerBlock(torch.nn.Module):
                     audio.cross_gate_timestep,
                     slice(0, 2),
                 )
-                ax_scaled = ax_norm3 * (1 + scale_ca_audio_a2v) + shift_ca_audio_a2v
+                a2v_ax_scaled = self.ada_zero_function(ax_pre_av, self.norm_eps, scale_ca_audio_a2v, shift_ca_audio_a2v)
                 del scale_ca_audio_a2v, shift_ca_audio_a2v
-                a2v_mask = perturbations.mask_like(PerturbationType.SKIP_A2V_CROSS_ATTN, self.idx, vx)
                 vx = vx + (
                     self.audio_to_video_attn(
-                        vx_scaled,
+                        a2v_vx_scaled,
-                        context=ax_scaled,
+                        context=a2v_ax_scaled,
                         pe=video.cross_positional_embeddings,
                         k_pe=audio.cross_positional_embeddings,
                     )
                     * gate_out_a2v
-                    * a2v_mask
+                    * video.cross_attn_perturbation_mask
                 )
-                del gate_out_a2v, a2v_mask, vx_scaled, ax_scaled
+                del gate_out_a2v, a2v_vx_scaled, a2v_ax_scaled
 
-            if run_v2a and not perturbations.all_in_batch(PerturbationType.SKIP_V2A_CROSS_ATTN, self.idx):
+            if run_v2a and not audio.cross_attn_skip_all:
                 scale_ca_audio_v2a, shift_ca_audio_v2a, gate_out_v2a = self.get_av_ca_ada_values(
                     self.scale_shift_table_a2v_ca_audio,
                     ax.shape[0],
@@ -329,7 +371,7 @@ class BasicAVTransformerBlock(torch.nn.Module):
                     audio.cross_gate_timestep,
                     slice(2, 4),
                 )
-                ax_scaled = ax_norm3 * (1 + scale_ca_audio_v2a) + shift_ca_audio_v2a
+                v2a_ax_scaled = self.ada_zero_function(ax_pre_av, self.norm_eps, scale_ca_audio_v2a, shift_ca_audio_v2a)
                 del scale_ca_audio_v2a, shift_ca_audio_v2a
                 scale_ca_video_v2a, shift_ca_video_v2a, _ = self.get_av_ca_ada_values(
                     self.scale_shift_table_a2v_ca_video,
@@ -338,28 +380,26 @@ class BasicAVTransformerBlock(torch.nn.Module):
                     video.cross_gate_timestep,
                     slice(2, 4),
                 )
-                vx_scaled = vx_norm3 * (1 + scale_ca_video_v2a) + shift_ca_video_v2a
+                v2a_vx_scaled = self.ada_zero_function(vx_pre_av, self.norm_eps, scale_ca_video_v2a, shift_ca_video_v2a)
                 del scale_ca_video_v2a, shift_ca_video_v2a
-                v2a_mask = perturbations.mask_like(PerturbationType.SKIP_V2A_CROSS_ATTN, self.idx, ax)
                 ax = ax + (
                     self.video_to_audio_attn(
-                        ax_scaled,
+                        v2a_ax_scaled,
-                        context=vx_scaled,
+                        context=v2a_vx_scaled,
                         pe=audio.cross_positional_embeddings,
                         k_pe=video.cross_positional_embeddings,
                     )
                     * gate_out_v2a
-                    * v2a_mask
+                    * audio.cross_attn_perturbation_mask
                 )
-                del gate_out_v2a, v2a_mask, ax_scaled, vx_scaled
+                del gate_out_v2a, v2a_vx_scaled, v2a_ax_scaled
-
+            del vx_pre_av, ax_pre_av
-            del vx_norm3, ax_norm3
 
         if run_vx:
             vshift_mlp, vscale_mlp, vgate_mlp = self.get_ada_values(
                 self.scale_shift_table, vx.shape[0], video.timesteps, slice(3, 6)
             )
-            vx_scaled = rms_norm(vx, eps=self.norm_eps) * (1 + vscale_mlp) + vshift_mlp
+            vx_scaled = self.ada_zero_function(vx, self.norm_eps, vscale_mlp, vshift_mlp)
             vx = vx + self.ff(vx_scaled) * vgate_mlp
 
             del vshift_mlp, vscale_mlp, vgate_mlp, vx_scaled
@@ -368,7 +408,7 @@ class BasicAVTransformerBlock(torch.nn.Module):
             ashift_mlp, ascale_mlp, agate_mlp = self.get_ada_values(
                 self.audio_scale_shift_table, ax.shape[0], audio.timesteps, slice(3, 6)
             )
-            ax_scaled = rms_norm(ax, eps=self.norm_eps) * (1 + ascale_mlp) + ashift_mlp
+            ax_scaled = self.ada_zero_function(ax, self.norm_eps, ascale_mlp, ashift_mlp)
             ax = ax + self.audio_ff(ax_scaled) * agate_mlp
 
             del ashift_mlp, ascale_mlp, agate_mlp, ax_scaled
@@ -377,7 +417,7 @@ class BasicAVTransformerBlock(torch.nn.Module):
 
 
 def apply_cross_attention_adaln(
-    x: torch.Tensor,
+    x_normed: torch.Tensor,
     context: torch.Tensor,
     attn: AttentionCallable,
     q_shift: torch.Tensor,
@@ -386,13 +426,17 @@ def apply_cross_attention_adaln(
     prompt_scale_shift_table: torch.Tensor,
     prompt_timestep: torch.Tensor,
     context_mask: torch.Tensor | None = None,
-    norm_eps: float = 1e-6,
 ) -> torch.Tensor:
-    batch_size = x.shape[0]
+    """Apply query/key AdaLN modulation then cross-attention.
+    ``x_normed`` is already RMS-normalized by ``post_sa_function``; this only
+    applies the affine (scale/shift) modulation, so the normalization is not
+    repeated here.
+    """
+    batch_size = x_normed.shape[0]
     shift_kv, scale_kv = (
-        prompt_scale_shift_table[None, None].to(device=x.device, dtype=x.dtype)
+        prompt_scale_shift_table[None, None].to(device=x_normed.device, dtype=x_normed.dtype)
         + prompt_timestep.reshape(batch_size, prompt_timestep.shape[1], 2, -1)
     ).unbind(dim=2)
-    attn_input = rms_norm(x, eps=norm_eps) * (1 + q_scale) + q_shift
+    attn_input = x_normed * (1 + q_scale) + q_shift
     encoder_hidden_states = context * (1 + scale_kv) + shift_kv
     return attn(attn_input, context=encoder_hidden_states, mask=context_mask) * q_gate

packages/ltx-core/src/ltx_core/model/transformer/transformer_args.py

@@ -2,6 +2,7 @@ from dataclasses import dataclass, replace
 
 import torch
 
+from ltx_core.guidance.perturbations import BatchedPerturbationConfig, PerturbationType
 from ltx_core.model.transformer.adaln import AdaLayerNormSingle
 from ltx_core.model.transformer.modality import Modality
 from ltx_core.model.transformer.rope import (
@@ -19,8 +20,8 @@ class TransformerArgs:
     context_mask: torch.Tensor
     timesteps: torch.Tensor
     embedded_timestep: torch.Tensor
-    positional_embeddings: torch.Tensor
+    positional_embeddings: tuple[torch.Tensor, torch.Tensor]
-    cross_positional_embeddings: torch.Tensor | None
+    cross_positional_embeddings: tuple[torch.Tensor, torch.Tensor] | None
     cross_scale_shift_timestep: torch.Tensor | None
     cross_gate_timestep: torch.Tensor | None
     enabled: bool
@@ -28,6 +29,57 @@ class TransformerArgs:
     self_attention_mask: torch.Tensor | None = (
         None  # Additive log-space self-attention bias (B, 1, T, T), None = full attention
     )
+    # Per-block perturbation state, precomputed by `LTXModel._process_transformer_blocks`
+    # so the block forward needs no per-block identity. The bool shortcuts
+    # (`*_all_perturbed`, `cross_attn_skip_all`) are Python bools that Dynamo specialises
+    # on — fine because they're stable across denoising steps for a fixed perturbation
+    # config.
+    self_attn_perturbation_mask: torch.Tensor | None = None
+    self_attn_all_perturbed: bool = False
+    cross_attn_perturbation_mask: torch.Tensor | None = None
+    cross_attn_skip_all: bool = False
+
+
+class BlockPerturbationsProcessor:
+    """Per-block preparation of ``TransformerArgs``.
+    The base implementation returns a copy of ``args`` with this block's
+    precomputed perturbation flags and masks attached. Subclasses can layer in
+    operations that must run on each block's inputs but stay outside the
+    compile boundary -- e.g. ``torch._dynamo.mark_dynamic`` for
+    shape-polymorphic block compilation (see ``compiling.py``). Swapping the
+    processor on an ``LTXModel`` instance is how compile transforms opt in to
+    such behaviour without baking it into the model's forward.
+    ``self_attn_perturbation_mask`` is None when all or none of the batch is
+    perturbed (the attention call can take the shortcut path). ``cross_attn_*``
+    is None when every sample skips the cross-attention entirely.
+    """
+
+    def __call__(
+        self,
+        args: "TransformerArgs",
+        perturbations: BatchedPerturbationConfig,
+        block_idx: int,
+        self_attn_type: PerturbationType,
+        cross_attn_type: PerturbationType,
+    ) -> "TransformerArgs":
+        all_self = perturbations.all_in_batch(self_attn_type, block_idx)
+        any_self = perturbations.any_in_batch(self_attn_type, block_idx)
+        self_mask: torch.Tensor | None = None
+        if any_self and not all_self:
+            self_mask = perturbations.mask(self_attn_type, block_idx)
+
+        all_cross = perturbations.all_in_batch(cross_attn_type, block_idx)
+        cross_mask: torch.Tensor | None = None
+        if not all_cross:
+            cross_mask = perturbations.mask(cross_attn_type, block_idx)
+
+        return replace(
+            args,
+            self_attn_perturbation_mask=self_mask,
+            self_attn_all_perturbed=all_self,
+            cross_attn_perturbation_mask=cross_mask,
+            cross_attn_skip_all=all_cross,
+        )
 
 
 class TransformerArgsPreprocessor:
@@ -98,9 +150,12 @@ class TransformerArgsPreprocessor:
         self, attention_mask: torch.Tensor | None, x_dtype: torch.dtype
     ) -> torch.Tensor | None:
         """Prepare self-attention mask by converting [0,1] values to additive log-space bias.
-        Input shape: (B, T, T) with values in [0, 1].
+        Input shape: 3D ``(B, T_q, T_k)`` with values in [0, 1]. The dense form
-        Output shape: (B, 1, T, T) with 0.0 for full attention and a large negative value
+        is ``(B, T, T)``; broadcastable forms like ``(1, 1, T)`` (key-only
-        for masked positions.
+        padding) or ``(B, 1, T)`` are also valid and yield a correspondingly
+        broadcastable output.
+        Output shape: ``(B, 1, T_q, T_k)`` (heads dim inserted) with 0.0 for
+        full attention and a large negative value for masked positions.
         Positions with attention_mask <= 0 are fully masked (mapped to the dtype's minimum
         representable value). Strictly positive entries are converted via log-space for
         smooth attenuation, with small values clamped for numerical stability.
@@ -120,7 +175,7 @@ class TransformerArgsPreprocessor:
         if positive.any():
             bias[positive] = torch.log(attention_mask[positive].clamp(min=eps)).to(x_dtype)
 
-        return bias.unsqueeze(1)  # (B, 1, T, T) for head broadcast
+        return bias.unsqueeze(1)  # (B, 1, T_q, T_k) for head broadcast
 
     def _prepare_positional_embeddings(
         self,
@@ -146,12 +201,46 @@ class TransformerArgsPreprocessor:
         )
         return pe
 
+    def _apply_patchify_proj(self, modality: Modality) -> torch.Tensor:
+        """Project patchified latents, concatenating any in-context conditioning channels first.
+
+        When ``patchify_proj`` expects more input features than the latent provides (a model built
+        with ``mask_conditioning_channels > 0``), the extra width is filled by ``modality.cond_channels``
+        (per-token, never noised). If those channels are absent they default to zeros, so a widened
+        model still runs and, with zero-initialized new projection columns, reproduces the base model's
+        output exactly.
+        """
+        latent = modality.latent
+        expected = self.patchify_proj.in_features
+        actual = latent.shape[-1]
+        if expected != actual:
+            missing = expected - actual
+            if missing < 0:
+                raise ValueError(
+                    f"patchify_proj expects {expected} input features but the latent already has {actual}."
+                )
+            cond_channels = modality.cond_channels
+            if cond_channels is None:
+                cond_channels = latent.new_zeros(latent.shape[0], latent.shape[1], missing)
+            elif cond_channels.shape[-1] != missing:
+                raise ValueError(
+                    f"cond_channels has {cond_channels.shape[-1]} channels but patchify_proj needs {missing} "
+                    f"extra input features (latent {actual} + cond {cond_channels.shape[-1]} != {expected})."
+                )
+            elif cond_channels.shape[1] != latent.shape[1]:
+                raise ValueError(
+                    f"cond_channels token length {cond_channels.shape[1]} must match latent token length "
+                    f"{latent.shape[1]}."
+                )
+            latent = torch.cat([latent, cond_channels.to(dtype=latent.dtype)], dim=-1)
+        return self.patchify_proj(latent)
+
     def prepare(
         self,
         modality: Modality,
         cross_modality: Modality | None = None,  # noqa: ARG002
     ) -> TransformerArgs:
-        x = self.patchify_proj(modality.latent)
+        x = self._apply_patchify_proj(modality)
         batch_size = x.shape[0]
         timestep, embedded_timestep = self._prepare_timestep(
             modality.timesteps, self.adaln, batch_size, modality.latent.dtype
@@ -244,10 +333,6 @@ class MultiModalTransformerArgsPreprocessor:
             if cross_modality.sigma.ndim != 1:
                 raise ValueError("Cross modality sigma must be a 1D tensor")
 
-        cross_timestep = cross_modality.sigma.view(
-            modality.timesteps.shape[0], 1, *[1] * len(modality.timesteps.shape[2:])
-        )
-
         cross_pe = self.simple_preprocessor._prepare_positional_embeddings(
             positions=modality.positions[:, 0:1, :],
             inner_dim=self.audio_cross_attention_dim,
@@ -258,7 +343,8 @@ class MultiModalTransformerArgsPreprocessor:
         )
 
         cross_scale_shift_timestep, cross_gate_timestep = self._prepare_cross_attention_timestep(
-            timestep=cross_timestep,
+            modality_timesteps=modality.timesteps,
+            cross_modality_sigma=cross_modality.sigma,
             timestep_scale_multiplier=self.simple_preprocessor.timestep_scale_multiplier,
             batch_size=transformer_args.x.shape[0],
             hidden_dtype=modality.latent.dtype,
@@ -273,23 +359,23 @@ class MultiModalTransformerArgsPreprocessor:
 
     def _prepare_cross_attention_timestep(
         self,
-        timestep: torch.Tensor | None,
+        modality_timesteps: torch.Tensor,
+        cross_modality_sigma: torch.Tensor,
         timestep_scale_multiplier: int,
         batch_size: int,
         hidden_dtype: torch.dtype,
     ) -> tuple[torch.Tensor, torch.Tensor]:
-        """Prepare cross attention timestep embeddings."""
+        """Prepare A-V cross-attention AdaLN inputs."""
-        timestep = timestep * timestep_scale_multiplier
-
         av_ca_factor = self.av_ca_timestep_scale_multiplier / timestep_scale_multiplier
 
         scale_shift_timestep, _ = self.cross_scale_shift_adaln(
-            timestep.flatten(),
+            (modality_timesteps * timestep_scale_multiplier).flatten(),
             hidden_dtype=hidden_dtype,
         )
         scale_shift_timestep = scale_shift_timestep.view(batch_size, -1, scale_shift_timestep.shape[-1])
+
         gate_noise_timestep, _ = self.cross_gate_adaln(
-            timestep.flatten() * av_ca_factor,
+            (cross_modality_sigma * timestep_scale_multiplier * av_ca_factor).flatten(),
             hidden_dtype=hidden_dtype,
         )
         gate_noise_timestep = gate_noise_timestep.view(batch_size, -1, gate_noise_timestep.shape[-1])

packages/ltx-core/src/ltx_core/model/video_vae/__init__.py

@@ -1,5 +1,6 @@
 """Video VAE package."""
 
+from ltx_core.model.video_vae.memory_efficient_decode import MEMORY_EFFICIENT_DECODE
 from ltx_core.model.video_vae.model_configurator import (
     VAE_DECODER_COMFY_KEYS_FILTER,
     VAE_ENCODER_COMFY_KEYS_FILTER,
@@ -10,6 +11,7 @@ from ltx_core.model.video_vae.tiling import SpatialTilingConfig, TemporalTilingC
 from ltx_core.model.video_vae.video_vae import VideoDecoder, VideoEncoder, get_video_chunks_number
 
 __all__ = [
+    "MEMORY_EFFICIENT_DECODE",
     "VAE_DECODER_COMFY_KEYS_FILTER",
     "VAE_ENCODER_COMFY_KEYS_FILTER",
     "SpatialTilingConfig",

packages/ltx-core/src/ltx_core/model/video_vae/memory_efficient_decode.py

@@ -0,0 +1,662 @@
+"""Memory-efficient VAE decoder operations.
+Reduces peak VRAM usage during video decoding through in-place operations
+and workspace buffer reuse.  The main optimizations are:
+1. **Workspace buffers** -- Pre-allocated tensors with temporal padding replace
+   dynamic padding (``F.pad`` / ``concatenate``) in ``CausalConv3d``.  A
+   workspace of shape ``[B, C, T+2, H, W]`` holds the data in positions
+   ``[1:-1]`` with replicate padding at ``[0]`` and ``[-1]``.
+2. **In-place temporal-chunked Conv3d** *(non-causal only)* -- The convolution
+   output is written back into the workspace buffer, avoiding a separate
+   output allocation.  Temporal chunking with boundary save/restore ensures
+   correct reads despite in-place writes.
+3. **In-place normalization and affine transforms** -- PixelNorm, scale/shift,
+   and SiLU are applied in-place on workspace views.
+4. **Free-before-conv** -- For ``DepthToSpaceUpsample`` blocks the input
+   tensor is freed before the convolution runs so that peak VRAM never holds
+   input *and* output simultaneously.
+Both causal and non-causal modes are supported.  Non-causal mode benefits
+from all four optimizations.  Causal mode benefits from optimizations 1, 3,
+and 4; in-place conv (2) is skipped because the asymmetric causal padding
+layout prevents clean in-place overwrites.
+Usage via the ``ModuleOps`` pattern (preferred)::
+    from ltx_core.model.video_vae import MEMORY_EFFICIENT_DECODE
+    builder = decoder_builder.with_module_ops(
+        (*decoder_builder.module_ops, MEMORY_EFFICIENT_DECODE)
+    )
+Or applied directly to an existing decoder::
+    from ltx_core.model.video_vae.memory_efficient_decode import (
+        enable_memory_efficient_decode,
+    )
+    enable_memory_efficient_decode(decoder)
+"""
+
+from __future__ import annotations
+
+import math
+from typing import TYPE_CHECKING
+
+import torch
+from einops import rearrange
+from torch import nn
+from torch.nn import functional as F
+
+from ltx_core.loader.module_ops import ModuleOps
+from ltx_core.model.common.normalization import PixelNorm
+from ltx_core.model.video_vae.convolution import CausalConv3d
+from ltx_core.model.video_vae.ops import unpatchify
+from ltx_core.model.video_vae.resnet import ResnetBlock3D, UNetMidBlock3D
+from ltx_core.model.video_vae.sampling import DepthToSpaceUpsample
+
+if TYPE_CHECKING:
+    from ltx_core.model.video_vae.video_vae import VideoDecoder
+
+
+# ---------------------------------------------------------------------------
+# Low-level helpers
+# ---------------------------------------------------------------------------
+
+
+def _memory_format_of(t: torch.Tensor, prefer_channels_last_3d: bool = False) -> torch.memory_format:
+    """Pick the memory format for a workspace allocation.
+    When ``prefer_channels_last_3d`` is True and ``t`` is 5D, return
+    ``channels_last_3d`` regardless of ``t``'s current strides -- the
+    workspace's ``.copy_(t)`` will transcribe the data into the new layout.
+    This is needed because intermediate tensors inside the decoder (after
+    ``rearrange`` + slice + residual add in ``_upsample_forward_efficient``)
+    are not NHWC-contiguous, so an auto-detect helper would silently fall
+    back to NCHW for every workspace after the first upsample.
+    Otherwise fall back to inspecting ``t``: ``channels_last_3d`` if ``t``
+    already uses it, else contiguous.
+    """
+    if prefer_channels_last_3d and t.dim() == 5:
+        return torch.channels_last_3d
+    if t.dim() == 5 and t.is_contiguous(memory_format=torch.channels_last_3d):
+        return torch.channels_last_3d
+    return torch.contiguous_format
+
+
+def _find_temporal_split_size(num_frames: int) -> int:
+    """Find chunk size for in-place temporal convolution.
+    The chunk size ensures the last chunk has at least 3 frames
+    (the temporal kernel size), avoiding degenerate chunks.
+    """
+    for s in range(16, 2, -1):
+        remainder = num_frames % s
+        if remainder == 0 or remainder >= 3:
+            return s
+
+    raise ValueError(
+        f"Unable to find a valid temporal split size for num_frames={num_frames}. "
+        "Expected a split size between 3 and 16 such that the final chunk is "
+        "either exact or has at least 3 frames."
+    )
+
+
+def _pad_workspace_temporal(workspace: torch.Tensor) -> None:
+    """Apply non-causal replicate padding to temporal boundaries.
+    Sets ``workspace[:, :, 0]`` to a copy of ``workspace[:, :, 1]`` and
+    ``workspace[:, :, -1]`` to a copy of ``workspace[:, :, -2]``.
+    """
+    workspace[:, :, 0, :, :].copy_(workspace[:, :, 1, :, :])
+    workspace[:, :, -1, :, :].copy_(workspace[:, :, -2, :, :])
+
+
+# ---------------------------------------------------------------------------
+# In-place Conv3d (non-causal only)
+# ---------------------------------------------------------------------------
+
+
+def inplace_conv3d_temporal_chunked(workspace: torch.Tensor, conv: nn.Conv3d) -> None:
+    """Run a 3x3x3 Conv3d in-place on a temporally-padded workspace.
+    The workspace has shape ``[B, C, T+2, H, W]`` where positions ``[1:-1]``
+    hold the real data and positions ``[0]`` and ``[-1]`` are padding slots.
+    The convolution must have ``kernel_size=(3,3,3)``, ``stride=(1,1,1)``,
+    ``padding=(0,1,1)`` -- no temporal padding, symmetric spatial padding.
+    The output (T frames) overwrites positions ``[1:-1]``.  Temporal chunking
+    with boundary save/restore ensures each chunk reads unmodified input even
+    though earlier chunks already wrote to the same buffer.
+    Only valid for **non-causal** mode (symmetric replicate padding).
+    Args:
+        workspace: Tensor ``[B, max(C_in, C_out), T+2, H, W]``.
+            Modified in-place; after the call ``workspace[:, :C_out, 1:-1]``
+            holds the convolution result.
+        conv: ``nn.Conv3d`` with the constraints above.
+    """
+    if conv.kernel_size != (3, 3, 3):
+        raise ValueError(f"Expected kernel_size=(3,3,3), got {conv.kernel_size}")
+    if conv.stride != (1, 1, 1):
+        raise ValueError(f"Expected stride=(1,1,1), got {conv.stride}")
+    if conv.padding != (0, 1, 1):
+        raise ValueError(f"Expected padding=(0,1,1), got {conv.padding}")
+
+    _pad_workspace_temporal(workspace)
+
+    total_frames = workspace.shape[2]
+    out_channels = conv.out_channels
+    in_channels = conv.in_channels
+
+    if total_frames > 16:
+        split_size = _find_temporal_split_size(total_frames)
+        num_splits = (total_frames + split_size - 1) // split_size
+    else:
+        split_size = total_frames - 1
+        num_splits = 1
+
+    # 1-frame buffers for saving / restoring boundary frames across chunks.
+    x_buf = torch.empty(
+        workspace.shape[0],
+        workspace.shape[1],
+        1,
+        workspace.shape[3],
+        workspace.shape[4],
+        device=workspace.device,
+        dtype=workspace.dtype,
+        memory_format=_memory_format_of(workspace),
+    )
+    o_buf = torch.empty_like(x_buf)
+
+    # Helper: extract a chunk and make it contiguous.  Workspace views can
+    # inherit strides > 2^31 from the full buffer, which makes Conv3d's
+    # reflect-padding path (F.pad) crash with "input tensor must fit into
+    # 32-bit index math".  A small .clone() per chunk avoids this.
+    needs_clone = workspace.untyped_storage().nbytes() > (2**31 - 1) * workspace.element_size()
+
+    def _chunk(t_start: int, t_end: int) -> torch.Tensor:
+        s = workspace[:, :in_channels, t_start:t_end]
+        return s.clone() if needs_clone else s
+
+    # --- First chunk ---
+    if num_splits > 1:
+        # Save the boundary now so the loop below can restore it.  Skipped
+        # when there is only one chunk: the loop never runs, and the save
+        # would be a wasted full HW slice copy.
+        x_buf[:, :, 0] = workspace[:, :, split_size - 1].clone()
+    workspace[:, :out_channels, 1:split_size] = conv(_chunk(0, split_size + 1))
+
+    # --- Remaining chunks ---
+    for i in range(1, num_splits):
+        start = i * split_size
+        end = min((i + 1) * split_size, total_frames - 1)
+
+        # Save the value at start-1 (now holds previous chunk's output).
+        o_buf[:, :, 0] = workspace[:, :, start - 1].clone()
+        # Restore the original input value needed by this chunk's conv.
+        workspace[:, :, start - 1] = x_buf[:, :, 0]
+        # Save the boundary for the *next* chunk before we overwrite it.
+        x_buf[:, :, 0] = workspace[:, :, end - 1].clone()
+
+        workspace[:, :out_channels, start:end] = conv(_chunk(start - 1, end + 1))
+
+        # Put back the previous chunk's output at the boundary.
+        workspace[:, :, start - 1] = o_buf[:, :, 0]
+
+
+# ---------------------------------------------------------------------------
+# Causal conv helper (free-before-conv)
+# ---------------------------------------------------------------------------
+
+
+def _causal_pad(x: torch.Tensor, pad_size: int) -> torch.Tensor:
+    """Build a causal-padded buffer of shape ``[B, C, T+pad_size, H, W]``.
+    Copies ``x`` into ``padded[:, :, pad_size:]`` and replicates the first
+    real frame into the leading ``pad_size`` slots.  The caller still owns
+    ``x`` after this returns.
+    """
+    padded = torch.empty(
+        x.shape[0],
+        x.shape[1],
+        x.shape[2] + pad_size,
+        x.shape[3],
+        x.shape[4],
+        device=x.device,
+        dtype=x.dtype,
+        memory_format=_memory_format_of(x),
+    )
+    padded[:, :, pad_size:].copy_(x)
+    for i in range(pad_size):
+        padded[:, :, i] = padded[:, :, pad_size]
+    return padded
+
+
+def _causal_pad_free_and_conv(x: torch.Tensor, causal_conv: CausalConv3d) -> torch.Tensor:
+    """Causal-pad *x*, free it, then run the raw ``nn.Conv3d``.
+    This avoids the peak where both the original and padded tensors are
+    live simultaneously (as happens inside ``CausalConv3d.forward``).
+    Args:
+        x: Input ``[B, C_in, T, H, W]``.  **Deleted** inside this function;
+            the caller must not use it afterwards.
+    Returns:
+        Convolution output ``[B, C_out, T, H, W]``.
+    """
+    padded = _causal_pad(x, causal_conv.time_kernel_size - 1)
+    del x
+    result = causal_conv.conv(padded)
+    del padded
+    return result
+
+
+# ---------------------------------------------------------------------------
+# In-place normalization
+# ---------------------------------------------------------------------------
+
+
+def _pixel_norm_inplace(x: torch.Tensor, eps: float = 1e-8) -> None:
+    """In-place RMS (pixel) normalization along the channel dimension."""
+    rms = torch.sqrt(torch.mean(x**2, dim=1, keepdim=True) + eps)
+    x.div_(rms)
+
+
+def _norm_inplace(norm: nn.Module, x: torch.Tensor) -> None:
+    """Apply *norm* in-place, using an optimised path for ``PixelNorm``."""
+    if isinstance(norm, PixelNorm):
+        _pixel_norm_inplace(x, eps=norm.eps)
+    else:
+        # GroupNorm or other -- fall back to allocating a temporary.
+        result = norm(x)
+        x.copy_(result)
+        del result
+
+
+# ---------------------------------------------------------------------------
+# Per-block efficient forwards
+# ---------------------------------------------------------------------------
+
+
+def _resnet_block_forward_inplace(
+    resnet: ResnetBlock3D,
+    workspace: torch.Tensor,
+    causal: bool,
+    timestep: torch.Tensor | None,
+    generator: torch.Generator | None,
+) -> None:
+    """Run a ``ResnetBlock3D`` in-place on a workspace buffer.
+    The workspace has shape ``[B, C, T+2, H, W]`` with real data in
+    ``[1:-1]``.  After this call ``workspace[:, :, 1:-1]`` holds the
+    residual-branch output ``F(x)`` (without the skip connection --
+    the caller adds it back to the hidden state).
+    Only valid when ``in_channels == out_channels`` (true for all
+    ``ResnetBlock3D`` instances inside a ``UNetMidBlock3D``).
+    """
+    if resnet.in_channels != resnet.out_channels:
+        raise ValueError(
+            "In-place resnet forward requires in_channels == out_channels, "
+            f"got {resnet.in_channels} != {resnet.out_channels}"
+        )
+
+    interior = workspace[:, :, 1:-1]
+
+    # --- norm1 + [ada scaling] + SiLU + conv1 ---
+    _norm_inplace(resnet.norm1, interior)
+
+    if resnet.timestep_conditioning and timestep is not None:
+        ada = resnet.scale_shift_table[None, ..., None, None, None].to(
+            device=interior.device, dtype=interior.dtype
+        ) + timestep.reshape(
+            interior.shape[0],
+            4,
+            -1,
+            timestep.shape[-3],
+            timestep.shape[-2],
+            timestep.shape[-1],
+        )
+        shift1, scale1, shift2, scale2 = ada.unbind(dim=1)
+        interior.mul_(1 + scale1).add_(shift1)
+
+    F.silu(interior, inplace=True)
+
+    if causal:
+        result = resnet.conv1(interior, causal=True)
+        interior.copy_(result)
+        del result
+    else:
+        inplace_conv3d_temporal_chunked(workspace, resnet.conv1.conv)
+
+    if resnet.inject_noise:
+        spatial_shape = interior.shape[-2:]
+        scale = resnet.per_channel_scale1.to(device=interior.device, dtype=interior.dtype)
+        noise = torch.randn(spatial_shape, device=interior.device, dtype=interior.dtype, generator=generator)
+        interior.add_((noise * scale)[None, :, None, ...])
+
+    # --- norm2 + [ada scaling] + SiLU + conv2 ---
+    _norm_inplace(resnet.norm2, interior)
+
+    if resnet.timestep_conditioning and timestep is not None:
+        interior.mul_(1 + scale2).add_(shift2)  # type: ignore[possibly-undefined]
+
+    F.silu(interior, inplace=True)
+    # dropout is always 0.0 during inference -- skip.
+
+    if causal:
+        result = resnet.conv2(interior, causal=True)
+        interior.copy_(result)
+        del result
+    else:
+        inplace_conv3d_temporal_chunked(workspace, resnet.conv2.conv)
+
+    if resnet.inject_noise:
+        spatial_shape = interior.shape[-2:]
+        scale = resnet.per_channel_scale2.to(device=interior.device, dtype=interior.dtype)
+        noise = torch.randn(spatial_shape, device=interior.device, dtype=interior.dtype, generator=generator)
+        interior.add_((noise * scale)[None, :, None, ...])
+
+
+def _midblock_forward_efficient(
+    block: UNetMidBlock3D,
+    hidden_states: torch.Tensor,
+    causal: bool,
+    timestep: torch.Tensor | None,
+    generator: torch.Generator | None,
+    prefer_channels_last_3d: bool = False,
+) -> torch.Tensor:
+    """Memory-efficient ``UNetMidBlock3D`` forward.
+    Allocates a single workspace buffer that is reused across all
+    ``ResnetBlock3D`` iterations.  For each block the workspace is
+    populated with the current hidden state, processed in-place, and
+    the result is added back (residual connection).
+    """
+    timestep_embed = None
+    if block.timestep_conditioning:
+        if timestep is None:
+            raise ValueError("'timestep' required when timestep_conditioning=True")
+        batch_size = hidden_states.shape[0]
+        timestep_embed = block.time_embedder(
+            timestep=timestep.flatten(),
+            hidden_dtype=hidden_states.dtype,
+        )
+        timestep_embed = timestep_embed.view(batch_size, timestep_embed.shape[-1], 1, 1, 1)
+
+    workspace = torch.empty(
+        hidden_states.shape[0],
+        hidden_states.shape[1],
+        hidden_states.shape[2] + 2,
+        hidden_states.shape[3],
+        hidden_states.shape[4],
+        device=hidden_states.device,
+        dtype=hidden_states.dtype,
+        memory_format=_memory_format_of(hidden_states, prefer_channels_last_3d),
+    )
+
+    for resnet in block.res_blocks:
+        workspace[:, :, 1:-1].copy_(hidden_states)
+        _resnet_block_forward_inplace(resnet, workspace, causal, timestep_embed, generator)
+        hidden_states.add_(workspace[:, :, 1:-1])
+
+    del workspace
+    return hidden_states
+
+
+def _upsample_forward_efficient(
+    block: DepthToSpaceUpsample,
+    x: torch.Tensor,
+    causal: bool,
+    prefer_channels_last_3d: bool = False,
+) -> torch.Tensor:
+    """Memory-efficient ``DepthToSpaceUpsample`` forward.
+    For non-causal mode the input is copied into a workspace and the
+    convolution runs in-place.  For causal mode the input is manually
+    padded and freed before the convolution runs.  Both paths avoid
+    the peak where input *and* output coexist.
+    """
+    if block.residual:
+        x_in = rearrange(
+            x,
+            "b (c p1 p2 p3) d h w -> b c (d p1) (h p2) (w p3)",
+            p1=block.stride[0],
+            p2=block.stride[1],
+            p3=block.stride[2],
+        )
+        num_repeat = math.prod(block.stride) // block.out_channels_reduction_factor
+        x_in = x_in.repeat(1, num_repeat, 1, 1, 1)
+        if block.stride[0] == 2:
+            x_in = x_in[:, :, 1:, :, :]
+
+    conv = block.conv.conv  # underlying nn.Conv3d inside CausalConv3d
+    in_channels = x.shape[1]
+    out_channels = conv.out_channels
+
+    if causal:
+        x = _causal_pad_free_and_conv(x, block.conv)
+    else:
+        mem_fmt = _memory_format_of(x, prefer_channels_last_3d)
+        workspace = torch.empty(
+            x.shape[0],
+            max(in_channels, out_channels),
+            x.shape[2] + 2,
+            x.shape[3],
+            x.shape[4],
+            device=x.device,
+            dtype=x.dtype,
+            memory_format=mem_fmt,
+        )
+        workspace[:, :in_channels, 1:-1].copy_(x)
+        del x
+        inplace_conv3d_temporal_chunked(workspace, conv)
+        x = workspace[:, :out_channels, 1:-1].contiguous(memory_format=mem_fmt)
+        del workspace
+
+    x = rearrange(
+        x,
+        "b (c p1 p2 p3) d h w -> b c (d p1) (h p2) (w p3)",
+        p1=block.stride[0],
+        p2=block.stride[1],
+        p3=block.stride[2],
+    )
+    if block.stride[0] == 2:
+        x = x[:, :, 1:, :, :]
+    if block.residual:
+        x.add_(x_in)
+        del x_in
+    return x
+
+
+# ---------------------------------------------------------------------------
+# Final norm + conv_out
+# ---------------------------------------------------------------------------
+
+
+def _final_norm_and_conv_out(
+    decoder: VideoDecoder,
+    sample: torch.Tensor,
+    causal: bool,
+    scaled_timestep: torch.Tensor | None,
+    batch_size: int,
+    prefer_channels_last_3d: bool = False,
+) -> torch.Tensor:
+    """Workspace-based final norm + [ada] + SiLU + conv_out + unpatchify."""
+    conv_out_mod: CausalConv3d = decoder.conv_out  # type: ignore[assignment]
+    conv_out = conv_out_mod.conv
+    feature_channels = sample.shape[1]
+
+    mem_fmt = _memory_format_of(sample, prefer_channels_last_3d)
+    workspace = torch.empty(
+        sample.shape[0],
+        max(feature_channels, conv_out.out_channels),
+        sample.shape[2] + 2,
+        sample.shape[3],
+        sample.shape[4],
+        device=sample.device,
+        dtype=sample.dtype,
+        memory_format=mem_fmt,
+    )
+    workspace[:, :feature_channels, 1:-1].copy_(sample)
+    del sample
+
+    interior = workspace[:, :feature_channels, 1:-1]
+    _norm_inplace(decoder.conv_norm_out, interior)
+
+    if decoder.timestep_conditioning:
+        embedded_timestep = decoder.last_time_embedder(
+            timestep=scaled_timestep.flatten(),
+            hidden_dtype=interior.dtype,
+        )
+        embedded_timestep = embedded_timestep.view(batch_size, embedded_timestep.shape[-1], 1, 1, 1)
+        ada_values = decoder.last_scale_shift_table[None, ..., None, None, None].to(
+            device=interior.device, dtype=interior.dtype
+        ) + embedded_timestep.reshape(
+            batch_size,
+            2,
+            -1,
+            embedded_timestep.shape[-3],
+            embedded_timestep.shape[-2],
+            embedded_timestep.shape[-1],
+        )
+        shift, scale = ada_values.unbind(dim=1)
+        interior.mul_(1 + scale).add_(shift)
+
+    F.silu(interior, inplace=True)
+
+    if causal:
+        # Causal: build padded tensor directly from the interior view,
+        # then free the workspace before running the conv.
+        padded = _causal_pad(interior, conv_out_mod.time_kernel_size - 1)
+        del workspace, interior
+        result = conv_out(padded)
+        del padded
+    else:
+        inplace_conv3d_temporal_chunked(workspace, conv_out)
+        result = workspace[:, : conv_out.out_channels, 1:-1].contiguous(memory_format=mem_fmt)
+        del workspace, interior
+
+    return unpatchify(result, patch_size_hw=decoder.patch_size, patch_size_t=1)
+
+
+# ---------------------------------------------------------------------------
+# Top-level efficient decoder forward
+# ---------------------------------------------------------------------------
+
+
+def _memory_efficient_forward(
+    decoder: VideoDecoder,
+    sample: torch.Tensor,
+    timestep: torch.Tensor | None = None,
+    generator: torch.Generator | None = None,
+) -> torch.Tensor:
+    """Full memory-efficient ``VideoDecoder.forward`` replacement.
+    Orchestrates the entire decode through workspace-based operations:
+    ``UNetMidBlock3D`` and ``DepthToSpaceUpsample`` blocks use efficient
+    paths; standalone ``ResnetBlock3D`` blocks fall back to the standard
+    forward.  The final norm + ada + SiLU + conv_out is also workspace-based.
+    All workspaces are allocated ``channels_last_3d`` so cuDNN's NHWC 3D
+    conv kernels run end-to-end. The caller (:func:`enable_memory_efficient_decode`)
+    is responsible for converting the input sample and decoder weights to NHWC.
+    """
+    causal = decoder.causal
+    batch_size = sample.shape[0]
+    sample = sample.to(next(decoder.parameters()).dtype)
+
+    # --- Noise injection and de-normalisation (identical to standard path) ---
+    if decoder.timestep_conditioning:
+        noise = (
+            torch.randn(sample.size(), generator=generator, dtype=sample.dtype, device=sample.device)
+            * decoder.decode_noise_scale
+        )
+        sample = noise + (1.0 - decoder.decode_noise_scale) * sample
+
+    sample = decoder.per_channel_statistics.un_normalize(sample)
+
+    if timestep is None and decoder.timestep_conditioning:
+        timestep = torch.full((batch_size,), decoder.decode_timestep, device=sample.device, dtype=sample.dtype)
+
+    # --- conv_in (latent tensor is small -- standard path is fine) ---
+    sample = decoder.conv_in(sample, causal=causal)
+
+    upscale_dtype = next(iter(decoder.up_blocks.parameters())).dtype
+    sample = sample.to(upscale_dtype)
+
+    scaled_timestep = None
+    if decoder.timestep_conditioning:
+        if timestep is None:
+            raise ValueError("'timestep' required when timestep_conditioning=True")
+        scaled_timestep = timestep * decoder.timestep_scale_multiplier.to(sample)
+
+    # Workspaces are unconditionally NHWC: rearrange + slice + residual-add
+    # inside _upsample_forward_efficient produces NCHW-default output, so
+    # per-tensor inspection would silently fall back to NCHW for every
+    # workspace after the first upsample.
+
+    # --- Up blocks (dispatch to efficient path per block type) ---
+    for up_block in decoder.up_blocks:
+        if isinstance(up_block, UNetMidBlock3D):
+            sample = _midblock_forward_efficient(
+                up_block,
+                sample,
+                causal=causal,
+                timestep=scaled_timestep if decoder.timestep_conditioning else None,
+                generator=generator,
+                prefer_channels_last_3d=True,
+            )
+        elif isinstance(up_block, DepthToSpaceUpsample):
+            sample = _upsample_forward_efficient(up_block, sample, causal=causal, prefer_channels_last_3d=True)
+        elif isinstance(up_block, ResnetBlock3D):
+            sample = up_block(sample, causal=causal, generator=generator)
+        else:
+            sample = up_block(sample, causal=causal)
+
+    return _final_norm_and_conv_out(decoder, sample, causal, scaled_timestep, batch_size, prefer_channels_last_3d=True)
+
+
+# ---------------------------------------------------------------------------
+# Public API
+# ---------------------------------------------------------------------------
+
+
+def enable_memory_efficient_decode(decoder: nn.Module) -> nn.Module:
+    """Patch a ``VideoDecoder`` to use the memory-efficient forward path.
+    The mem-efficient path runs the decoder in ``channels_last_3d`` memory
+    format: weights and inputs are converted on first call so cuDNN's NHWC
+    3D conv kernels are used (~2x faster, avoids the large vol2col scratch
+    buffer of the NCHW path).
+    The original ``forward`` is saved as ``decoder._original_forward`` so
+    that it can be restored later with :func:`disable_memory_efficient_decode`.
+    """
+    # Import here to avoid circular dependency at module level.
+    from ltx_core.model.video_vae.video_vae import VideoDecoder  # noqa: PLC0415
+
+    if not isinstance(decoder, VideoDecoder):
+        raise TypeError(f"Expected VideoDecoder, got {type(decoder).__name__}")
+
+    if hasattr(decoder, "_original_forward"):
+        return decoder
+
+    original_forward = decoder.forward
+    weights_converted = False
+
+    def efficient_forward(
+        sample: torch.Tensor,
+        timestep: torch.Tensor | None = None,
+        generator: torch.Generator | None = None,
+    ) -> torch.Tensor:
+        nonlocal weights_converted
+        if sample.dim() == 5:
+            if not weights_converted:
+                # Lazy: weights are real by first-call time (meta -> loader -> here).
+                decoder.to(memory_format=torch.channels_last_3d)
+                weights_converted = True
+            sample = sample.to(memory_format=torch.channels_last_3d)
+        return _memory_efficient_forward(decoder, sample, timestep, generator)
+
+    decoder._original_forward = original_forward  # type: ignore[attr-defined]
+    decoder.forward = efficient_forward  # type: ignore[assignment]
+    return decoder
+
+
+def disable_memory_efficient_decode(decoder: nn.Module) -> nn.Module:
+    """Restore the original ``forward`` method on a patched ``VideoDecoder``."""
+    if hasattr(decoder, "_original_forward"):
+        decoder.forward = decoder._original_forward  # type: ignore[attr-defined]
+        del decoder._original_forward  # type: ignore[attr-defined]
+    return decoder
+
+
+def _is_video_decoder(model: nn.Module) -> bool:
+    """Matcher for the ``MEMORY_EFFICIENT_DECODE`` module op."""
+    from ltx_core.model.video_vae.video_vae import VideoDecoder  # noqa: PLC0415
+
+    return isinstance(model, VideoDecoder)
+
+
+MEMORY_EFFICIENT_DECODE = ModuleOps(
+    name="memory_efficient_vae_decode",
+    matcher=_is_video_decoder,
+    mutator=enable_memory_efficient_decode,
+)

packages/ltx-core/src/ltx_core/model/video_vae/tiling.py

@@ -64,6 +64,6 @@ class TilingConfig:
     @classmethod
     def default(cls) -> "TilingConfig":
         return cls(
-            spatial_config=SpatialTilingConfig(tile_size_in_pixels=512, tile_overlap_in_pixels=64),
+            spatial_config=SpatialTilingConfig(tile_size_in_pixels=768, tile_overlap_in_pixels=64),
-            temporal_config=TemporalTilingConfig(tile_size_in_frames=64, tile_overlap_in_frames=24),
+            temporal_config=TemporalTilingConfig(tile_size_in_frames=80, tile_overlap_in_frames=24),
         )

packages/ltx-core/src/ltx_core/model/video_vae/video_vae.py

@@ -259,6 +259,7 @@ class VideoEncoder(nn.Module):
         Args:
             sample: Input video (B, C, F, H, W). F should be 1 + 8*k (e.g., 1, 9, 17, 25, 33...).
                 If not, the encoder crops the last frames to the nearest valid length.
+                Should be normalized to [-1, 1] range before encoding.
         Returns:
             Normalized latent means (B, 128, F', H', W') where F' = 1+(F-1)/8, H' = H/32, W' = W/32.
             Example: (B, 3, 33, 512, 512) -> (B, 128, 5, 16, 16).
@@ -605,8 +606,8 @@ class VideoDecoder(nn.Module):
         # many video frames and pixels correspond to a single latent cell.
         self.video_downscale_factors = SpatioTemporalScaleFactors(
             time=8,
-            width=32,
             height=32,
+            width=32,
         )
 
         self.patch_size = patch_size
@@ -905,34 +906,22 @@ class VideoDecoder(nn.Module):
         latent: torch.Tensor,
         tiling_config: TilingConfig | None = None,
         generator: torch.Generator | None = None,
-        *,
-        output_dtype: torch.dtype = torch.uint8,
     ) -> Iterator[torch.Tensor]:
-        """Decode a video latent tensor, yielding chunks ``[f, h, w, c]``.
+        """Decode a video latent tensor, yielding float chunks ``[f, h, w, c]`` in ``[0, 1]``.
         Subclasses (e.g. ``DistributedVideoDecoder``) may override this to
         control eagerness or distribution across ranks.
-        Args:
-            output_dtype: Target dtype for output tensors.  ``torch.uint8``
-                (default) maps the decoder's ``[-1, 1]`` output to
-                ``[0, 255]``.  Any floating dtype returns ``[0, 1]`` cast
-                to that dtype.
         """
 
-        def _convert(frames: torch.Tensor) -> torch.Tensor:
+        def to_rgb(frames: torch.Tensor) -> torch.Tensor:
-            # rearrange materializes a new contiguous tensor for this permutation,
-            # so in-place ops below do not mutate the caller's data.
             video = rearrange(frames[0], "c f h w -> f h w c")
-            video.add_(1.0).mul_(0.5).clamp_(0.0, 1.0)
+            return video.add_(1.0).mul_(0.5).clamp_(0.0, 1.0)
-            if output_dtype == torch.uint8:
-                return video.mul_(255.0).to(torch.uint8)
-            return video.to(output_dtype)
 
         if tiling_config is not None:
             for frames in self.tiled_decode(latent, tiling_config, generator=generator):
-                yield _convert(frames)
+                yield to_rgb(frames)
         else:
             decoded = self(latent, generator=generator)
-            yield _convert(decoded)
+            yield to_rgb(decoded)
 
     def _group_tiles_by_temporal_slice(self, tiles: List[Tile]) -> List[List[Tile]]:
         """Group tiles by their temporal output slice."""

packages/ltx-core/src/ltx_core/multigpu/__init__.py

@@ -0,0 +1,11 @@
+"""
+Multi-GPU utilities for LTX models.
+This package provides utilities for running LTX models across multiple GPUs
+using tiled data-parallel techniques and sharded state-dict utilities.
+"""
+
+from ltx_core.multigpu import transformer, vae
+from ltx_core.multigpu.sharded_sd import ShardedSD
+from ltx_core.tiling import DimensionTilingConfig, TileCountConfig
+
+__all__ = ["DimensionTilingConfig", "ShardedSD", "TileCountConfig", "transformer", "vae"]

packages/ltx-core/src/ltx_core/multigpu/gemma/__init__.py

@@ -0,0 +1,6 @@
+"""Multi-GPU utilities for the Gemma text encoder."""
+
+from ltx_core.multigpu.gemma.accelerate_wrapper import AccelerateGemmaWrapper
+from ltx_core.multigpu.gemma.loader import load_gemma_with_device_map
+
+__all__ = ["AccelerateGemmaWrapper", "load_gemma_with_device_map"]

packages/ltx-core/src/ltx_core/multigpu/gemma/accelerate_wrapper.py

@@ -0,0 +1,30 @@
+"""Accelerate-based Gemma text encoder wrapper for multi-GPU inference.
+One rank (``src_rank``) holds the real ``GemmaTextEncoder`` loaded with
+``device_map="auto"``; other ranks hold a lightweight stub. Every public
+method runs on the source rank and broadcasts results to all ranks via the
+provided NCCL process group.
+The ``broadcast_group`` should cover **all ranks that need the
+embeddings** (typically the transformer group or world group).
+"""
+
+from __future__ import annotations
+
+import torch
+
+from ltx_core.multigpu.gemma.broadcast_wrapper import BroadcastGemmaWrapper
+
+
+class AccelerateGemmaWrapper(BroadcastGemmaWrapper):
+    """Source-rank encode + NCCL broadcast around a sharded ``GemmaTextEncoder``."""
+
+    def encode(
+        self,
+        prompts: list[str],
+        padding_side: str = "left",
+    ) -> list[tuple[tuple[torch.Tensor, ...], torch.Tensor]]:
+        """Fuse all prompts into one Gemma call on the source rank, broadcast each output."""
+        if self._rank == self._src_rank:
+            local_outputs = self._encoder.encode(prompts, padding_side)
+        else:
+            local_outputs = [(None, None)] * len(prompts)
+        return [self._broadcast_encoder_output(hs, mask, self._src_rank) for hs, mask in local_outputs]

packages/ltx-core/src/ltx_core/multigpu/gemma/batch_parallel_wrapper.py

@@ -0,0 +1,75 @@
+"""Batch-parallel Gemma text encoder wrapper for multi-GPU inference.
+Each rank holds a full :class:`GemmaTextEncoder` replica resident on its
+own GPU.  ``encode`` partitions the prompt list across ranks (each rank
+encodes a disjoint slice) and broadcasts every prompt's outputs from its
+encoding rank to all other ranks, so all ranks end up with the full list
+in the original order.
+``enhance_t2v`` / ``enhance_i2v`` (inherited) involve sampling, so they
+execute on ``src_rank`` only and the generated string is broadcast.
+"""
+
+from __future__ import annotations
+
+import torch
+import torch.distributed as dist
+
+from ltx_core.multigpu.gemma.broadcast_wrapper import BroadcastGemmaWrapper
+from ltx_core.text_encoders.gemma.encoders.base_encoder import GemmaTextEncoder
+
+
+def _partition(total: int, world_size: int) -> list[int]:
+    """Spread ``total`` items across ``world_size`` ranks; remainder lands on the first ranks."""
+    base, rem = divmod(total, world_size)
+    return [base + (1 if i < rem else 0) for i in range(world_size)]
+
+
+class BatchParallelGemmaWrapper(BroadcastGemmaWrapper):
+    """Per-rank Gemma replica; ``encode`` parallelises a batch across ranks."""
+
+    _encoder: GemmaTextEncoder  # always resident on every rank, unlike the base's optional encoder
+
+    def __init__(
+        self,
+        encoder: GemmaTextEncoder,
+        broadcast_group: dist.ProcessGroup | None,
+        src_rank: int,
+        dtype: torch.dtype = torch.bfloat16,
+        device: torch.device | None = None,
+    ) -> None:
+        """Wrap a per-rank Gemma replica for batch-parallel encoding.
+        Args:
+            encoder: Full Gemma replica; required and resident on every rank (unlike
+                the base, where it is optional and real only on ``src_rank``).
+            broadcast_group: NCCL group spanning the ranks that share the encode work.
+            src_rank: Rank within ``broadcast_group`` that runs the inherited sampling
+                methods (``enhance_t2v`` / ``enhance_i2v``); ``encode`` uses every rank.
+            dtype: Target dtype for output tensors.
+            device: Target device for output tensors; defaults to the current CUDA device.
+        """
+        super().__init__(encoder, broadcast_group, src_rank, dtype, device)
+        self._world_size = dist.get_world_size(broadcast_group)
+
+    def encode(
+        self,
+        prompts: list[str],
+        padding_side: str = "left",
+    ) -> list[tuple[tuple[torch.Tensor, ...], torch.Tensor]]:
+        """Partition prompts across ranks, encode in parallel, broadcast per-prompt outputs.
+        With B prompts on W ranks, each rank gets ``ceil(B/W)`` or ``floor(B/W)``
+        prompts; the typical pos+neg case (B=2, W=2) gives one prompt per rank,
+        running both Gemma forwards concurrently on different GPUs.
+        """
+        n = len(prompts)
+        if n == 0:
+            return []
+        counts = _partition(n, self._world_size)
+        start = sum(counts[: self._rank])
+        local_prompts = prompts[start : start + counts[self._rank]]
+        local_outputs = self._encoder.encode(local_prompts, padding_side) if local_prompts else []
+
+        all_outputs: list[tuple[tuple[torch.Tensor, ...], torch.Tensor]] = []
+        for owner_rank, owner_count in enumerate(counts):
+            for slot in range(owner_count):
+                hs, mask = local_outputs[slot] if owner_rank == self._rank else (None, None)
+                all_outputs.append(self._broadcast_encoder_output(hs, mask, owner_rank))
+        return all_outputs

packages/ltx-core/src/ltx_core/multigpu/gemma/broadcast_wrapper.py

@@ -0,0 +1,108 @@
+"""Shared base for the multi-GPU Gemma wrappers: src-rank gating + NCCL broadcast."""
+
+from __future__ import annotations
+
+import torch
+import torch.distributed as dist
+
+from ltx_core.text_encoders.gemma.encoders.base_encoder import GemmaTextEncoder
+
+
+class BroadcastGemmaWrapper(torch.nn.Module):
+    """Encoder/group plumbing, prompt enhancement, and result broadcast.
+    Subclasses implement ``encode`` (the stub below raises).
+    Args:
+        encoder: The encoder; real on ``src_rank``, may be ``None`` elsewhere.
+        broadcast_group: NCCL group covering ranks that need the embeddings.
+        src_rank: Rank *within* ``broadcast_group`` that holds the real encoder and runs
+            the sampling-based ``enhance_*`` methods; its results are broadcast to every
+            other rank in the group. Builders derive it from a global driver rank via
+            ``dist.get_group_rank``.
+        dtype: Target dtype for output tensors.
+        device: Target device for output tensors.
+    """
+
+    def __init__(
+        self,
+        encoder: GemmaTextEncoder | None,
+        broadcast_group: dist.ProcessGroup | None,
+        src_rank: int,
+        dtype: torch.dtype = torch.bfloat16,
+        device: torch.device | None = None,
+    ) -> None:
+        super().__init__()
+        if device is None and torch.cuda.is_available():
+            device = torch.device("cuda", torch.cuda.current_device())
+        self._encoder = encoder
+        self._group = broadcast_group
+        self._src_rank = src_rank
+        self._rank = dist.get_rank(broadcast_group)
+        self._dtype = dtype
+        self._device = device
+
+    def encode(
+        self,
+        prompts: list[str],
+        padding_side: str = "left",
+    ) -> list[tuple[tuple[torch.Tensor, ...], torch.Tensor]]:
+        """Encode a batch of prompts to per-prompt hidden states; implemented by subclasses."""
+        raise NotImplementedError
+
+    def enhance_t2v(
+        self,
+        prompt: str,
+        max_new_tokens: int = 512,
+        system_prompt: str | None = None,
+        seed: int = 10,
+    ) -> str:
+        result = None
+        if self._rank == self._src_rank:
+            result = self._encoder.enhance_t2v(prompt, max_new_tokens, system_prompt, seed)
+        return self._broadcast_str(result)
+
+    def enhance_i2v(
+        self,
+        prompt: str,
+        image: torch.Tensor,
+        max_new_tokens: int = 512,
+        system_prompt: str | None = None,
+        seed: int = 10,
+    ) -> str:
+        result = None
+        if self._rank == self._src_rank:
+            result = self._encoder.enhance_i2v(prompt, image, max_new_tokens, system_prompt, seed)
+        return self._broadcast_str(result)
+
+    def _broadcast_str(self, value: str | None) -> str:
+        obj_list: list[str | None] = [value]
+        dist.broadcast_object_list(obj_list, src=self._src_rank, group=self._group)
+        result = obj_list[0]
+        assert result is not None, "broadcast returned None; check src_rank/broadcast_group"
+        return result
+
+    def _broadcast_encoder_output(
+        self,
+        hidden_states: tuple[torch.Tensor, ...] | None,
+        attention_mask: torch.Tensor | None,
+        src_rank: int,
+    ) -> tuple[tuple[torch.Tensor, ...], torch.Tensor]:
+        """Broadcast hidden states + attention mask via NCCL from ``src_rank``."""
+        if self._rank == src_rank:
+            meta = [{"hs_shapes": [h.shape for h in hidden_states], "mask_shape": attention_mask.shape}]
+        else:
+            meta = [None]
+        dist.broadcast_object_list(meta, src=src_rank, group=self._group)
+        info = meta[0]
+
+        if self._rank != src_rank:
+            hidden_states = tuple(torch.empty(s, device=self._device, dtype=self._dtype) for s in info["hs_shapes"])
+            attention_mask = torch.empty(info["mask_shape"], device=self._device, dtype=torch.long)
+        else:
+            hidden_states = tuple(h.to(device=self._device, dtype=self._dtype) for h in hidden_states)
+            attention_mask = attention_mask.to(device=self._device)
+
+        for h in hidden_states:
+            dist.broadcast(h, src=src_rank, group=self._group)
+        dist.broadcast(attention_mask, src=src_rank, group=self._group)
+
+        return hidden_states, attention_mask

packages/ltx-core/src/ltx_core/multigpu/gemma/loader.py

@@ -0,0 +1,55 @@
+"""Load GemmaTextEncoder with Accelerate ``device_map="auto"``.
+The Gemma LLM backbone is spread across available CUDA devices using
+HuggingFace Accelerate's automatic device placement.
+Mirrors the ``PromptEncoder`` text-encoder loading in
+``ltx_pipelines.utils.blocks`` but uses ``device_map="auto"`` instead of
+placing the entire model on a single GPU.
+"""
+
+from __future__ import annotations
+
+import logging
+
+import torch
+from transformers import AutoImageProcessor, Gemma3ForConditionalGeneration, Gemma3Processor
+
+from ltx_core.text_encoders.gemma.encoders.base_encoder import GemmaTextEncoder
+from ltx_core.text_encoders.gemma.tokenizer import LTXVGemmaTokenizer
+from ltx_core.utils import find_matching_file
+
+logger = logging.getLogger(__name__)
+
+
+def load_gemma_with_device_map(
+    gemma_root_path: str,
+    dtype: torch.dtype = torch.bfloat16,
+) -> GemmaTextEncoder:
+    """Load GemmaTextEncoder with the LLM backbone spread across GPUs.
+    Uses ``Gemma3ForConditionalGeneration.from_pretrained(device_map="auto")``
+    to distribute layers across available CUDA devices.
+    Args:
+        gemma_root_path: Path to Gemma model directory.
+        dtype: Data type for model weights.
+    """
+    model_folder = str(find_matching_file(gemma_root_path, "model*.safetensors").parent)
+    tokenizer_path = str(find_matching_file(gemma_root_path, "tokenizer.model").parent)
+    processor_path = str(find_matching_file(gemma_root_path, "preprocessor_config.json").parent)
+
+    logger.info("Loading Gemma LLM with device_map='auto'...")
+    gemma_model = Gemma3ForConditionalGeneration.from_pretrained(
+        model_folder,
+        dtype=dtype,
+        device_map="auto",
+        local_files_only=True,
+    )
+
+    tokenizer = LTXVGemmaTokenizer(tokenizer_path, 1024)
+    image_processor = AutoImageProcessor.from_pretrained(processor_path, local_files_only=True, use_fast=False)
+    processor = Gemma3Processor(image_processor=image_processor, tokenizer=tokenizer.tokenizer)
+
+    return GemmaTextEncoder(
+        model=gemma_model,
+        tokenizer=tokenizer,
+        processor=processor,
+        dtype=dtype,
+    )

packages/ltx-core/src/ltx_core/multigpu/sharded_sd.py

@@ -0,0 +1,191 @@
+"""Sharded state dict with distributed weight backup.
+Each rank stores ~1/N of the model weights. Bucketed broadcasts
+restore weights into a target state dict using only a small,
+caller-provided staging buffer.
+"""
+
+from __future__ import annotations
+
+import hashlib
+from dataclasses import dataclass
+
+import torch
+import torch.distributed as dist
+
+
+def _stable_owner(key: str, world: int) -> int:
+    """Deterministic rank assignment (same across all processes)."""
+    h = hashlib.md5(key.encode("utf-8")).digest()
+    return int.from_bytes(h[:8], "little") % world
+
+
+def _nbytes(t: torch.Tensor) -> int:
+    return t.numel() * t.element_size()
+
+
+@dataclass
+class ShardedSD:
+    """Sharded state dict with distributed weight backup.
+    Distributes model weights across ranks for memory-efficient backup
+    and restoration.  Can be used for any scenario where a full state dict
+    needs to be restored from sharded storage (e.g. LoRA hot-swap, weight
+    rollback, checkpoint recovery).
+    - Deterministic ownership: ``MD5(key) % world_size``
+    - Local storage only for owned keys (VRAM ≈ 1/world_size of model)
+    - Bucketed broadcast using a single small staging buffer
+    Usage::
+        backup = ShardedSD.from_state_dict(model.state_dict(), group)
+        staging = torch.empty(64 * 1024 * 1024, dtype=torch.uint8, device=device)
+        backup.broadcast_shards_into(target_sd, staging)    # cooperative: all ranks must call
+    """
+
+    keys: tuple[str, ...]
+    """All parameter keys in the original state dict, in insertion order."""
+    key_sizes: dict[str, int]
+    """Byte size of each parameter tensor (numel * element_size)."""
+    owner_of: dict[str, int]
+    """Maps each key to the rank that stores it."""
+    local_shard: dict[str, torch.Tensor]
+    """Tensors owned by this rank (subset of the full state dict)."""
+    rank: int
+    """This process's rank within the group."""
+    world: int
+    """Total number of ranks in the group."""
+    group: dist.ProcessGroup
+    """NCCL process group used for broadcast operations."""
+    _owner_groups: dict[int, list[str]]
+    """Keys grouped by owning rank, sorted by descending tensor size."""
+
+    @classmethod
+    def from_state_dict(
+        cls,
+        sd: dict[str, torch.Tensor],
+        group: dist.ProcessGroup,
+        clone: bool = True,
+    ) -> ShardedSD:
+        rank = dist.get_rank(group)
+        world = dist.get_world_size(group)
+
+        keys = tuple(sd.keys())
+        # Co-locate .weight_scale with its .weight on the same rank.
+        owner_of: dict[str, int] = {}
+        for k in keys:
+            if k.endswith(".weight_scale"):
+                parent = k.replace(".weight_scale", ".weight")
+                if parent in sd:
+                    owner_of[k] = _stable_owner(parent, world)
+                    continue
+            owner_of[k] = _stable_owner(k, world)
+
+        key_sizes = {k: _nbytes(v) for k, v in sd.items()}
+
+        local_shard: dict[str, torch.Tensor] = {}
+        for k, v in sd.items():
+            if owner_of[k] == rank:
+                local_shard[k] = v.clone() if clone else v
+
+        owner_groups: dict[int, list[str]] = {r: [] for r in range(world)}
+        for k in keys:
+            owner_groups[owner_of[k]].append(k)
+        for r in range(world):
+            owner_groups[r].sort(key=lambda kk: key_sizes[kk], reverse=True)
+
+        return cls(
+            keys=keys,
+            key_sizes=key_sizes,
+            owner_of=owner_of,
+            local_shard=local_shard,
+            rank=rank,
+            world=world,
+            group=group,
+            _owner_groups=owner_groups,
+        )
+
+    def broadcast_shards_into(
+        self,
+        target_sd: dict[str, torch.Tensor],
+        staging: torch.Tensor,
+    ) -> None:
+        """Broadcast stored weights from sharded backup into *target_sd*.
+        This is a **cooperative operation** — all ranks in the process group
+        must call it simultaneously.
+        *staging* is a caller-owned ``uint8`` scratch buffer; its size sets the
+        broadcast granularity (tensors larger than it split across rounds). It
+        may be shared by instances that never broadcast at the same time. Writes
+        directly into existing tensors in *target_sd*.
+        """
+        if staging.dtype != torch.uint8 or staging.numel() == 0:
+            raise ValueError("staging must be a non-empty uint8 buffer")
+        for owner, klist in self._owner_groups.items():
+            if klist:
+                self._broadcast_group(owner, klist, target_sd, staging)
+
+    def _broadcast_group(
+        self,
+        owner: int,
+        keys: list[str],
+        target_sd: dict[str, torch.Tensor],
+        staging: torch.Tensor,
+    ) -> None:
+        """Pack & broadcast params from *owner*, splitting tensors across rounds."""
+        rounds = self._plan_rounds(keys, staging.numel())
+
+        for round_chunks in rounds:
+            filled = 0
+            if self.rank == owner:
+                for k, offset, chunk_size in round_chunks:
+                    src = self.local_shard[k]
+                    if not src.is_contiguous():
+                        raise RuntimeError(f"ShardedSD: local shard tensor '{k}' is not contiguous")
+                    src_bytes = src.view(torch.uint8).view(-1)
+                    staging[filled : filled + chunk_size].copy_(
+                        src_bytes[offset : offset + chunk_size], non_blocking=True
+                    )
+                    filled += chunk_size
+            else:
+                filled = sum(chunk_size for (_, _, chunk_size) in round_chunks)
+
+            if filled == 0:
+                continue
+            view = staging[:filled]
+            dist.broadcast(view, src=owner, group=self.group)
+
+            cursor = 0
+            for k, offset, chunk_size in round_chunks:
+                dst = target_sd[k]
+                if not dst.is_contiguous():
+                    raise RuntimeError(f"ShardedSD: target tensor '{k}' is not contiguous")
+                dst_bytes = dst.view(torch.uint8).view(-1)
+                dst_bytes[offset : offset + chunk_size].copy_(staging[cursor : cursor + chunk_size], non_blocking=True)
+                cursor += chunk_size
+
+    def _plan_rounds(self, keys: list[str], capacity: int) -> list[list[tuple[str, int, int]]]:
+        """Build rounds that pack a *capacity*-byte buffer, splitting tensors if needed.
+        Returns a list of rounds, each containing ``(key, byte_offset, chunk_bytes)`` tuples.
+        """
+        rounds: list[list[tuple[str, int, int]]] = []
+        current: list[tuple[str, int, int]] = []
+        used = 0
+
+        for k in keys:
+            remaining = self.key_sizes[k]
+            offset = 0
+
+            while remaining > 0:
+                space = capacity - used
+                if space == 0:
+                    rounds.append(current)
+                    current = []
+                    used = 0
+                    space = capacity
+
+                chunk = min(remaining, space)
+                current.append((k, offset, chunk))
+                used += chunk
+                offset += chunk
+                remaining -= chunk
+
+        if current:
+            rounds.append(current)
+
+        return rounds

packages/ltx-core/src/ltx_core/multigpu/transformer/__init__.py

@@ -0,0 +1,13 @@
+"""
+Multi-GPU transformer utilities for LTX models.
+This module provides utilities for running LTX transformer models across multiple GPUs
+using tiled data parallelism.
+"""
+
+from ltx_core.multigpu.transformer.tiled_data_parallel import (
+    TiledDataParallelModelWrapper,
+)
+
+__all__ = [
+    "TiledDataParallelModelWrapper",
+]

packages/ltx-core/src/ltx_core/multigpu/transformer/attention.py

@@ -0,0 +1,270 @@
+import torch
+import torch.distributed as dist
+
+from ltx_core.model.transformer.attention import AttentionCallable, MaskedAttentionCallable
+
+# Mirrors the kernel's DEFAULT_BARRIER_TIMEOUT_SECONDS (configs.cuh), which All2All converts to
+# cycles via the device peak SM clock. Stored so the timeout can be read back to reset after a raise.
+_DEFAULT_ALL2ALL_TIMEOUT_SECONDS = 10.0
+
+
+class AttentionManager:
+    def __init__(
+        self,
+        max_tokens: int,
+        num_heads: int,
+        head_dim: int,
+        tensor_dtype: torch.dtype,
+        group: torch.distributed.ProcessGroup,
+        copy_out_: bool = False,
+    ) -> None:
+        # Lazy: ltx_kernels is an optional GPU-only dep, and this constructor already
+        # requires CUDA -- so importing it here (not at module scope) keeps the multigpu
+        # modules importable without the kernels installed (e.g. CPU CI test collection).
+        from ltx_kernels import All2All  # noqa: PLC0415
+
+        self.rank = dist.get_rank(group)
+        self.world_size = dist.get_world_size(group)
+        self.max_tokens = max_tokens
+        hidden_dim = num_heads * head_dim
+        num_sms = torch.cuda.get_device_properties(self.rank).multi_processor_count
+        self.copy_out = copy_out_
+        buffer_seqlen = (max_tokens + self.world_size - 1) // self.world_size
+        self.all2all_heads, self.all2all_q = (
+            All2All(
+                rank=self.rank,
+                world_size=self.world_size,
+                seqlen=buffer_seqlen,
+                hidden_dim=hidden_dim,
+                num_sms=num_sms,
+                tensor_dtype=tensor_dtype,
+                group=group,
+            )
+            for _ in range(2)
+        )
+        self.all2all_k, self.all2all_v = (
+            All2All(
+                rank=self.rank,
+                world_size=self.world_size,
+                seqlen=buffer_seqlen,
+                hidden_dim=hidden_dim,
+                num_sms=num_sms,
+                tensor_dtype=tensor_dtype,
+                group=group,
+            )
+            if not self.copy_out
+            else self.all2all_q
+            for _ in range(2)
+        )
+        self.group = group
+        self._all2all_timeout_seconds = _DEFAULT_ALL2ALL_TIMEOUT_SECONDS
+
+    def set_seqlen_all2all(self, seqlens: list[int]) -> None:
+        # Route through the wrappers so the registered custom ops' fake-impl
+        # shape info gets updated alongside the C++ runtime's rank_tokens.
+        self.all2all_q.set_rank_tokens(seqlens)
+        self.all2all_k.set_rank_tokens(seqlens)
+        self.all2all_v.set_rank_tokens(seqlens)
+        self.all2all_heads.set_rank_tokens(seqlens)
+
+    @property
+    def all2all_timeout_seconds(self) -> float:
+        """The all2all barrier (deadlock-detection) timeout, in seconds, applied to every instance."""
+        return self._all2all_timeout_seconds
+
+    @all2all_timeout_seconds.setter
+    def all2all_timeout_seconds(self, seconds: float) -> None:
+        # Raise it for the first ``torch.compile`` forward -- where one rank's recompile can delay its
+        # all2all kernel launch past the steady-state timeout, tripping the barrier -- then reset to
+        # the prior value. ``all2all_k``/``all2all_v`` may alias ``all2all_q`` (copy-out path);
+        # setting twice is idempotent. Fan out first (it validates) so a rejected value leaves the
+        # stored steady-state value untouched.
+        for a2a in (self.all2all_q, self.all2all_k, self.all2all_v, self.all2all_heads):
+            a2a.set_timeout_seconds(seconds)
+        self._all2all_timeout_seconds = seconds
+
+    def send_recv_qkv(
+        self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor
+    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        t_q = self.all2all_q.send_recv_heads(q, copy_out=self.copy_out)
+        t_k = self.all2all_k.send_recv_heads(k, copy_out=self.copy_out)
+        t_v = self.all2all_v.send_recv_heads(v, copy_out=self.copy_out)
+        return t_q, t_k, t_v
+
+    def gather_heads(self, heads_local: torch.Tensor) -> torch.Tensor:
+        out = self.all2all_heads.gather_heads(heads_local, copy_out=self.copy_out)
+        return out
+
+
+class _All2AllRedistribute:
+    """Shared redistribute/gather pipeline for self-attention SP wrappers.
+    Folds the head dim view-and-shuffle so the masked and unmasked variants only
+    have to choose how to invoke the inner attention (with or without the mask
+    kwarg) -- the rest of the SP plumbing is identical.
+    """
+
+    def __init__(self, manager: AttentionManager) -> None:
+        self.manager = manager
+
+    def redistribute(
+        self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, heads: int
+    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, int, int]:
+        if heads % self.manager.world_size != 0:
+            raise ValueError(f"heads ({heads}) must be divisible by world_size ({self.manager.world_size})")
+
+        head_dim = q.shape[-1] // heads
+        q = q.view(q.shape[0], q.shape[1], heads, head_dim)
+        k = k.view(k.shape[0], k.shape[1], heads, head_dim)
+        v = v.view(v.shape[0], v.shape[1], heads, head_dim)
+
+        t_q, t_k, t_v = self.manager.send_recv_qkv(q, k, v)
+        local_heads = heads // self.manager.world_size
+
+        # `flatten` / `unflatten` collapse only the head dims, avoiding a `-1` in the
+        # seq position -- that would otherwise be ambiguous if the seq is 0 for a
+        # zero-token modality.
+        t_q = t_q.flatten(-2)
+        t_k = t_k.flatten(-2)
+        t_v = t_v.flatten(-2)
+        return t_q, t_k, t_v, local_heads, head_dim
+
+    def gather(self, hidden_states: torch.Tensor, local_heads: int, head_dim: int) -> torch.Tensor:
+        hidden_states = hidden_states.unflatten(-1, (local_heads, head_dim))
+        hidden_states = self.manager.gather_heads(hidden_states)
+        return hidden_states.flatten(-2)
+
+
+class All2AllAttention(AttentionCallable):
+    def __init__(self, manager: AttentionManager, original_attention: AttentionCallable):
+        self._sp = _All2AllRedistribute(manager)
+        self.original_attention = original_attention
+
+    def __call__(
+        self,
+        q: torch.Tensor,
+        k: torch.Tensor,
+        v: torch.Tensor,
+        heads: int,
+    ) -> torch.Tensor:
+        t_q, t_k, t_v, local_heads, head_dim = self._sp.redistribute(q, k, v, heads)
+        hidden_states = self.original_attention(q=t_q, k=t_k, v=t_v, heads=local_heads)
+        return self._sp.gather(hidden_states, local_heads, head_dim)
+
+
+class MaskedAll2AllAttention(MaskedAttentionCallable):
+    def __init__(self, manager: AttentionManager, original_attention: MaskedAttentionCallable):
+        self._sp = _All2AllRedistribute(manager)
+        self.original_attention = original_attention
+
+    def __call__(
+        self,
+        q: torch.Tensor,
+        k: torch.Tensor,
+        v: torch.Tensor,
+        heads: int,
+        mask: torch.Tensor,
+    ) -> torch.Tensor:
+        t_q, t_k, t_v, local_heads, head_dim = self._sp.redistribute(q, k, v, heads)
+        hidden_states = self.original_attention(q=t_q, k=t_k, v=t_v, heads=local_heads, mask=mask)
+        return self._sp.gather(hidden_states, local_heads, head_dim)
+
+
+class _AudioAll2AllRedistribute:
+    """Shared redistribute/gather pipeline for audio cross-attention SP wrappers.
+    Q is sliced locally per rank (no cross-rank shuffle on Q because the audio
+    sequence length is small enough to replicate); K/V are redistributed across
+    ranks via ``send_recv_heads``; outputs are gathered via
+    ``all_gather_into_tensor`` along the head dimension. The masked and unmasked
+    variants share this plumbing and only differ in how they invoke the inner
+    attention.
+    """
+
+    def __init__(self, manager: AttentionManager) -> None:
+        self.manager = manager
+
+    def redistribute(
+        self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, heads: int
+    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, int, int]:
+        if heads % self.manager.world_size != 0:
+            raise ValueError(f"heads ({heads}) must be divisible by world_size ({self.manager.world_size})")
+
+        head_dim = q.shape[-1] // heads
+        heads_per_rank = heads // self.manager.world_size
+        rank = self.manager.rank
+
+        q = q.view(q.shape[0], q.shape[1], heads, head_dim)
+        k = k.view(k.shape[0], k.shape[1], heads, head_dim)
+        v = v.view(v.shape[0], v.shape[1], heads, head_dim)
+
+        t_q = q[:, :, heads_per_rank * rank : heads_per_rank * (rank + 1), :].clone()
+        t_k = self.manager.all2all_k.send_recv_heads(k, copy_out=self.manager.copy_out)
+        t_v = self.manager.all2all_v.send_recv_heads(v, copy_out=self.manager.copy_out)
+
+        # `flatten` / `unflatten` collapse only the head dims, avoiding a `-1` in the
+        # seq position -- that would otherwise be ambiguous if the seq is 0 for a
+        # zero-token modality.
+        t_q = t_q.flatten(-2)
+        t_k = t_k.flatten(-2)
+        t_v = t_v.flatten(-2)
+        return t_q, t_k, t_v, heads_per_rank, head_dim
+
+    def gather(self, hidden_states: torch.Tensor, heads_per_rank: int, head_dim: int) -> torch.Tensor:
+        # (B, S, heads_per_rank, head_dim). Move head dim to dim 0 so all_gather_into_tensor
+        # gathers along it; permute back after the collective.
+        hidden_states = hidden_states.unflatten(-1, (heads_per_rank, head_dim)).permute(2, 0, 1, 3).contiguous()
+        gathered = torch.empty(
+            (heads_per_rank * self.manager.world_size, *hidden_states.shape[1:]),
+            dtype=hidden_states.dtype,
+            device=hidden_states.device,
+        )
+        dist.all_gather_into_tensor(gathered, hidden_states, group=self.manager.group)
+        # (heads, B, S, head_dim) -> (B, S, heads, head_dim) -> (B, S, heads * head_dim)
+        return gathered.permute(1, 2, 0, 3).flatten(-2)
+
+
+class AudioAll2AllAttention(AttentionCallable):
+    """All2All attention for audio cross-attention (video_to_audio).
+    Q is sliced locally per rank, K/V are redistributed via send_recv_heads,
+    then outputs are gathered via all_gather across the head dimension.
+    """
+
+    def __init__(self, manager: AttentionManager, original_attention: AttentionCallable):
+        self._sp = _AudioAll2AllRedistribute(manager)
+        self.original_attention = original_attention
+
+    def __call__(
+        self,
+        q: torch.Tensor,
+        k: torch.Tensor,
+        v: torch.Tensor,
+        heads: int,
+    ) -> torch.Tensor:
+        t_q, t_k, t_v, heads_per_rank, head_dim = self._sp.redistribute(q, k, v, heads)
+        hidden_states = self.original_attention(q=t_q, k=t_k, v=t_v, heads=heads_per_rank)
+        return self._sp.gather(hidden_states, heads_per_rank, head_dim)
+
+
+class MaskedAudioAll2AllAttention(MaskedAttentionCallable):
+    """Masked counterpart to :class:`AudioAll2AllAttention`.
+    No current caller invokes A2V / V2A cross-attention with a mask, so the SP
+    mutator pre-installs an unmasked-only :class:`AudioAll2AllAttention` and the
+    masked slot stays at the model default. Defined now so adding masked audio
+    cross-attention later is just an SP-mutator change, not a missing-piece
+    discovery.
+    """
+
+    def __init__(self, manager: AttentionManager, original_attention: MaskedAttentionCallable):
+        self._sp = _AudioAll2AllRedistribute(manager)
+        self.original_attention = original_attention
+
+    def __call__(
+        self,
+        q: torch.Tensor,
+        k: torch.Tensor,
+        v: torch.Tensor,
+        heads: int,
+        mask: torch.Tensor,
+    ) -> torch.Tensor:
+        t_q, t_k, t_v, heads_per_rank, head_dim = self._sp.redistribute(q, k, v, heads)
+        hidden_states = self.original_attention(q=t_q, k=t_k, v=t_v, heads=heads_per_rank, mask=mask)
+        return self._sp.gather(hidden_states, heads_per_rank, head_dim)

packages/ltx-core/src/ltx_core/multigpu/transformer/sequence_parallel.py

@@ -0,0 +1,300 @@
+"""
+Multi-GPU inference wrapper for LTX transformer models.
+This module provides utilities for running LTX model inference across multiple GPUs
+using sequence parallelism. It:
+- Tiles the video inputs across GPUs in the sequence (token) dimension
+- Patches video self-attention operations with all2all attention
+- Runs the model forward pass on each GPU with its local tile
+- Gathers all tokens back to all GPUs after the forward pass
+"""
+
+from dataclasses import replace
+from itertools import accumulate
+
+import torch
+
+from ltx_core.guidance.perturbations import BatchedPerturbationConfig
+from ltx_core.loader.module_ops import ModuleOps
+from ltx_core.model.transformer.attention import Attention
+from ltx_core.model.transformer.modality import Modality
+from ltx_core.model.transformer.model import LTXModel
+from ltx_core.model.transformer.transformer import BasicAVTransformerBlock
+from ltx_core.multigpu.transformer.attention import (
+    All2AllAttention,
+    AttentionManager,
+    AudioAll2AllAttention,
+    MaskedAll2AllAttention,
+    MaskedAudioAll2AllAttention,
+)
+
+
+def compute_sequence_partition(
+    total_tokens: int,
+    world_size: int,
+) -> list[int]:
+    """
+    Compute uniform per-rank token counts.
+    Requires ``total_tokens % world_size == 0`` — callers must pad up-front via
+    :func:`pad_modality_for_uniform_sharding`. Uniform sharding lets the
+    All2All custom-op fakes derive output shapes symbolically from input shapes
+    (``x.shape[1] * world_size`` / ``x.shape[1] // world_size``) instead of from
+    Python int args.
+    """
+    if total_tokens % world_size != 0:
+        raise ValueError(
+            f"compute_sequence_partition expects uniform sharding: total_tokens "
+            f"({total_tokens}) must be divisible by world_size ({world_size}). "
+            f"Pad the modality up-front."
+        )
+    per_rank = total_tokens // world_size
+    return [per_rank] * world_size
+
+
+def pad_modality_for_uniform_sharding(
+    modality: Modality,
+    world_size: int,
+) -> tuple[Modality, int]:
+    """Pad the seq dim up to the next multiple of ``world_size`` and attach a
+    padding-aware attention bias so the padded keys are ignored.
+    Returns ``(padded_modality, original_seq_len)``. If no padding is needed
+    the original modality is returned unchanged.
+    """
+    t_orig = modality.latent.shape[1]
+    pad = (-t_orig) % world_size
+    if pad == 0:
+        return modality, t_orig
+
+    t_padded = t_orig + pad
+    b = modality.latent.shape[0]
+    device = modality.latent.device
+    dtype = modality.latent.dtype
+
+    latent_pad = torch.zeros(b, pad, modality.latent.shape[2], dtype=dtype, device=device)
+    latent = torch.cat([modality.latent, latent_pad], dim=1)
+
+    timesteps_pad_shape = list(modality.timesteps.shape)
+    timesteps_pad_shape[1] = pad
+    timesteps_pad = torch.zeros(timesteps_pad_shape, dtype=modality.timesteps.dtype, device=modality.timesteps.device)
+    timesteps = torch.cat([modality.timesteps, timesteps_pad], dim=1)
+
+    positions_pad_shape = list(modality.positions.shape)
+    positions_pad_shape[2] = pad
+    positions_pad = torch.zeros(positions_pad_shape, dtype=modality.positions.dtype, device=modality.positions.device)
+    positions = torch.cat([modality.positions, positions_pad], dim=2)
+
+    if modality.attention_mask is None:
+        # Key-only padding mask in the canonical [0, 1] form: 1 on valid keys,
+        # 0 on padded keys. Shape (1, 1, T_padded) broadcasts across batch and
+        # queries -- O(T) memory instead of materialising a dense (B, T, T)
+        # matrix just to mask `pad` (< world_size) keys.
+        # `_prepare_self_attention_mask` does the standard 3D -> 4D log-space
+        # conversion and produces a (1, 1, 1, T_padded) bias.
+        attention_mask = torch.ones(1, 1, t_padded, dtype=torch.float32, device=device)
+        attention_mask[:, :, t_orig:] = 0.0
+    else:
+        # User-supplied (B, T, T) [0, 1] mask: extend with padded rows/cols.
+        # Padded query rows attend to all valid keys so their softmax stays
+        # well-defined (the outputs are sliced off after the gather, but a
+        # fully-masked row would produce NaN).
+        old = modality.attention_mask
+        attention_mask = torch.zeros(b, t_padded, t_padded, dtype=old.dtype, device=old.device)
+        attention_mask[:, :t_orig, :t_orig] = old
+        attention_mask[:, t_orig:, :t_orig] = 1.0
+
+    padded = replace(
+        modality,
+        latent=latent,
+        timesteps=timesteps,
+        positions=positions,
+        attention_mask=attention_mask,
+    )
+    return padded, t_orig
+
+
+def compute_sequence_offsets(token_counts: list[int]) -> list[int]:
+    """
+    Compute the starting offset for each rank's token partition.
+    Args:
+        token_counts: List of token counts per rank.
+    Returns:
+        List of starting offsets for each rank.
+    """
+    return [0, *accumulate(token_counts[:-1])]
+
+
+def tile_modality_for_rank(
+    modality: Modality,
+    rank: int,
+    world_size: int,
+) -> tuple[Modality, list[int]]:
+    """
+    Tile a modality's tensors for a specific GPU rank.
+    Splits the sequence dimension (dim 1 for latent/timesteps, dim 2 for positions)
+    across GPUs, returning the local tile for the given rank.
+    Args:
+        modality: The modality to tile.
+        rank: Current GPU rank.
+        world_size: Total number of GPUs.
+    Returns:
+        Tuple of (tiled_modality, token_counts_per_rank).
+    """
+    total_tokens = modality.latent.shape[1]
+    token_counts = compute_sequence_partition(total_tokens, world_size)
+    offsets = compute_sequence_offsets(token_counts)
+
+    start = offsets[rank]
+    end = start + token_counts[rank]
+
+    # Tile latent: (B, T, D) -> (B, T_local, D)
+    tiled_latent = modality.latent[:, start:end, :]
+
+    # Tile timesteps: (B, T) -> (B, T_local)
+    tiled_timesteps = modality.timesteps[:, start:end]
+
+    # Tile positions: (B, 3, T, 2) -> (B, 3, T_local, 2)
+    tiled_positions = modality.positions[:, :, start:end, :]
+
+    tiled_modality = replace(
+        modality,
+        latent=tiled_latent,
+        timesteps=tiled_timesteps,
+        positions=tiled_positions,
+    )
+
+    return tiled_modality, token_counts
+
+
+def gather_output_tokens(
+    local_output: torch.Tensor,
+    token_counts: list[int],
+    group: torch.distributed.ProcessGroup | None = None,
+) -> torch.Tensor:
+    """
+    Gather output tokens from all GPUs back into a single tensor.
+    Args:
+        local_output: Local output tensor of shape (B, T_local, D).
+        token_counts: Number of tokens on each rank.
+        group: Process group for communication. If None, uses default group.
+    Returns:
+        Gathered tensor of shape (B, T_total, D) on all ranks.
+    """
+    world_size = len(token_counts)
+    batch_size = local_output.shape[0]
+    hidden_dim = local_output.shape[2]
+
+    # Prepare output tensors for all_gather
+    max_tokens = max(token_counts)
+
+    # Pad local output to max size for uniform all_gather
+    padded_local = torch.zeros(
+        batch_size,
+        max_tokens,
+        hidden_dim,
+        dtype=local_output.dtype,
+        device=local_output.device,
+    )
+    padded_local[:, : local_output.shape[1], :] = local_output
+
+    # All gather padded outputs
+    gathered_list = [torch.zeros_like(padded_local) for _ in range(world_size)]
+    torch.distributed.all_gather(gathered_list, padded_local, group=group)
+
+    # Extract actual tokens (remove padding) and concatenate
+    outputs = []
+    for i, count in enumerate(token_counts):
+        outputs.append(gathered_list[i][:, :count, :])
+
+    return torch.cat(outputs, dim=1)
+
+
+def create_video_self_attention_module_ops(
+    attention_manager: AttentionManager,
+) -> ModuleOps:
+    """
+    Create ModuleOps for patching video self-attention with all2all attention.
+    This patches the `attn1` attribute on BasicAVTransformerBlock instances,
+    which is the video self-attention module.
+    Args:
+        attention_manager: The AttentionManager instance for all2all communication.
+    Returns:
+        ModuleOps that can be used to patch the model.
+    """
+
+    def mutator(module: torch.nn.Module) -> torch.nn.Module:
+        for block in module.transformer_blocks:
+            if not isinstance(block, BasicAVTransformerBlock):
+                continue
+
+            # Video self-attention: ``Attention.forward`` may receive a non-None
+            # ``mask`` (``video.self_attention_mask``), so wrap both slots; the
+            # branch in ``Attention.forward`` then routes to whichever wrapper
+            # corresponds to the actual call.
+            if hasattr(block, "attn1"):
+                attn1 = block.attn1
+                if isinstance(attn1, Attention):
+                    attn1.attention_function = All2AllAttention(attention_manager, attn1.attention_function)
+                    attn1.masked_attention_function = MaskedAll2AllAttention(
+                        attention_manager, attn1.masked_attention_function
+                    )
+            # video_to_audio cross-attention: no current caller passes a mask
+            # (see ``BasicAVTransformerBlock.forward``), so the masked branch
+            # is dead code today. Wrap both slots anyway so that if a future
+            # caller adds a mask, the SP plumbing is already in place rather
+            # than silently bypassing All2All on that path.
+            if hasattr(block, "video_to_audio_attn"):
+                video_to_audio_attn = block.video_to_audio_attn
+                if isinstance(video_to_audio_attn, Attention):
+                    video_to_audio_attn.attention_function = AudioAll2AllAttention(
+                        attention_manager, video_to_audio_attn.attention_function
+                    )
+                    video_to_audio_attn.masked_attention_function = MaskedAudioAll2AllAttention(
+                        attention_manager, video_to_audio_attn.masked_attention_function
+                    )
+        return module
+
+    return ModuleOps(
+        name="video_self_attention_all2all",
+        matcher=lambda module: isinstance(module, LTXModel),
+        mutator=mutator,
+    )
+
+
+class SequenceParallelModelWrapper(torch.nn.Module):
+    def __init__(self, model: torch.nn.Module, attention_manager: AttentionManager):
+        super().__init__()
+        self.model = model
+        self.attention_manager = attention_manager
+
+    @property
+    def num_blocks(self) -> int:
+        return self.model.num_blocks
+
+    def forward(
+        self, video: Modality | None, audio: Modality | None, perturbations: BatchedPerturbationConfig | None
+    ) -> tuple[torch.Tensor | None, torch.Tensor | None]:
+        if video is None:
+            return self.model(video, audio, perturbations)
+
+        # Pad the video seq dim up to a multiple of world_size so all ranks get
+        # equal shards. The attention mask we attach makes the padded keys
+        # invisible to attention; padded rows are sliced off after the gather.
+        video, t_orig = pad_modality_for_uniform_sharding(video, self.attention_manager.world_size)
+
+        video_tile, token_counts = tile_modality_for_rank(
+            video, self.attention_manager.rank, self.attention_manager.world_size
+        )
+        total_tokens = sum(token_counts)
+        if total_tokens > self.attention_manager.max_tokens:
+            raise ValueError(
+                f"Total video token count ({total_tokens}) exceeds attention_manager max_tokens "
+                f"({self.attention_manager.max_tokens}). Use a smaller resolution or fewer frames."
+            )
+        self.attention_manager.set_seqlen_all2all(token_counts)
+        torch.distributed.barrier(self.attention_manager.group)
+        video, audio = self.model(video_tile, audio, perturbations)
+        video = gather_output_tokens(video, token_counts, self.attention_manager.group)
+        # Unpad: drop the rows we added in `pad_modality_for_uniform_sharding` to make
+        # the seq dim divisible by world_size, restoring the caller's original length.
+        if video.shape[1] != t_orig:
+            video = video[:, :t_orig, :]
+        return video, audio

packages/ltx-core/src/ltx_core/multigpu/transformer/tiled_data_parallel.py

@@ -0,0 +1,99 @@
+"""Tiled Data Parallel model wrapper for the LTX transformer.
+Each GPU processes one or more tiles of the patchified
+``(frames, height, width)`` latent.  Tiles are assigned to ranks via
+round-robin, so the number of tiles may exceed the number of GPUs.
+Tiles may overlap; overlapping regions are blended with trapezoidal
+masks so that seam artefacts are suppressed.  Each rank accumulates
+its assigned tiles locally, then a single ``all_reduce`` synchronises
+the blended output across all ranks.
+Conditioning tokens (appended after the generated tokens) are filtered
+per tile: only tokens whose positions overlap with the tile's spatial
+extent (or that have negative time coordinates) are included.
+"""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+import torch
+import torch.distributed as dist
+
+from ltx_core.modality_tiling import VideoModalityTilingHelper
+from ltx_core.model.transformer.modality import Modality
+from ltx_core.tiling import TileCountConfig
+from ltx_core.tools import VideoLatentTools
+
+if TYPE_CHECKING:
+    from ltx_core.guidance.perturbations import BatchedPerturbationConfig
+
+
+class TiledDataParallelModelWrapper(torch.nn.Module):
+    """Wraps an ``X0Model`` for tiled data parallelism.
+    Tiles are distributed across ranks via round-robin, allowing more
+    tiles than GPUs (e.g. 16 tiles on 4 GPUs = 4 tiles per rank).
+    Each rank processes its assigned tiles sequentially, blending each
+    into a full-size accumulator.  A single ``all_reduce(SUM)`` after
+    all local tiles produces the final result (blend masks sum to 1
+    globally across all tiles).
+    Audio is processed untiled on every tile forward; the outputs are
+    summed via ``all_reduce`` and divided by the total tile count so
+    that all ranks stay in sync.
+    """
+
+    def __init__(
+        self,
+        model: torch.nn.Module,
+        *,
+        video_tools: VideoLatentTools,
+        tiling: TileCountConfig,
+        group: dist.ProcessGroup,
+        normalize_positions: bool = True,
+    ) -> None:
+        super().__init__()
+        self.model = model
+        self.group = group
+        self.world_size = dist.get_world_size(group)
+        self._normalize_positions = normalize_positions
+        self._helper = VideoModalityTilingHelper(tiling, video_tools)
+        all_tiles = self._helper.tiles
+        rank = dist.get_rank(group)
+        self._tiles = [t for i, t in enumerate(all_tiles) if i % self.world_size == rank]
+
+    @property
+    def num_blocks(self) -> int:
+        return self.model.num_blocks
+
+    def forward(
+        self,
+        video: Modality | None,
+        audio: Modality | None,
+        perturbations: BatchedPerturbationConfig | None,
+    ) -> tuple[torch.Tensor | None, torch.Tensor | None]:
+        if video is None:
+            return self.model(video, audio, perturbations)
+
+        # Each rank processes its assigned tiles and accumulates locally.
+        denoised_video: torch.Tensor | None = None
+        denoised_audio: torch.Tensor | None = None
+        for tile in self._tiles:
+            tiled_video, ctx = self._helper.tile_modality(video, tile, normalize_positions=self._normalize_positions)
+            tile_out, audio_out = self.model(tiled_video, audio, perturbations)
+            blended = self._helper.blend(tile_out, tile, ctx)
+            denoised_video = blended if denoised_video is None else denoised_video + blended
+            if audio_out is not None:
+                denoised_audio = audio_out if denoised_audio is None else denoised_audio + audio_out
+
+        assert denoised_video is not None
+
+        # All-reduce: sum blended tiles across ranks (masks sum to 1 globally).
+        denoised_video = denoised_video.contiguous()
+        dist.all_reduce(denoised_video, op=dist.ReduceOp.SUM, group=self.group)
+
+        # Average audio across all tile forwards (each saw different video context).
+        if denoised_audio is not None:
+            total_tiles = len(self._helper.tiles)
+            denoised_audio = denoised_audio.contiguous()
+            dist.all_reduce(denoised_audio, op=dist.ReduceOp.SUM, group=self.group)
+            denoised_audio = denoised_audio / total_tiles
+
+        return denoised_video, denoised_audio

packages/ltx-core/src/ltx_core/multigpu/vae/__init__.py

@@ -0,0 +1,5 @@
+"""Multi-GPU utilities for VAE decoding."""
+
+from ltx_core.multigpu.vae.distributed_decoder import DistributedVideoDecoder
+
+__all__ = ["DistributedVideoDecoder"]

packages/ltx-core/src/ltx_core/multigpu/vae/distributed_decoder.py

@@ -0,0 +1,307 @@
+"""Distributed video decoder that partitions the latent across ranks.
+Tiles are assigned to ranks via round-robin, so the number of tiles
+may exceed the number of GPUs (e.g. 16 tiles on 4 GPUs = 4 tiles per
+rank).  Each rank decodes its assigned tiles sequentially.  Workers
+put their list of decoded tiles into a ``mp.Queue`` (CUDA IPC —
+zero-copy handle sharing).  The driver collects all tiles, blends
+overlap zones, and returns temporal batches distributed across devices.
+The tiling configuration comes from ``MGPUConfig.vae_tiling`` (set at
+construction time), NOT from the pipeline's SGPU tiling kwarg.  MGPU
+tiling controls parallelism; SGPU tiling controls single-GPU VRAM
+management — they are independent concerns.
+"""
+
+from __future__ import annotations
+
+import logging
+from collections.abc import Callable, Iterator
+from dataclasses import dataclass
+from typing import TYPE_CHECKING
+
+import torch
+import torch.distributed as dist
+from einops import rearrange
+from torch.multiprocessing import Queue
+
+from ltx_core.model.video_vae.tiling import TilingConfig
+from ltx_core.model.video_vae.video_vae import (
+    VideoDecoder,
+    map_spatial_slice,
+    map_temporal_slice,
+    to_mapping_operation,
+)
+from ltx_core.tiling import (
+    Tile,
+    create_tiles,
+    split_by_count,
+    split_by_count_temporal_causal,
+)
+from ltx_core.types import SpatioTemporalScaleFactors, VideoLatentShape
+
+if TYPE_CHECKING:
+    from ltx_core.tiling import TileCountConfig
+
+logger = logging.getLogger(__name__)
+
+
+# ------------------------------------------------------------------
+# Data structures
+# ------------------------------------------------------------------
+
+
+@dataclass(frozen=True)
+class DecodedTile:
+    """A VAE-decoded tile with pixel-space placement.
+    Attributes:
+        pixels: ``[F_tile, H_tile, W_tile, C]`` in the decoder's native dtype.
+        pixel_tile: Carries ``out_coords`` (f, h, w slices) and ``blend_mask``.
+    """
+
+    pixels: torch.Tensor
+    pixel_tile: Tile
+
+
+# ------------------------------------------------------------------
+# Tile construction helpers
+# ------------------------------------------------------------------
+
+
+def _to_decoded_tile(
+    raw: torch.Tensor,
+    tile: Tile,
+) -> DecodedTile:
+    """Convert raw decoder output ``[B, C, F, H, W]`` to a :class:`DecodedTile`.
+    Rearranges to ``[F, H, W, C]`` and normalises ``[-1, 1] → [0, 1]``.
+    """
+    pixels = rearrange(raw[0], "c f h w -> f h w c")
+    pixels = ((pixels + 1.0) / 2.0).clamp(0.0, 1.0)
+    return DecodedTile(pixels=pixels, pixel_tile=tile)
+
+
+# ------------------------------------------------------------------
+# Tile assembly
+# ------------------------------------------------------------------
+
+
+def compute_summed_weights(
+    tiles: list[DecodedTile],
+    total_frames: int,
+    output_height: int,
+    output_width: int,
+) -> torch.Tensor:
+    """Build the ``[F, H, W]`` denominator for weighted blending."""
+    weights = torch.zeros(total_frames, output_height, output_width)
+    for tile in tiles:
+        f_slice, h_slice, w_slice = tile.pixel_tile.out_coords
+        weights[f_slice, h_slice, w_slice] += tile.pixel_tile.blend_mask
+    return weights.clamp(min=1e-8)
+
+
+def gather_frames(
+    tiles: list[DecodedTile],
+    total_frames: int,
+    output_height: int,
+    output_width: int,
+    num_temporal_batches: int,
+    world_size: int,
+    weights: torch.Tensor,
+    device_fn: Callable[[int], str | torch.device] | None = None,
+) -> Iterator[torch.Tensor]:
+    """Assemble decoded tiles into temporal batches distributed across GPUs.
+    Each temporal batch is allocated on the device returned by *device_fn(batch_index)*.
+    By default batches are placed round-robin on ``cuda:0`` … ``cuda:<world_size-1>``.
+    """
+    if device_fn is None:
+        device_fn = lambda b: f"cuda:{b % world_size}"  # noqa: E731
+
+    batch_size = (total_frames + num_temporal_batches - 1) // num_temporal_batches
+
+    for b in range(num_temporal_batches):
+        batch_range = slice(b * batch_size, min((b + 1) * batch_size, total_frames))
+        batch_len = batch_range.stop - batch_range.start
+        if batch_len <= 0:
+            break
+
+        device = device_fn(b)
+        dtype = tiles[0].pixels.dtype
+        output = torch.zeros(batch_len, output_height, output_width, 3, device=device, dtype=dtype)
+
+        for tile in tiles:
+            f_slice, h_slice, w_slice = tile.pixel_tile.out_coords
+
+            overlap = slice(max(batch_range.start, f_slice.start), min(batch_range.stop, f_slice.stop))
+            if overlap.start >= overlap.stop:
+                continue
+
+            tile_frames = slice(overlap.start - f_slice.start, overlap.stop - f_slice.start)
+            out_frames = slice(overlap.start - batch_range.start, overlap.stop - batch_range.start)
+
+            blend = tile.pixel_tile.blend_mask[tile_frames].to(device=device)
+            output[out_frames, h_slice, w_slice, :] += tile.pixels[tile_frames].to(device=device) * blend[:, :, :, None]
+
+        batch_weights = weights[batch_range.start : batch_range.stop].to(device=device)
+        output.div_(batch_weights[:, :, :, None])
+        yield output
+
+
+# ------------------------------------------------------------------
+# Main class
+# ------------------------------------------------------------------
+
+
+class DistributedVideoDecoder(torch.nn.Module):
+    """Distributed VAE decoder with queue-based tile collection.
+    All ranks decode their latent tile in parallel.  Workers send
+    their :class:`DecodedTile` to the driver rank via the shared
+    ``mp.Queue`` (CUDA IPC — zero-copy).  The driver collects all
+    tiles, blends overlapping regions, and returns temporal batches
+    as an iterator.
+    Parameters
+    ----------
+    decoder:
+        The real (local) ``VideoDecoder`` instance.
+    queue:
+        ``mp.Queue`` shared across all ranks for CUDA IPC tile transfer.
+    vae_group:
+        NCCL process group for the VAE ranks. Used to derive
+        ``rank`` and ``world_size`` within the group.
+    vae_tiling:
+        MGPU tiling config that determines how the latent is split.
+    driver_rank:
+        Group-local rank of the driver process (the rank that collects
+        and assembles tiles).
+    """
+
+    def __init__(
+        self,
+        decoder: VideoDecoder,
+        queue: Queue,  # type: ignore[type-arg]
+        vae_group: dist.ProcessGroup,
+        vae_tiling: TileCountConfig,
+        driver_rank: int = 0,
+    ) -> None:
+        super().__init__()
+        self.decoder = decoder
+        self.queue = queue
+        self.vae_group = vae_group
+        self.rank = dist.get_rank(vae_group)
+        self.world_size = dist.get_world_size(vae_group)
+        self.vae_tiling = vae_tiling
+        self.driver_rank = driver_rank
+
+    def forward(
+        self,
+        sample: torch.Tensor,
+        timestep: torch.Tensor | None = None,
+        generator: torch.Generator | None = None,
+    ) -> torch.Tensor:
+        """Non-tiled path: fall back to local decode."""
+        return self.decoder(sample, timestep, generator)
+
+    def decode_video(
+        self,
+        latent: torch.Tensor,
+        tiling_config: TilingConfig | None = None,
+        generator: torch.Generator | None = None,
+        device_fn: Callable[[int], str | torch.device] | None = None,
+    ) -> Iterator[torch.Tensor]:
+        """Distributed decode — all ranks decode, driver assembles.
+        Not a generator so that worker side-effects (decode + queue.put)
+        execute eagerly regardless of whether the caller iterates.
+        1. Each rank decodes its latent tile (with optional intra-GPU tiling).
+        2. Workers send their :class:`DecodedTile` to the driver via the queue.
+        3. The driver collects all tiles, blends overlaps, and returns
+           temporal batches distributed across GPUs.
+        """
+        if (
+            self.vae_tiling.frames.num_tiles > 1
+            and tiling_config is not None
+            and tiling_config.temporal_config is not None
+        ):
+            raise ValueError(
+                "Cannot combine multi-GPU temporal tiling (vae_tiling.frames.num_tiles > 1) "
+                "with single-GPU temporal tiling (tiling_config.temporal_config). "
+                "Use only one to avoid causal decoding conflicts."
+            )
+
+        latent_shape = VideoLatentShape.from_torch_shape(latent.shape)
+        scale = self.decoder.video_downscale_factors
+        full_shape = latent_shape.upscale(scale)
+
+        # Phase 1: each rank decodes its assigned tiles.
+        my_tiles = self._decode_tiles(latent, latent_shape, scale, generator, tiling_config)
+
+        # Phase 2: workers send tiles to driver.
+        if self.rank != self.driver_rank:
+            self.queue.put((self.rank, my_tiles))
+            return iter([])
+
+        # Phase 3: driver collects and assembles.
+        all_tiles = self._collect_tiles(my_tiles)
+        weights = compute_summed_weights(all_tiles, full_shape.frames, full_shape.height, full_shape.width)
+        batches = gather_frames(
+            all_tiles,
+            full_shape.frames,
+            full_shape.height,
+            full_shape.width,
+            self.world_size,
+            self.world_size,
+            weights,
+            device_fn=device_fn,
+        )
+        return batches
+
+    # ------------------------------------------------------------------
+    # Private helpers
+    # ------------------------------------------------------------------
+
+    def _decode_tiles(
+        self,
+        latent: torch.Tensor,
+        latent_shape: VideoLatentShape,
+        scale: SpatioTemporalScaleFactors,
+        generator: torch.Generator | None,
+        tiling_config: TilingConfig | None = None,
+    ) -> list[DecodedTile]:
+        """Decode this rank's assigned latent tiles and convert to :class:`DecodedTile` list."""
+        all_tiles = create_tiles(
+            torch.Size([latent_shape.frames, latent_shape.height, latent_shape.width]),
+            splitters=[
+                split_by_count_temporal_causal(self.vae_tiling.frames.num_tiles, self.vae_tiling.frames.overlap),
+                split_by_count(self.vae_tiling.height.num_tiles, self.vae_tiling.height.overlap),
+                split_by_count(self.vae_tiling.width.num_tiles, self.vae_tiling.width.overlap),
+            ],
+            mappers=[
+                to_mapping_operation(map_temporal_slice, scale.time),
+                to_mapping_operation(map_spatial_slice, scale.height),
+                to_mapping_operation(map_spatial_slice, scale.width),
+            ],
+        )
+        my_tiles = [t for i, t in enumerate(all_tiles) if i % self.world_size == self.rank]
+        decoded = []
+        for tile in my_tiles:
+            latent_slice = latent[:, :, tile.in_coords[0], tile.in_coords[1], tile.in_coords[2]]
+            if tiling_config is not None:
+                chunks = list(self.decoder.tiled_decode(latent_slice, tiling_config, generator=generator))
+                raw = torch.cat(chunks, dim=2)
+            else:
+                raw = self.decoder.forward(latent_slice, generator=generator)
+            decoded.append(_to_decoded_tile(raw, tile))
+        return decoded
+
+    def _collect_tiles(self, driver_tiles: list[DecodedTile]) -> list[DecodedTile]:
+        """Collect tiles from all workers via the queue. Returns flat list of all tiles.
+        Sorted by rank so the downstream reduction in ``gather_frames`` /
+        ``compute_summed_weights`` (in-place ``+=`` over overlapping pixel
+        regions) processes tiles in a fixed order. Queue-arrival order would
+        otherwise vary run-to-run and yield 1-ulp bf16 drift from
+        non-associative floating-point summation.
+        """
+        per_rank: dict[int, list[DecodedTile]] = {self.driver_rank: driver_tiles}
+        for _ in range(self.world_size - 1):
+            worker_rank, worker_tiles = self.queue.get()
+            per_rank[worker_rank] = worker_tiles
+        result: list[DecodedTile] = []
+        for rank in sorted(per_rank):
+            result.extend(per_rank[rank])
+        return result

packages/ltx-core/src/ltx_core/quantization/__init__.py

@@ -2,15 +2,16 @@ from ltx_core.quantization.fp8_cast import (
     TRANSFORMER_LINEAR_DOWNCAST_MAP,
     UPCAST_DURING_INFERENCE,
     UpcastWithStochasticRounding,
+    fp8_cast_fuse_rule,
 )
-from ltx_core.quantization.fp8_scaled_mm import FP8_PREPARE_MODULE_OPS, FP8_TRANSPOSE_SD_OPS
+from ltx_core.quantization.fp8_scaled_mm import fp8_scaled_mm_fuse_rule
 from ltx_core.quantization.policy import QuantizationPolicy
 
 __all__ = [
-    "FP8_PREPARE_MODULE_OPS",
-    "FP8_TRANSPOSE_SD_OPS",
     "TRANSFORMER_LINEAR_DOWNCAST_MAP",
     "UPCAST_DURING_INFERENCE",
     "QuantizationPolicy",
     "UpcastWithStochasticRounding",
+    "fp8_cast_fuse_rule",
+    "fp8_scaled_mm_fuse_rule",
 ]

packages/ltx-core/src/ltx_core/quantization/blockwise/__init__.py

@@ -0,0 +1,47 @@
+"""Public API for blockwise FP8/FP6 quantization.
+The implementation lives in :mod:`._impl`, which imports the compiled
+``ltx_kernels.blockwise`` kernels at top level. This module deliberately defers
+that import so that ``ltx_core.quantization.blockwise`` remains importable
+without those kernels built; the gate fires only when one of the policy builders
+is actually called.
+"""
+
+from ltx_core.quantization.policy import QuantizationPolicy
+
+__all__ = ["build_fp6_policy", "build_fp8_policy"]
+
+
+def _import_impl():  # noqa: ANN202 - internal helper
+    try:
+        from ltx_core.quantization.blockwise import _impl  # noqa: PLC0415
+
+        return _impl
+    except ImportError as e:
+        raise RuntimeError(
+            "ltx-kernels not built; blockwise FP8/FP6 quantization requires it. "
+            "Build it on a CUDA host with `uv sync --group kernels` (or "
+            "`uv pip install -e packages/ltx-kernels --no-build-isolation`) before "
+            "calling build_fp8_policy() / build_fp6_policy()."
+        ) from e
+
+
+def build_fp8_policy() -> QuantizationPolicy:
+    """Build a blockwise FP8 quantization policy. Raises ``RuntimeError`` if ``ltx-kernels`` is not built."""
+    impl = _import_impl()
+    return QuantizationPolicy(
+        sd_ops=impl.build_sd_ops_fp8(),
+        module_ops=(impl.build_module_ops_fp8(),),
+        model_configurator=impl.BlockwiseFP8LTXModelConfigurator,
+        fuse_rule=impl.fuse_rule_fp8,
+    )
+
+
+def build_fp6_policy() -> QuantizationPolicy:
+    """Build a blockwise FP6 quantization policy. Raises ``RuntimeError`` if ``ltx-kernels`` is not built."""
+    impl = _import_impl()
+    return QuantizationPolicy(
+        sd_ops=impl.build_sd_ops_fp6(),
+        module_ops=(impl.build_module_ops_fp6(),),
+        model_configurator=impl.BlockwiseFP6LTXModelConfigurator,
+        fuse_rule=impl.fuse_rule_fp6,
+    )

packages/ltx-core/src/ltx_core/quantization/blockwise/_impl.py

@@ -0,0 +1,431 @@
+"""Implementation of blockwise FP8/FP6 quantization. Depends on ``ltx_kernels``.
+This module imports the compiled ``ltx_kernels.blockwise`` kernels at top level
+— without them built, simply importing this file raises :class:`ImportError`.
+The intended access path is through ``ltx_core.quantization.blockwise.__init__``
+which catches that and re-raises as a clean :class:`RuntimeError`. Do not import
+this module directly from non-quantization code.
+"""
+
+from typing import Callable, ClassVar, List, NamedTuple, Protocol, Type
+
+import torch
+from ltx_kernels.blockwise.functional import (
+    blockwise_dequantize,
+    blockwise_quantize_adanorm_triton,
+    blockwise_quantize_rms_fma_triton,
+    fp6_blockwise_quantize_weights_torch,
+    fp6_pack_tensor,
+    fp6_unpack_tensor,
+    fp8_blockwise_quantize_weights_torch,
+    gated_attention_triton,
+    rms_norm_rope,
+    rms_norm_split_rope,
+)
+from ltx_kernels.blockwise.linear import BlockwiseFP6Linear, BlockwiseFP8Linear
+from torch import nn
+
+from ltx_core.loader.fuse_loras import FuseRule, bf16_fuse_rule
+from ltx_core.loader.module_ops import ModuleOps
+from ltx_core.loader.primitives import StateDict
+from ltx_core.loader.sd_ops import KeyValueOperationResult, SDOps
+from ltx_core.model.model_protocol import ModelConfigurator
+from ltx_core.model.transformer import LTXModel
+from ltx_core.model.transformer.model_configurator import LTXModelConfigurator, LTXVideoOnlyModelConfigurator
+from ltx_core.model.transformer.ops import (
+    AdaZeroCallable,
+    GatedAttentionCallable,
+    PostSACallable,
+    PreAttentionCallable,
+)
+from ltx_core.model.transformer.rope import LTXRopeType
+from ltx_core.model.transformer.transformer import TransformerOpsConfig
+
+
+class FromLinearProtocol(Protocol):
+    """Protocol for nn.Module subclasses that can be constructed from an nn.Linear."""
+
+    @classmethod
+    def from_linear(cls, linear: nn.Linear, transform_weights: bool = True) -> nn.Module: ...
+
+
+class BlockwiseQuantizedWeight(NamedTuple):
+    """Result of blockwise quantization: a quantized weight tensor and its per-block scale.
+    For FP8: ``weight`` is ``float8_e4m3fn``, ``scale`` is ``float32`` shaped
+    ``[out // 128, in // 128]``.
+    For FP6: ``weight`` is packed ``uint8`` shaped ``[out, (in // 4) * 3]``,
+    ``scale`` is ``float32`` shaped ``[out // 128, in // 128]``.
+    """
+
+    weight: torch.Tensor
+    scale: torch.Tensor
+
+
+EXCLUDED_LAYER_SUBSTRINGS = (
+    "patchify_proj",
+    "adaln_single",
+    "av_ca_video_scale_shift_adaln_single",
+    "av_ca_a2v_gate_adaln_single",
+    "caption_projection",
+    "proj_out",
+    "audio_patchify_proj",
+    "audio_adaln_single",
+    "av_ca_audio_scale_shift_adaln_single",
+    "av_ca_v2a_gate_adaln_single",
+    "audio_caption_projection",
+    "audio_proj_out",
+    "to_gate_logits",
+    "scale_shift_table",
+)
+
+
+_QUANTIZABLE_FLOAT_DTYPES = (torch.bfloat16, torch.float16, torch.float32)
+
+
+def _is_quantizable_float(x: torch.Tensor | torch.dtype) -> bool:
+    """Whether ``x`` is an unquantized high-precision float (bf16 / fp16 / fp32).
+    FP8 / FP6 weights are floats too but they're already in a quantized layout
+    and must not be re-quantized.
+    """
+    dtype = x.dtype if isinstance(x, torch.Tensor) else x
+    return dtype in _QUANTIZABLE_FLOAT_DTYPES
+
+
+def _should_skip_layer(layer_name: str, excluded_layer_substrings: tuple[str, ...]) -> bool:
+    return any(substring in layer_name for substring in excluded_layer_substrings)
+
+
+def _replace_linear_modules(model: torch.nn.Module, linear_cls: Type[FromLinearProtocol]) -> torch.nn.Module:
+    skip_list = ["to_gate_logits", "scale_shift_table"]
+    for name, module in model.named_modules():
+        if "transformer_block" in name and isinstance(module, torch.nn.Linear):
+            if _should_skip_layer(name, skip_list):
+                continue
+            *parent_path, child_name = name.split(".")
+            parent = model
+            for part in parent_path:
+                parent = getattr(parent, part)
+            setattr(
+                parent,
+                child_name,
+                linear_cls.from_linear(module, False),
+            )
+            del module.weight
+            del module.bias
+            torch.cuda.empty_cache()
+    return model
+
+
+# ---------------------------------------------------------------------------
+# Weight quantization helpers
+# ---------------------------------------------------------------------------
+
+
+def _blockwise_quantize_weight_helper(
+    value: torch.Tensor,
+    quant_fn: Callable[[torch.Tensor, int], tuple[torch.Tensor, torch.Tensor]],
+    pack_fn: Callable[[torch.Tensor], torch.Tensor],
+) -> BlockwiseQuantizedWeight:
+    orig_device = value.device
+    w_quant, w_scales = quant_fn(value.cuda())
+    return BlockwiseQuantizedWeight(
+        weight=pack_fn(w_quant).to(device=orig_device),
+        scale=w_scales.to(device=orig_device),
+    )
+
+
+def _fp8_blockwise_quantize_weight(value: torch.Tensor) -> BlockwiseQuantizedWeight:
+    return _blockwise_quantize_weight_helper(value, fp8_blockwise_quantize_weights_torch, lambda x: x)
+
+
+def _fp6_blockwise_quantize_weight(value: torch.Tensor) -> BlockwiseQuantizedWeight:
+    return _blockwise_quantize_weight_helper(value, fp6_blockwise_quantize_weights_torch, fp6_pack_tensor)
+
+
+def _create_weight_quantize_op(
+    excluded_layer_substrings: tuple[str, ...],
+    quantization_func: Callable[[torch.Tensor], BlockwiseQuantizedWeight],
+) -> Callable[[str, torch.Tensor], list[KeyValueOperationResult]]:
+    """KeyValueOperation that blockwise-quantizes a 2D BF16 ``.weight`` and emits ``.weight_scale``."""
+
+    def quantize_weight(key: str, value: torch.Tensor) -> list[KeyValueOperationResult]:
+        if _should_skip_layer(key, excluded_layer_substrings):
+            return [KeyValueOperationResult(key, value)]
+        if value.dim() != 2 or not _is_quantizable_float(value):
+            return [KeyValueOperationResult(key, value)]
+        quantized = quantization_func(value)
+        scale_key = key.replace(".weight", ".weight_scale")
+        return [
+            KeyValueOperationResult(key, quantized.weight),
+            KeyValueOperationResult(scale_key, quantized.scale),
+        ]
+
+    return quantize_weight
+
+
+def _create_bias_to_fp32_op(
+    excluded_layer_substrings: tuple[str, ...],
+) -> Callable[[str, torch.Tensor], list[KeyValueOperationResult]]:
+    """KeyValueOperation that casts a ``.bias`` tensor to FP32.
+    ``BlockwiseFP{8,6}Linear`` registers ``.bias`` as float32; the load-time
+    cast keeps the checkpoint's BF16 bias compatible with that param dtype.
+    """
+
+    def bias_to_fp32(key: str, value: torch.Tensor) -> list[KeyValueOperationResult]:
+        if _should_skip_layer(key, excluded_layer_substrings):
+            return [KeyValueOperationResult(key, value)]
+        return [KeyValueOperationResult(key, value.float())]
+
+    return bias_to_fp32
+
+
+# ---------------------------------------------------------------------------
+# Q8 activation callables (formerly in model.transformer.ops)
+# ---------------------------------------------------------------------------
+
+
+class Q8KernelsPreAttention(PreAttentionCallable):
+    def __call__(
+        self,
+        q: torch.Tensor,
+        k: torch.Tensor,
+        attn_module: nn.Module,
+        mask: torch.Tensor | None,  # noqa: ARG002
+        pe: torch.Tensor | None,
+        k_pe: torch.Tensor | None,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        if attn_module.rope_type == LTXRopeType.INTERLEAVED:
+            rope_func = rms_norm_rope
+        elif attn_module.rope_type == LTXRopeType.SPLIT:
+            rope_func = rms_norm_split_rope
+        else:
+            raise ValueError(f"Invalid rope type: {attn_module.rope_type}")
+
+        if pe is not None:
+            k_pe = k_pe if k_pe is not None else pe
+            q = rope_func(q, pe[0], pe[1], attn_module.q_norm.weight, False)
+            k = rope_func(k, k_pe[0], k_pe[1], attn_module.k_norm.weight, False)
+        else:
+            q = attn_module.q_norm(q)
+            k = attn_module.k_norm(k)
+        return q, k
+
+
+class Q8KernelsAdaZeroFunction(AdaZeroCallable):
+    def __call__(
+        self,
+        x: torch.Tensor,
+        eps: float,  # noqa: ARG002
+        scale: torch.Tensor,
+        shift: torch.Tensor,
+    ) -> torch.Tensor:
+        return blockwise_quantize_adanorm_triton(x, None, scale, shift, torch.float8_e4m3fn, 1.0)
+
+
+class Q8KernelsPostSAFunction(PostSACallable):
+    def __call__(
+        self,
+        x: torch.Tensor,
+        y: torch.Tensor,
+        norm_weights: torch.Tensor | None,  # noqa: ARG002
+        eps: float,  # noqa: ARG002
+        gate: torch.Tensor,
+    ) -> List[torch.Tensor]:
+        # Dequantize the fused result: the cross-attention AdaLN path applies a BF16
+        # scale/shift, which cannot operate on the (fp8, scales) payload.
+        normed_fp8 = blockwise_quantize_rms_fma_triton(x, y, gate)
+        return x, blockwise_dequantize(normed_fp8)
+
+
+class Q8KernelsGatedAttention(GatedAttentionCallable):
+    def __call__(
+        self,
+        x: torch.Tensor,
+        attn_out: torch.Tensor,
+        attn_module: nn.Module,
+    ) -> torch.Tensor:
+        # Self-attention path: ``x`` arrives as the ``(fp8, scales)`` tuple
+        # produced by Q8KernelsAdaZeroFunction. Cross-attention path
+        # (apply_cross_attention_adaln) feeds plain BF16, so dequantize only
+        # when needed.
+        if isinstance(x, tuple):
+            x = blockwise_dequantize(x)
+        gate_logits = attn_module.to_gate_logits(x)
+        return gated_attention_triton(attn_out, gate_logits)
+
+
+# ---------------------------------------------------------------------------
+# Fuse rules
+# ---------------------------------------------------------------------------
+
+
+_BLOCK = 128
+
+
+def _blockwise_dequantize_2d(weight_fp8: torch.Tensor, weight_scale: torch.Tensor) -> torch.Tensor:
+    """Dequantize a 2D blockwise-FP8 weight ``[out, in]`` with per-block scale
+    ``[out//128, in//128]`` to BF16.
+    ``ltx_kernels.blockwise.blockwise_dequantize`` is built for 3D activations where
+    scales are ``[b*s, in//128]`` — one row per token. Weights are block-
+    quantized along the row dim too, so we expand the row axis 128x via
+    ``repeat_interleave`` and reuse the kernel.
+    """
+    out_features, in_features = weight_fp8.shape
+    scales_per_row = weight_scale.repeat_interleave(_BLOCK, dim=0)
+    return blockwise_dequantize((weight_fp8.unsqueeze(0), scales_per_row)).view(out_features, in_features)
+
+
+def _blockwise_fp8_fuse(
+    key: str,
+    weight: torch.Tensor,
+    deltas: torch.Tensor,
+    model_sd: StateDict,
+) -> dict[str, torch.Tensor]:
+    """Dequantize the FP8 weight + per-block scale to BF16, add the BF16 delta,
+    and re-quantize blockwise. Both ``.weight`` and the companion
+    ``.weight_scale`` are emitted so the loaded layer matches what
+    ``BlockwiseFP8Linear`` expects.
+    Excluded layers (see ``EXCLUDED_LAYER_SUBSTRINGS``) stay BF16 and have no
+    ``.weight_scale`` companion — for those, fall back to a plain bf16 fuse.
+    """
+    scale_key = key.replace(".weight", ".weight_scale")
+    if scale_key not in model_sd.sd:
+        return bf16_fuse_rule(key, weight, deltas, model_sd)
+    weight_scale = model_sd.sd[scale_key]
+    bf16_weight = _blockwise_dequantize_2d(weight, weight_scale)
+    merged = bf16_weight + deltas.to(dtype=bf16_weight.dtype)
+    new_fp8_weight, new_weight_scale = fp8_blockwise_quantize_weights_torch(merged.cuda())
+    return {
+        key: new_fp8_weight.to(device=weight.device),
+        scale_key: new_weight_scale.to(device=weight.device),
+    }
+
+
+def _blockwise_fp6_fuse(
+    key: str,
+    weight: torch.Tensor,
+    deltas: torch.Tensor,
+    model_sd: StateDict,
+) -> dict[str, torch.Tensor]:
+    """Mirror ``BlockwiseFP6Linear.fp8weight`` for the dequant side: unpack the
+    packed ``uint8`` weight to ``float8_e4m3fn``, dequantize via the per-block
+    scale to BF16, add the BF16 delta, re-quantize to FP6, and pack back to
+    uint8. Both ``.weight`` (packed uint8) and ``.weight_scale`` are emitted.
+    Note: ``fp6_unpack_tensor`` restores the dropped e_1/e_2 exponent bits as 0,
+    so the dequant->add->requant round-trip is lossy on those bits even when no
+    LoRA delta is applied. This matches what ``BlockwiseFP6Linear`` already does
+    at inference time via its ``fp8weight`` property, so the fused weight is
+    numerically consistent with the unfused inference path.
+    Excluded layers (see ``EXCLUDED_LAYER_SUBSTRINGS``) stay BF16 and have no
+    ``.weight_scale`` companion — for those, fall back to a plain bf16 fuse.
+    """
+    scale_key = key.replace(".weight", ".weight_scale")
+    if scale_key not in model_sd.sd:
+        return bf16_fuse_rule(key, weight, deltas, model_sd)
+    weight_scale = model_sd.sd[scale_key]
+    # Packed shape is [out, (in // 4) * 3]; recover in_features.
+    original_n = weight.shape[-1] * 4 // 3
+    fp8_view = fp6_unpack_tensor(weight, original_n).view(torch.float8_e4m3fn)
+    bf16_weight = _blockwise_dequantize_2d(fp8_view, weight_scale)
+    merged = bf16_weight + deltas.to(dtype=bf16_weight.dtype)
+    new_fp8, new_scale = fp6_blockwise_quantize_weights_torch(merged)
+    new_packed = fp6_pack_tensor(new_fp8.view(torch.uint8))
+    return {
+        key: new_packed.to(device=weight.device),
+        scale_key: new_scale.to(device=weight.device),
+    }
+
+
+# ---------------------------------------------------------------------------
+# Configurators (TransformerOpsConfig with Q8 activation callables)
+# ---------------------------------------------------------------------------
+
+
+def _build_blockwise_ops_config() -> TransformerOpsConfig:
+    return TransformerOpsConfig.from_functions(
+        preattention=Q8KernelsPreAttention(),
+        gated_attention=Q8KernelsGatedAttention(),
+        ada_zero=Q8KernelsAdaZeroFunction(),
+        post_sa=Q8KernelsPostSAFunction(),
+    )
+
+
+# FP6 is weight-only; activation ops match FP8.
+_BLOCKWISE_OPS = _build_blockwise_ops_config()
+
+
+class BlockwiseFP8LTXModelConfigurator(ModelConfigurator[LTXModel]):
+    BASE: ClassVar[type[ModelConfigurator[LTXModel]]] = LTXModelConfigurator
+    OPS: ClassVar[TransformerOpsConfig] = _BLOCKWISE_OPS
+
+    @classmethod
+    def from_config(cls, config: dict) -> LTXModel:
+        return cls.BASE.from_config(config, ops=cls.OPS)
+
+
+class BlockwiseFP8LTXVideoOnlyModelConfigurator(BlockwiseFP8LTXModelConfigurator):
+    BASE = LTXVideoOnlyModelConfigurator
+
+
+class BlockwiseFP6LTXModelConfigurator(BlockwiseFP8LTXModelConfigurator):
+    pass
+
+
+class BlockwiseFP6LTXVideoOnlyModelConfigurator(BlockwiseFP8LTXVideoOnlyModelConfigurator):
+    pass
+
+
+# ---------------------------------------------------------------------------
+# SDOps / ModuleOps / FuseRule assembly
+# ---------------------------------------------------------------------------
+
+
+def build_sd_ops_fp8() -> SDOps:
+    return (
+        SDOps("blockwise_fp8_weights")
+        .with_kv_operation(
+            _create_weight_quantize_op(EXCLUDED_LAYER_SUBSTRINGS, _fp8_blockwise_quantize_weight),
+            key_prefix="transformer_blocks.",
+            key_suffix=".weight",
+        )
+        .with_kv_operation(
+            _create_bias_to_fp32_op(EXCLUDED_LAYER_SUBSTRINGS),
+            key_prefix="transformer_blocks.",
+            key_suffix=".bias",
+        )
+    )
+
+
+def build_sd_ops_fp6() -> SDOps:
+    return (
+        SDOps("blockwise_fp6_weights")
+        .with_kv_operation(
+            _create_weight_quantize_op(EXCLUDED_LAYER_SUBSTRINGS, _fp6_blockwise_quantize_weight),
+            key_prefix="transformer_blocks.",
+            key_suffix=".weight",
+        )
+        .with_kv_operation(
+            _create_bias_to_fp32_op(EXCLUDED_LAYER_SUBSTRINGS),
+            key_prefix="transformer_blocks.",
+            key_suffix=".bias",
+        )
+    )
+
+
+def build_module_ops_fp8() -> ModuleOps:
+    return ModuleOps(
+        name="blockwise_fp8_prepare_for_loading",
+        matcher=lambda model: isinstance(model, LTXModel),
+        mutator=lambda model: _replace_linear_modules(model, BlockwiseFP8Linear),
+    )
+
+
+def build_module_ops_fp6() -> ModuleOps:
+    return ModuleOps(
+        name="blockwise_fp6_prepare_for_loading",
+        matcher=lambda model: isinstance(model, LTXModel),
+        mutator=lambda model: _replace_linear_modules(model, BlockwiseFP6Linear),
+    )
+
+
+fuse_rule_fp8 = FuseRule(aggregation_dtype=torch.bfloat16, fuse_fn=_blockwise_fp8_fuse)
+fuse_rule_fp6 = FuseRule(aggregation_dtype=torch.bfloat16, fuse_fn=_blockwise_fp6_fuse)

packages/ltx-core/src/ltx_core/quantization/fp8_cast.py

@@ -1,14 +1,26 @@
+from pathlib import Path
+
+import safetensors
 import torch
 
+from ltx_core.loader.fuse_loras import FuseRule, bf16_fuse_rule
+from ltx_core.loader.kernels import TRITON_AVAILABLE
 from ltx_core.loader.module_ops import ModuleOps
+from ltx_core.loader.primitives import StateDict
 from ltx_core.loader.sd_ops import KeyValueOperationResult, SDOps
 from ltx_core.model.transformer.model import LTXModel
+from ltx_core.quantization.policy import QuantizationPolicy
 
 BLOCK_SIZE = 1024
 
 
-def _fused_add_round_launch(target_weight: torch.Tensor, original_weight: torch.Tensor, seed: int) -> torch.Tensor:
+def fused_add_round_launch(target_weight: torch.Tensor, original_weight: torch.Tensor, seed: int) -> torch.Tensor:
-    # Lazy import triton - only available on CUDA platforms
+    if not TRITON_AVAILABLE:
+        raise RuntimeError(
+            "fused_add_round_launch requires Triton, which is not available on this platform. "
+            "Callers should gate on ltx_core.loader.kernels.TRITON_AVAILABLE and use a "
+            "deterministic-rounding fallback instead."
+        )
     import triton  # noqa: PLC0415
 
     from ltx_core.loader.kernels import fused_add_round_kernel  # noqa: PLC0415
@@ -53,10 +65,13 @@ def _upcast_and_round(
     """
     Upcast the weight to the given dtype and optionally apply stochastic rounding.
     Input weight needs to have float8_e4m3fn or float8_e5m2 dtype.
+    Stochastic rounding is implemented via a Triton kernel. When Triton is not
+    available (e.g., on Windows), this falls back to deterministic (nearest)
+    rounding via ``weight.to(dtype)``.
     """
-    if not with_stochastic_rounding:
+    if not with_stochastic_rounding or not TRITON_AVAILABLE or weight.device.type != "cuda":
         return weight.to(dtype)
-    return _fused_add_round_launch(torch.zeros_like(weight, dtype=dtype), weight, seed)
+    return fused_add_round_launch(torch.zeros_like(weight, dtype=dtype), weight, seed)
 
 
 class Fp8CastLinear(torch.nn.Linear):
@@ -82,68 +97,87 @@ class Fp8CastLinear(torch.nn.Linear):
 def _replace_fwd_with_upcast(layer: torch.nn.Linear, with_stochastic_rounding: bool = False, seed: int = 0) -> None:
     """
     Intended to be applied via __class__ reassignment to existing nn.Linear
-    instances so that their parameter and buffer tensors are preserved in-place,
+    instances. Forward remains defined at the class level, which is required for
-    avoiding re-instantiation. Forward remains defined at the class level, which
+    torch.compile compatibility — instance-level closure monkey-patches cause
-    is required for torch.compile compatibility — instance-level closure
+    graph breaks.
-    monkey-patches cause graph breaks.
+    Also retypes ``weight`` and ``bias`` to fp8 so the meta param dtype matches
+    the post-load tensor dtype (sd_ops downcasts checkpoint bf16 -> fp8 at load).
+    Block streaming relies on this to derive pool buffer layout from the meta
+    model without an eager checkpoint read.
     """
     layer.__class__ = Fp8CastLinear
     layer._with_stochastic_rounding = with_stochastic_rounding
     layer._seed = seed
+    layer.weight = torch.nn.Parameter(
+        torch.empty(layer.weight.shape, dtype=torch.float8_e4m3fn, device=layer.weight.device),
+        requires_grad=layer.weight.requires_grad,
+    )
+    if layer.bias is not None:
+        layer.bias = torch.nn.Parameter(
+            torch.empty(layer.bias.shape, dtype=torch.float8_e4m3fn, device=layer.bias.device),
+            requires_grad=layer.bias.requires_grad,
+        )
+
+
+# Module-name suffixes for the Linears that participate in fp8 cast. Used by
+# both the upcast matcher and the sd_ops downcast map so the two cannot drift.
+# - ``.to_q`` / ``.to_k`` / ``.to_v`` / ``.to_out.0`` have a leading dot so they
+#   only match the attention Linears at ``...attnN.to_q`` etc.
+# - ``ff.net.0.proj`` / ``ff.net.2`` are intentionally **dotless** so they match
+#   both video FF (``...ff.net.0.proj``) and audio FF (``...audio_ff.net.0.proj``).
+_FP8_CAST_KEY_PREFIX = "transformer_blocks."
+_FP8_CAST_LINEAR_SUFFIXES: tuple[str, ...] = (
+    ".to_q",
+    ".to_k",
+    ".to_v",
+    ".to_out.0",
+    "ff.net.0.proj",
+    "ff.net.2",
+)
+
+
+def _is_fp8_cast_linear(module_name: str) -> bool:
+    """Return True if *module_name* names a Linear that should be fp8-cast."""
+    if _FP8_CAST_KEY_PREFIX not in module_name:
+        return False
+    return any(module_name.endswith(suffix) for suffix in _FP8_CAST_LINEAR_SUFFIXES)
 
 
 def _amend_forward_with_upcast(
     model: torch.nn.Module, with_stochastic_rounding: bool = False, seed: int = 0
 ) -> torch.nn.Module:
     """
-    Replace the forward method of the model's Linear layers to forward
+    Replace the forward method of the fp8-cast Linear layers (per
-    with upcast and optional stochastic rounding.
+    :data:`_FP8_CAST_LINEAR_SUFFIXES`) to forward with upcast and optional
+    stochastic rounding.
+    Only the Linears whose weights are downcast by :data:`TRANSFORMER_LINEAR_DOWNCAST_MAP`
+    are retyped. Linears outside that subset (e.g. ``to_gate_logits``) are left as
+    plain ``nn.Linear`` so the meta-model param dtype matches the loaded checkpoint
+    dtype.
     """
-    for m in model.modules():
+    for name, m in model.named_modules():
-        if isinstance(m, (torch.nn.Linear)):
+        if isinstance(m, torch.nn.Linear) and _is_fp8_cast_linear(name):
             _replace_fwd_with_upcast(m, with_stochastic_rounding, seed)
     return model
 
 
-TRANSFORMER_LINEAR_DOWNCAST_MAP = (
+def _build_transformer_linear_downcast_map() -> SDOps:
-    SDOps("TRANSFORMER_LINEAR_DOWNCAST_MAP")
+    """Build the sd_ops downcast map from the same suffix registry as the matcher."""
-    .with_kv_operation(
+    ops = SDOps("TRANSFORMER_LINEAR_DOWNCAST_MAP")
-        key_prefix="transformer_blocks.", key_suffix=".to_q.weight", operation=_naive_weight_or_bias_downcast
+    for suffix in _FP8_CAST_LINEAR_SUFFIXES:
+        ops = ops.with_kv_operation(
+            key_prefix=_FP8_CAST_KEY_PREFIX,
+            key_suffix=suffix + ".weight",
+            operation=_naive_weight_or_bias_downcast,
+        ).with_kv_operation(
+            key_prefix=_FP8_CAST_KEY_PREFIX,
+            key_suffix=suffix + ".bias",
+            operation=_naive_weight_or_bias_downcast,
         )
-    .with_kv_operation(
+    return ops
-        key_prefix="transformer_blocks.", key_suffix=".to_q.bias", operation=_naive_weight_or_bias_downcast
+
-    )
+
-    .with_kv_operation(
+TRANSFORMER_LINEAR_DOWNCAST_MAP = _build_transformer_linear_downcast_map()
-        key_prefix="transformer_blocks.", key_suffix=".to_k.weight", operation=_naive_weight_or_bias_downcast
-    )
-    .with_kv_operation(
-        key_prefix="transformer_blocks.", key_suffix=".to_k.bias", operation=_naive_weight_or_bias_downcast
-    )
-    .with_kv_operation(
-        key_prefix="transformer_blocks.", key_suffix=".to_v.weight", operation=_naive_weight_or_bias_downcast
-    )
-    .with_kv_operation(
-        key_prefix="transformer_blocks.", key_suffix=".to_v.bias", operation=_naive_weight_or_bias_downcast
-    )
-    .with_kv_operation(
-        key_prefix="transformer_blocks.", key_suffix=".to_out.0.weight", operation=_naive_weight_or_bias_downcast
-    )
-    .with_kv_operation(
-        key_prefix="transformer_blocks.", key_suffix=".to_out.0.bias", operation=_naive_weight_or_bias_downcast
-    )
-    .with_kv_operation(
-        key_prefix="transformer_blocks.", key_suffix="ff.net.0.proj.weight", operation=_naive_weight_or_bias_downcast
-    )
-    .with_kv_operation(
-        key_prefix="transformer_blocks.", key_suffix="ff.net.0.proj.bias", operation=_naive_weight_or_bias_downcast
-    )
-    .with_kv_operation(
-        key_prefix="transformer_blocks.", key_suffix="ff.net.2.weight", operation=_naive_weight_or_bias_downcast
-    )
-    .with_kv_operation(
-        key_prefix="transformer_blocks.", key_suffix="ff.net.2.bias", operation=_naive_weight_or_bias_downcast
-    )
-)
 
 UPCAST_DURING_INFERENCE = ModuleOps(
     name="upcast_fp8_during_linear_forward",
@@ -165,3 +199,139 @@ class UpcastWithStochasticRounding(ModuleOps):
             matcher=lambda model: isinstance(model, LTXModel),
             mutator=lambda model: _amend_forward_with_upcast(model, True, seed),
         )
+
+
+def fuse_cast_fp8_weight(
+    delta_bf16: torch.Tensor,
+    weight_fp8: torch.Tensor,
+) -> torch.Tensor:
+    """Return ``(delta_bf16 + dequantize(weight_fp8)).to(weight_fp8.dtype)``.
+    CUDA with Triton uses stochastic rounding via the fused kernel; otherwise
+    falls back to a deterministic bf16 add. ``delta_bf16`` is the bf16
+    accumulator and is mutated in place.
+    """
+    if delta_bf16.dtype != torch.bfloat16:
+        raise ValueError(f"delta_bf16 must be bfloat16, got {delta_bf16.dtype}")
+    if str(weight_fp8.device).startswith("cuda") and TRITON_AVAILABLE:
+        fused_add_round_launch(delta_bf16, weight_fp8, seed=0)
+    else:
+        delta_bf16.add_(weight_fp8.to(dtype=torch.bfloat16))
+    return delta_bf16.to(dtype=weight_fp8.dtype)
+
+
+def _fp8_cast_fuse(
+    key: str,
+    weight: torch.Tensor,
+    deltas: torch.Tensor,
+    model_sd: StateDict,
+) -> dict[str, torch.Tensor]:
+    """Cast the dequantized FP8 weight + BF16 deltas back to ``weight.dtype``
+    (FP8) via the fused-add-round kernel on CUDA.
+    Only a subset of linears are FP8-downcast (see ``TRANSFORMER_LINEAR_DOWNCAST_MAP``);
+    LoRAs may also target layers left in BF16 (e.g. audio ``add_q/k/v_proj``, cross-modal
+    projections). For those, fall back to a plain BF16 fuse.
+    """
+    if weight.dtype not in (torch.float8_e4m3fn, torch.float8_e5m2):
+        return bf16_fuse_rule(key, weight, deltas, model_sd)
+    return {key: fuse_cast_fp8_weight(deltas, weight)}
+
+
+fp8_cast_fuse_rule = FuseRule(aggregation_dtype=torch.bfloat16, fuse_fn=_fp8_cast_fuse)
+
+
+# Raw safetensors storage prefix shared by every diffusion-transformer
+# parameter (and every prequant `*_scale` sibling). Verified against
+# ltx-2.3-22b-{dev,distilled}-fp8.safetensors: 2924/2924 and 2992/2992 of
+# the scale keys start with this exact prefix.
+_RAW_DIFFUSION_MODEL_PREFIX = "model.diffusion_model."
+
+
+def _read_scales(checkpoint_path: str | Path) -> dict[str, torch.Tensor]:
+    """Return ``{post_rename_param_key: scale_tensor}`` for every prequant
+    ``*_scale`` sibling in *checkpoint_path*.
+    Keys are returned in the post-rename form the loader will pass to the
+    sd-op (e.g. ``transformer_blocks.0.attn1.to_q.weight``) -- the raw
+    ``model.diffusion_model.`` prefix and the ``_scale`` suffix are both
+    stripped. Catches both ``.weight_scale`` and ``.bias_scale``; the
+    latter is absent in the current LTX-2.3 prequant checkpoints but
+    accepted for forward compatibility.
+    """
+    out: dict[str, torch.Tensor] = {}
+    with safetensors.safe_open(str(checkpoint_path), framework="pt", device="cpu") as h:
+        raw_keys = h.keys()
+        for k in raw_keys:
+            if not k.endswith("_scale"):
+                continue
+            if not k.startswith(_RAW_DIFFUSION_MODEL_PREFIX):
+                raise ValueError(
+                    f"Scale key {k!r} does not start with the expected raw prefix {_RAW_DIFFUSION_MODEL_PREFIX!r}"
+                )
+            param_key = k.removeprefix(_RAW_DIFFUSION_MODEL_PREFIX).removesuffix("_scale")
+            out[param_key] = h.get_tensor(k)
+    return out
+
+
+def _build_prequant_fold_sd_ops(scales: dict[str, torch.Tensor]) -> SDOps:
+    """Build sd-ops that fold prequant ``*_scale`` siblings into their parent
+    tensor at load time.
+    *scales* is keyed by the **post-rename** param key (e.g.
+    ``transformer_blocks.0.attn1.to_q.weight``); see :func:`_read_scales`.
+    Four ``with_kv_operation`` entries (symmetric for ``.weight`` and ``.bias``):
+    * ``.weight`` / ``.bias`` -> if a sibling scale exists in *scales*, fold;
+      then delegate to ``TRANSFORMER_LINEAR_DOWNCAST_MAP`` (downcast covered
+      Linears, pass everything else through). Without a scale, delegate
+      directly.
+    * ``.weight_scale`` / ``.bias_scale`` -> drop (the scale is consumed by
+      the fold). Raises if the scale key doesn't correspond to a known
+      entry in *scales* -- that means the file shipped a scale we didn't
+      pre-register, which would silently desync the fold.
+    """
+
+    def _on_param(param_key: str, value: torch.Tensor) -> list[KeyValueOperationResult]:
+        scale = scales.get(param_key)
+        if scale is None:
+            return TRANSFORMER_LINEAR_DOWNCAST_MAP.apply_to_key_value(param_key, value)
+        scale = scale.to(device=value.device)
+        if scale.ndim != 0:
+            raise ValueError(f"Unsupported scale shape {tuple(scale.shape)} for {param_key}")
+        bf16 = (value.to(torch.float32) * scale).to(torch.bfloat16)
+        # Delegate the final fp8-vs-bf16 decision to the downcast map: Linears
+        # outside the fp8 subset (e.g. to_gate_logits) stay bf16 to match the
+        # plain nn.Linear that the upcast matcher leaves untouched.
+        return TRANSFORMER_LINEAR_DOWNCAST_MAP.apply_to_key_value(param_key, bf16)
+
+    def _drop_scale(scale_key: str, _value: torch.Tensor) -> list[KeyValueOperationResult]:
+        param_key = scale_key.removesuffix("_scale")
+        if param_key not in scales:
+            raise ValueError(
+                f"Scale key {scale_key!r} has no matching entry in the prequant scales dict; "
+                f"_read_scales and the loader's rename map have drifted"
+            )
+        return []
+
+    # Register the drop ops first so the dict-membership sanity check is the
+    # earliest sd-op that can fire on a scale key -- we crash on a stray scale
+    # before any silently mismatched fold has a chance to land in the state
+    # dict. Registration order is irrelevant for correctness (no overlap
+    # between matchers) but communicates intent.
+    return (
+        SDOps("FP8_CAST_PREQUANT_AWARE")
+        .with_kv_operation(key_suffix=".weight_scale", operation=_drop_scale)
+        .with_kv_operation(key_suffix=".bias_scale", operation=_drop_scale)
+        .with_kv_operation(key_suffix=".weight", operation=_on_param)
+        .with_kv_operation(key_suffix=".bias", operation=_on_param)
+    )
+
+
+def build_policy(checkpoint_path: str | Path) -> QuantizationPolicy:
+    """FP8 casting with upcasting during inference.
+    *checkpoint_path* is required (mirroring ``fp8_scaled_mm.build_policy``).
+    For prequantized fp8 checkpoints, sibling ``*_scale`` tensors (weight or
+    bias) are folded into the parent at load time.
+    """
+    scales = _read_scales(checkpoint_path)
+    return QuantizationPolicy(
+        sd_ops=_build_prequant_fold_sd_ops(scales),
+        module_ops=(UPCAST_DURING_INFERENCE,),
+        fuse_rule=fp8_cast_fuse_rule,
+    )

packages/ltx-core/src/ltx_core/quantization/fp8_scaled_mm.py

@@ -1,11 +1,23 @@
+import json
+import struct
 from typing import Callable
 
 import torch
 from torch import nn
 
+from ltx_core.loader.fuse_loras import FuseRule, bf16_fuse_rule
 from ltx_core.loader.module_ops import ModuleOps
-from ltx_core.loader.sd_ops import KeyValueOperationResult, SDOps
+from ltx_core.loader.primitives import StateDict
 from ltx_core.model.transformer import LTXModel
+from ltx_core.quantization.policy import QuantizationPolicy
+
+
+def _read_safetensors_dtypes(path: str) -> dict[str, str]:
+    """Return ``{tensor_name: dtype_string}`` from the safetensors header."""
+    with open(path, "rb") as f:
+        header_size = struct.unpack("<Q", f.read(8))[0]
+        header = json.loads(f.read(header_size).decode("utf-8"))
+    return {k: v["dtype"] for k, v in header.items() if k != "__metadata__"}
 
 
 class FP8Linear(nn.Module):
@@ -25,11 +37,8 @@ class FP8Linear(nn.Module):
         self.in_features = in_features
         self.out_features = out_features
 
-        fp8_shape = (in_features, out_features)
+        self.weight = nn.Parameter(torch.empty((out_features, in_features), dtype=torch.float8_e4m3fn, device=device))
-        self.weight = nn.Parameter(torch.empty(fp8_shape, dtype=torch.float8_e4m3fn, device=device))
-        # Weight scale for FP8 dequantization (shape matches checkpoint format)
         self.weight_scale = nn.Parameter(torch.empty((), dtype=torch.float32, device=device))
-        # Input scale for static quantization (pre-quantized checkpoints)
         self.input_scale = nn.Parameter(torch.empty((), dtype=torch.float32, device=device))
 
         if bias:
@@ -40,31 +49,25 @@ class FP8Linear(nn.Module):
     def forward(self, x: torch.Tensor) -> torch.Tensor:
         origin_shape = x.shape
 
-        # Static quantization: use pre-computed scale
+        # Clamp before cast: out-of-range values cast to NaN/saturated FP8, which
-        qinput, cur_input_scale = torch.ops.tensorrt_llm.static_quantize_e4m3_per_tensor(x, self.input_scale)
+        # produces black-screen output on some checkpoints (e.g. ltx-2-19b-dev-fp8).
-
+        fp8_min = torch.finfo(torch.float8_e4m3fn).min
-        # Flatten to 2D for matmul
+        fp8_max = torch.finfo(torch.float8_e4m3fn).max
+        qinput = torch.clamp(x * self.input_scale.reciprocal(), fp8_min, fp8_max).to(torch.float8_e4m3fn)
         if qinput.dim() == 3:
             qinput = qinput.reshape(-1, qinput.shape[-1])
-
+        output = torch._scaled_mm(
-        # FP8 scaled matmul
-        output = torch.ops.trtllm.cublas_scaled_mm(
             qinput,
-            self.weight,
+            self.weight.t(),
-            scale_a=cur_input_scale,
+            scale_a=self.input_scale,
             scale_b=self.weight_scale,
-            bias=None,
             out_dtype=x.dtype,
+            use_fast_accum=True,
         )
 
-        # Add bias
         if self.bias is not None:
-            bias = self.bias
+            output = output + self.bias.to(output.dtype)
-            if bias.dtype != output.dtype:
-                bias = bias.to(output.dtype)
-            output = output + bias
 
-        # Restore original shape
         if output.dim() != len(origin_shape):
             output_shape = list(origin_shape)
             output_shape[-1] = output.shape[-1]
@@ -74,15 +77,7 @@ class FP8Linear(nn.Module):
 
 
 def quantize_weight_to_fp8_per_tensor(weight: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
-    """
+    """Quantize a weight tensor to ``float8_e4m3fn`` with a per-tensor scale."""
-    Quantize a weight tensor to FP8 (float8_e4m3fn) using per-tensor scaling.
-    Args:
-        weight: The weight tensor to quantize (any dtype, will be cast to float32)
-    Returns:
-        Tuple of (quantized_weight, weight_scale):
-        - quantized_weight: FP8 tensor, transposed for cublas_scaled_mm
-        - weight_scale: Per-tensor scale factor (reciprocal of quantization scale)
-    """
     weight_fp32 = weight.to(torch.float32)
 
     fp8_min = torch.finfo(torch.float8_e4m3fn).min
@@ -96,7 +91,6 @@ def quantize_weight_to_fp8_per_tensor(weight: torch.Tensor) -> tuple[torch.Tenso
         weight_fp32: torch.Tensor, scale: torch.Tensor, fp8_min: torch.Tensor, fp8_max: torch.Tensor
     ) -> tuple[torch.Tensor, torch.Tensor]:
         quantized_weight = torch.clamp(weight_fp32 * scale, min=fp8_min, max=fp8_max).to(torch.float8_e4m3fn)
-        quantized_weight = quantized_weight.t()
         weight_scale = scale.reciprocal()
         return quantized_weight, weight_scale
 
@@ -104,36 +98,8 @@ def quantize_weight_to_fp8_per_tensor(weight: torch.Tensor) -> tuple[torch.Tenso
     return quantized_weight, weight_scale
 
 
-def _should_skip_layer(layer_name: str, excluded_layer_substrings: tuple[str, ...]) -> bool:
-    return any(substring in layer_name for substring in excluded_layer_substrings)
-
-
-EXCLUDED_LAYER_SUBSTRINGS = (
-    "patchify_proj",
-    "adaln_single",
-    "av_ca_video_scale_shift_adaln_single",
-    "av_ca_a2v_gate_adaln_single",
-    "caption_projection",
-    "proj_out",
-    "audio_patchify_proj",
-    "audio_adaln_single",
-    "av_ca_audio_scale_shift_adaln_single",
-    "av_ca_v2a_gate_adaln_single",
-    "audio_caption_projection",
-    "audio_proj_out",
-    "transformer_blocks.0.",
-    *[f"transformer_blocks.{i}." for i in range(43, 48)],
-)
-
-
 def _linear_to_fp8linear(layer: nn.Linear) -> FP8Linear:
-    """
+    """Create an ``FP8Linear`` matching the shape/bias of *layer*."""
-    Create an FP8Linear layer from an nn.Linear layer.
-    Args:
-        layer: The nn.Linear layer to convert (typically on meta device)
-    Returns:
-        A new FP8Linear with the same configuration
-    """
     return FP8Linear(
         in_features=layer.in_features,
         out_features=layer.out_features,
@@ -142,15 +108,14 @@ def _linear_to_fp8linear(layer: nn.Linear) -> FP8Linear:
     )
 
 
-def _apply_fp8_prepare_to_model(model: nn.Module, excluded_layer_substrings: tuple[str, ...]) -> nn.Module:
+def _swap_linears_to_fp8(model: nn.Module, should_swap: Callable[[str], bool]) -> nn.Module:
-    """Replace nn.Linear layers with FP8Linear in the module tree."""
+    """Replace nn.Linear layers with FP8Linear where ``should_swap(name)`` returns True."""
     replacements: list[tuple[nn.Module, str, nn.Linear]] = []
 
     for name, module in model.named_modules():
         if not isinstance(module, nn.Linear) or isinstance(module, FP8Linear):
             continue
-
+        if not should_swap(name):
-        if _should_skip_layer(name, excluded_layer_substrings):
             continue
 
         if "." in name:
@@ -168,40 +133,71 @@ def _apply_fp8_prepare_to_model(model: nn.Module, excluded_layer_substrings: tup
     return model
 
 
-def _create_transpose_kv_operation(
+def get_fp8_swap_module_ops(checkpoint_path: str) -> tuple[ModuleOps, ...]:
-    excluded_layer_substrings: tuple[str, ...],
+    """Return the FP8 swap ``ModuleOps`` for layers whose ``.weight`` is ``F8_E4M3``
-) -> Callable[[str, torch.Tensor], list[KeyValueOperationResult]]:
+    and which have a sibling ``.weight_scale`` tensor in the checkpoint.
-    def transpose_if_matches(key: str, value: torch.Tensor) -> list[KeyValueOperationResult]:
+    Raises ``ValueError`` if no such layers are found — that combination is ambiguous
-        # Only process .weight keys
+    (a BF16 checkpoint with this policy would load as a no-op).
-        if not key.endswith(".weight"):
+    """
-            return [KeyValueOperationResult(key, value)]
+    dtypes = _read_safetensors_dtypes(checkpoint_path)
+    fp8_scale_paths = frozenset(
+        key.removesuffix(".weight_scale")
+        for key in dtypes
+        if key.endswith(".weight_scale") and dtypes.get(key.removesuffix(".weight_scale") + ".weight") == "F8_E4M3"
+    )
+    if not fp8_scale_paths:
+        raise ValueError(
+            f"fp8_scaled_mm requires a pre-quantized checkpoint with F8_E4M3 .weight + .weight_scale "
+            f"tensors, but {checkpoint_path!r} has none. Use QuantizationPolicy.fp8_cast() for BF16 checkpoints."
+        )
 
-        # Only transpose 2D FP8 tensors (Linear weights)
+    def _should_swap(name: str) -> bool:
-        if value.dim() != 2 or value.dtype != torch.float8_e4m3fn:
+        suffix = "." + name
-            return [KeyValueOperationResult(key, value)]
+        return any(p == name or p.endswith(suffix) for p in fp8_scale_paths)
 
-        # Check if the layer is excluded
+    return (
-        layer_name = key.rsplit(".weight", 1)[0]
+        ModuleOps(
-        if _should_skip_layer(layer_name, excluded_layer_substrings):
+            name="fp8_swap_linears",
-            return [KeyValueOperationResult(key, value)]
-
-        # Transpose to cuBLAS layout (in, out)
-        transposed_weight = value.t()
-
-        return [KeyValueOperationResult(key, transposed_weight)]
-
-    return transpose_if_matches
-
-
-FP8_TRANSPOSE_SD_OPS = SDOps("fp8_transpose_weights").with_kv_operation(
-    _create_transpose_kv_operation(EXCLUDED_LAYER_SUBSTRINGS),
-    key_prefix="transformer_blocks.",
-    key_suffix=".weight",
-)
-
-
-FP8_PREPARE_MODULE_OPS = ModuleOps(
-    name="fp8_prepare_for_loading",
             matcher=lambda model: isinstance(model, LTXModel),
-    mutator=lambda model: _apply_fp8_prepare_to_model(model, EXCLUDED_LAYER_SUBSTRINGS),
+            mutator=lambda model: _swap_linears_to_fp8(model, _should_swap),
-)
+        ),
+    )
+
+
+def _fp8_scaled_mm_fuse(
+    key: str,
+    weight: torch.Tensor,
+    deltas: torch.Tensor,
+    model_sd: StateDict,
+) -> dict[str, torch.Tensor]:
+    """Dequantize via ``weight.float() * weight_scale``, add the BF16 delta,
+    and re-quantize to FP8 with a fresh per-tensor scale.
+    Layers that were not swapped to scaled FP8 (e.g. small embedder linears
+    excluded from the auto-discovered swap set) stay BF16 and have no
+    ``.weight_scale`` companion -- for those, fall back to a plain bf16 fuse.
+    """
+    scale_key = key.replace(".weight", ".weight_scale")
+    if scale_key not in model_sd.sd:
+        return bf16_fuse_rule(key, weight, deltas, model_sd)
+    weight_scale = model_sd.sd[scale_key]
+    original_weight = weight.to(torch.float32) * weight_scale
+    new_weight = original_weight + deltas.to(torch.float32)
+    new_fp8_weight, new_weight_scale = quantize_weight_to_fp8_per_tensor(new_weight)
+    return {key: new_fp8_weight, scale_key: new_weight_scale}
+
+
+fp8_scaled_mm_fuse_rule = FuseRule(aggregation_dtype=torch.bfloat16, fuse_fn=_fp8_scaled_mm_fuse)
+
+
+def build_policy(checkpoint_path: str) -> QuantizationPolicy:
+    """FP8 scaled matmul for checkpoints pre-quantized with per-tensor scales.
+    The set of layers to swap to ``FP8Linear`` is discovered from the
+    checkpoint's ``.weight_scale`` tensors via suffix-matching against the
+    model's named modules. Requires a pre-quantized checkpoint; for BF16
+    checkpoints, use :func:`ltx_core.quantization.fp8_cast.build_policy`.
+    """
+    return QuantizationPolicy(
+        sd_ops=None,
+        module_ops=get_fp8_swap_module_ops(checkpoint_path),
+        fuse_rule=fp8_scaled_mm_fuse_rule,
+    )

packages/ltx-core/src/ltx_core/quantization/policy.py

@@ -1,39 +1,24 @@
 from dataclasses import dataclass
 
+from ltx_core.loader.fuse_loras import FuseRule, bf16_fuse_rule
 from ltx_core.loader.module_ops import ModuleOps
 from ltx_core.loader.sd_ops import SDOps
-from ltx_core.quantization.fp8_cast import TRANSFORMER_LINEAR_DOWNCAST_MAP, UPCAST_DURING_INFERENCE
+from ltx_core.model.model_protocol import ModelConfigurator
-from ltx_core.quantization.fp8_scaled_mm import FP8_PREPARE_MODULE_OPS, FP8_TRANSPOSE_SD_OPS
+from ltx_core.model.transformer.model import LTXModel
 
 
 @dataclass(frozen=True)
 class QuantizationPolicy:
     """Configuration for model quantization during loading.
     Attributes:
-        sd_ops: State dict operations for weight transformation.
+        sd_ops: State-dict operations applied to each tensor during load.
-        module_ops: Post-load module transformations.
+        module_ops: Post-load module transformations applied to the meta model.
+        model_configurator: Configurator class to use when constructing the transformer.
+        fuse_rule: How LoRA deltas merge into this policy's weight layout.
+            Default ``bf16_fuse_rule`` is used when no policy is configured.
     """
 
     sd_ops: SDOps | None = None
     module_ops: tuple[ModuleOps, ...] = ()
-
+    model_configurator: type[ModelConfigurator[LTXModel]] | None = None
-    @classmethod
+    fuse_rule: FuseRule = bf16_fuse_rule
-    def fp8_cast(cls) -> "QuantizationPolicy":
-        """Create policy using FP8 casting with upcasting during inference."""
-        return cls(
-            sd_ops=TRANSFORMER_LINEAR_DOWNCAST_MAP,
-            module_ops=(UPCAST_DURING_INFERENCE,),
-        )
-
-    @classmethod
-    def fp8_scaled_mm(cls) -> "QuantizationPolicy":
-        """Create policy using FP8 scaled matrix multiplication."""
-        try:
-            import tensorrt_llm  # noqa: F401, PLC0415
-        except ImportError as e:
-            raise ImportError("tensorrt_llm is not installed, skipping FP8 scaled MM quantization") from e
-
-        return cls(
-            sd_ops=FP8_TRANSPOSE_SD_OPS,
-            module_ops=(FP8_PREPARE_MODULE_OPS,),
-        )

packages/ltx-core/src/ltx_core/text_encoders/gemma/embeddings_connector.py

@@ -18,7 +18,7 @@ class _BasicTransformerBlock1D(torch.nn.Module):
         dim: int,
         heads: int,
         dim_head: int,
-        rope_type: LTXRopeType = LTXRopeType.INTERLEAVED,
+        rope_type: LTXRopeType = LTXRopeType.SPLIT,
         apply_gated_attention: bool = False,
     ):
         super().__init__()
@@ -39,7 +39,7 @@ class _BasicTransformerBlock1D(torch.nn.Module):
     def forward(
         self,
         hidden_states: torch.Tensor,
-        attention_mask: torch.Tensor | None = None,
+        additive_attention_mask: torch.Tensor | None = None,
         pe: torch.Tensor | None = None,
     ) -> torch.Tensor:
         # Notice that normalization is always applied before the real computation in the following blocks.
@@ -49,8 +49,8 @@ class _BasicTransformerBlock1D(torch.nn.Module):
 
         norm_hidden_states = norm_hidden_states.squeeze(1)
 
-        # 2. Self-Attention
+        # 2. Self-Attention — `mask` is the kernel-boundary name for the additive mask.
-        attn_output = self.attn1(norm_hidden_states, mask=attention_mask, pe=pe)
+        attn_output = self.attn1(norm_hidden_states, mask=additive_attention_mask, pe=pe)
 
         hidden_states = attn_output + hidden_states
         if hidden_states.ndim == 4:
@@ -84,7 +84,7 @@ class Embeddings1DConnector(torch.nn.Module):
         causal_temporal_positioning (bool): If True, uses causal attention (default=False).
         num_learnable_registers (int | None): Number of learnable registers to replace padded tokens. If None, disables
             register replacement. (default=128)
-        rope_type (LTXRopeType): The RoPE variant to use (default=DEFAULT_ROPE_TYPE).
+        rope_type (LTXRopeType): The RoPE variant to use.
         double_precision_rope (bool): Use double precision rope calculation (default=False).
     """
 
@@ -99,7 +99,7 @@ class Embeddings1DConnector(torch.nn.Module):
         positional_embedding_max_pos: list[int] | None = None,
         causal_temporal_positioning: bool = False,
         num_learnable_registers: int | None = 128,
-        rope_type: LTXRopeType = LTXRopeType.INTERLEAVED,
+        rope_type: LTXRopeType = LTXRopeType.SPLIT,
         double_precision_rope: bool = False,
         apply_gated_attention: bool = False,
     ):
@@ -133,51 +133,40 @@ class Embeddings1DConnector(torch.nn.Module):
             )
 
     def _replace_padded_with_learnable_registers(
-        self, hidden_states: torch.Tensor, attention_mask: torch.Tensor
+        self, hidden_states: torch.Tensor, additive_attention_mask: torch.Tensor
     ) -> tuple[torch.Tensor, torch.Tensor]:
-        assert hidden_states.shape[1] % self.num_learnable_registers == 0, (
+        batch_size, seq_len, _ = hidden_states.shape
-            f"Hidden states sequence length {hidden_states.shape[1]} must be divisible by num_learnable_registers "
-            f"{self.num_learnable_registers}."
-        )
 
-        num_registers_duplications = hidden_states.shape[1] // self.num_learnable_registers
+        assert seq_len % self.num_learnable_registers == 0
-        learnable_registers = torch.tile(self.learnable_registers, (num_registers_duplications, 1))
-        attention_mask_binary = (attention_mask.squeeze(1).squeeze(1).unsqueeze(-1) >= -9000.0).int()
 
-        non_zero_hidden_states = hidden_states[:, attention_mask_binary.squeeze().bool(), :]
+        registers = self.learnable_registers.repeat(seq_len // self.num_learnable_registers, 1).to(hidden_states.dtype)
-        non_zero_nums = non_zero_hidden_states.shape[1]
+        registers = registers.unsqueeze(0).expand(batch_size, -1, -1)  # (B, seq_len, hidden_dim)
-        pad_length = hidden_states.shape[1] - non_zero_nums
+        binary_mask = additive_attention_mask[:, 0, 0, :].unsqueeze(-1) >= 0
-        adjusted_hidden_states = torch.nn.functional.pad(non_zero_hidden_states, pad=(0, 0, 0, pad_length), value=0)
+        binary_mask = binary_mask.to(hidden_states.dtype)
-        flipped_mask = torch.flip(attention_mask_binary, dims=[1])
+        hidden_states = binary_mask * hidden_states + (1 - binary_mask) * registers
-        hidden_states = flipped_mask * adjusted_hidden_states + (1 - flipped_mask) * learnable_registers
 
-        attention_mask = torch.full_like(
+        return hidden_states, torch.zeros_like(additive_attention_mask)
-            attention_mask,
-            0.0,
-            dtype=attention_mask.dtype,
-            device=attention_mask.device,
-        )
-
-        return hidden_states, attention_mask
 
     def forward(
         self,
         hidden_states: torch.Tensor,
-        attention_mask: torch.Tensor | None = None,
+        additive_attention_mask: torch.Tensor | None = None,
     ) -> tuple[torch.Tensor, torch.Tensor]:
-        """
+        """Forward pass of Embeddings1DConnector.
-        Forward pass of Embeddings1DConnector.
         Args:
-            hidden_states (torch.Tensor): Input tensor of embeddings (shape [batch, seq_len, feature_dim]).
+            hidden_states: (B, S, D) input embeddings.
-            attention_mask (torch.Tensor|None): Optional mask for valid tokens (shape compatible with hidden_states).
+            additive_attention_mask: optional additive mask of shape (B, 1, 1, S), where
+                valid = 0.0 and padding = -torch.finfo(dtype).max.
         Returns:
-            tuple[torch.Tensor, torch.Tensor]: Processed features and the corresponding (possibly modified) mask.
+            (hidden_states, additive_attention_mask)
         """
         if self.num_learnable_registers:
-            hidden_states, attention_mask = self._replace_padded_with_learnable_registers(hidden_states, attention_mask)
+            hidden_states, additive_attention_mask = self._replace_padded_with_learnable_registers(
+                hidden_states, additive_attention_mask
+            )
 
         indices_grid = torch.arange(hidden_states.shape[1], dtype=torch.float32, device=hidden_states.device)
-        indices_grid = indices_grid[None, None, :]
+        indices_grid = indices_grid[None, None, :].expand(hidden_states.shape[0], -1, -1)
         freq_grid_generator = generate_freq_grid_np if self.double_precision_rope else generate_freq_grid_pytorch
         freqs_cis = precompute_freqs_cis(
             indices_grid=indices_grid,
@@ -191,11 +180,11 @@ class Embeddings1DConnector(torch.nn.Module):
         )
 
         for block in self.transformer_1d_blocks:
-            hidden_states = block(hidden_states, attention_mask=attention_mask, pe=freqs_cis)
+            hidden_states = block(hidden_states, additive_attention_mask=additive_attention_mask, pe=freqs_cis)
 
         hidden_states = rms_norm(hidden_states)
 
-        return hidden_states, attention_mask
+        return hidden_states, additive_attention_mask
 
 
 class Embeddings1DConnectorConfigurator(ModelConfigurator[Embeddings1DConnector]):
@@ -204,7 +193,7 @@ class Embeddings1DConnectorConfigurator(ModelConfigurator[Embeddings1DConnector]
     @classmethod
     def from_config(cls: type[Embeddings1DConnector], config: dict) -> Embeddings1DConnector:
         transformer_config = config.get("transformer", {})
-        rope_type = LTXRopeType(transformer_config.get("rope_type", "interleaved"))
+        rope_type = LTXRopeType(transformer_config.get("rope_type", "split"))
         double_precision_rope = transformer_config.get("frequencies_precision", False) == "float64"
         pe_max_pos = transformer_config.get("connector_positional_embedding_max_pos", [1])
 
@@ -231,7 +220,7 @@ class AudioEmbeddings1DConnectorConfigurator(ModelConfigurator[Embeddings1DConne
     @classmethod
     def from_config(cls: type[Embeddings1DConnector], config: dict) -> Embeddings1DConnector:
         transformer_config = config.get("transformer", {})
-        rope_type = LTXRopeType(transformer_config.get("rope_type", "interleaved"))
+        rope_type = LTXRopeType(transformer_config.get("rope_type", "split"))
         double_precision_rope = transformer_config.get("frequencies_precision", False) == "float64"
         pe_max_pos = transformer_config.get("connector_positional_embedding_max_pos", [1])
 

packages/ltx-core/src/ltx_core/text_encoders/gemma/embeddings_processor.py

@@ -19,12 +19,32 @@ def convert_to_additive_mask(attention_mask: torch.Tensor, dtype: torch.dtype) -
     ) * torch.finfo(dtype).max
 
 
-def _to_binary_mask(encoded: torch.Tensor, encoded_mask: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
+def _compute_right_pad_order(additive_mask: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
-    """Convert connector output mask to binary mask and apply to encoded tensor."""
+    """Compute the index permutation that places valid tokens before pads in each row.
-    binary_mask = (encoded_mask < 0.000001).to(torch.int64)
+    Stable sort: valid tokens keep their relative order. Idempotent for inputs already
-    binary_mask = binary_mask.reshape([encoded.shape[0], encoded.shape[1], 1])
+    right-padded. The sort and reordered mask depend only on the mask, so they can be
-    encoded = encoded * binary_mask
+    computed once and reused across multiple feature tensors that share the mask.
-    return encoded, binary_mask
+    Args:
+        additive_mask: (B, 1, 1, S) additive mask, ``0.0`` for valid, ``-finfo.max`` for pad.
+    Returns:
+        ``(sort_idx, reordered_additive_mask)``: ``sort_idx`` is (B, S); the reordered mask
+        has the same shape as the input.
+    """
+    binary = (additive_mask[:, 0, 0, :] >= 0).to(torch.int32)  # (B, S)
+    sort_idx = torch.argsort(binary, dim=-1, descending=True, stable=True)  # (B, S)
+    new_binary = torch.gather(binary, 1, sort_idx)
+    new_additive = (new_binary.to(additive_mask.dtype) - 1) * torch.finfo(additive_mask.dtype).max
+    return sort_idx, new_additive[:, None, None, :]
+
+
+def _apply_right_pad_order(features: torch.Tensor, sort_idx: torch.Tensor) -> torch.Tensor:
+    """Apply a precomputed right-pad permutation (from ``_compute_right_pad_order``) to features."""
+    return torch.gather(features, 1, sort_idx.unsqueeze(-1).expand_as(features))
+
+
+def _to_binary_mask(encoded_mask: torch.Tensor, lead_shape: tuple[int, int]) -> torch.Tensor:
+    """Convert connector output mask to a binary (0/1) mask shaped ``(B, S, 1)`` for broadcasting."""
+    return (encoded_mask < 0.000001).to(torch.int64).reshape([lead_shape[0], lead_shape[1], 1])
 
 
 class EmbeddingsProcessor(nn.Module):
@@ -57,12 +77,19 @@ class EmbeddingsProcessor(nn.Module):
         if self.audio_connector is None and audio_features is not None:
             raise ValueError("Audio features were provided but no audio connector is configured.")
 
-        video_encoded, video_mask = self.video_connector(video_features, additive_attention_mask)
+        # Connectors expect right-padded input ([valid, pad]). Normalize layout here so the
-        video_encoded, binary_mask = _to_binary_mask(video_encoded, video_mask)
+        # upstream tokenizer can keep using either side without coupling to the connector.
+        # The sort index depends only on the mask, so compute it once and reuse for audio.
+        sort_idx, mask_for_connector = _compute_right_pad_order(additive_attention_mask)
+        video_features = _apply_right_pad_order(video_features, sort_idx)
+        video_encoded, video_mask = self.video_connector(video_features, mask_for_connector)
+        binary_mask = _to_binary_mask(video_mask, video_encoded.shape[:2])
+        video_encoded = video_encoded * binary_mask
 
         audio_encoded = None
         if self.audio_connector is not None:
-            audio_encoded, _ = self.audio_connector(audio_features, additive_attention_mask)
+            audio_features = _apply_right_pad_order(audio_features, sort_idx)
+            audio_encoded, _ = self.audio_connector(audio_features, mask_for_connector)
 
         return video_encoded, audio_encoded, binary_mask.squeeze(-1)
 

packages/ltx-core/src/ltx_core/text_encoders/gemma/encoders/base_encoder.py

@@ -30,21 +30,31 @@ class GemmaTextEncoder(torch.nn.Module):
 
     def encode(
         self,
-        text: str,
+        prompts: list[str],
         padding_side: str = "left",  # noqa: ARG002
-    ) -> tuple[tuple[torch.Tensor, ...], torch.Tensor]:
+    ) -> list[tuple[tuple[torch.Tensor, ...], torch.Tensor]]:
-        """Run Gemma LLM and return raw hidden states + attention mask.
+        """Run a single fused Gemma forward over a batch of prompts.
-        Calls the inner model (self.model.model) to skip lm_head logits computation (~500 MiB saving).
+        Calls the inner model (self.model.model) to skip lm_head logits computation
-        Returns:
+        (~500 MiB saving). The tokenizer pads every prompt to ``max_length`` (1024),
-            (hidden_states, attention_mask) where hidden_states is a tuple of per-layer tensors.
+        so the inputs stack into a single ``[N, 1024]`` batch with no further padding
+        logic; per-prompt outputs are sliced back to ``[1, 1024, D]`` / ``[1, 1024]``
+        in the original order.
         """
-        token_pairs = self.tokenizer.tokenize_with_weights(text)["gemma"]
+        if not prompts:
-        input_ids = torch.tensor([[t[0] for t in token_pairs]], device=self.model.device)
+            return []
-        attention_mask = torch.tensor([[w[1] for w in token_pairs]], device=self.model.device)
+        tokenized = [self.tokenizer.tokenize_with_weights(t)["gemma"] for t in prompts]
+        input_ids = torch.tensor(
+            [[tok for tok, _ in pairs] for pairs in tokenized],
+            device=self.model.device,
+        )
+        attention_mask = torch.tensor(
+            [[w for _, w in pairs] for pairs in tokenized],
+            device=self.model.device,
+        )
         outputs = self.model.model(input_ids=input_ids, attention_mask=attention_mask, output_hidden_states=True)
         hidden_states = outputs.hidden_states
         del outputs
-        return hidden_states, attention_mask
+        return [(tuple(h[i : i + 1] for h in hidden_states), attention_mask[i : i + 1]) for i in range(len(prompts))]
 
     # --- Prompt enhancement methods ---
 
@@ -65,7 +75,9 @@ class GemmaTextEncoder(torch.nn.Module):
         pad_token_id = self.processor.tokenizer.pad_token_id if self.processor.tokenizer.pad_token_id is not None else 0
         model_inputs = _pad_inputs_for_attention_alignment(model_inputs, pad_token_id=pad_token_id)
 
-        with torch.inference_mode(), torch.random.fork_rng(devices=[self.model.device]):
+        # fork_rng device pinning is only supported for CUDA; MPS/CPU fork CPU RNG only.
+        fork_devices = [self.model.device] if self.model.device.type == "cuda" else []
+        with torch.inference_mode(), torch.random.fork_rng(devices=fork_devices):
             torch.manual_seed(seed)
             outputs = self.model.generate(
                 **model_inputs,
@@ -183,7 +195,7 @@ def module_ops_from_gemma_root(gemma_root: str) -> tuple[ModuleOps, ...]:
         return module
 
     def load_processor(module: GemmaTextEncoder) -> GemmaTextEncoder:
-        image_processor = AutoImageProcessor.from_pretrained(processor_root, local_files_only=True)
+        image_processor = AutoImageProcessor.from_pretrained(processor_root, local_files_only=True, use_fast=False)
         if not module.tokenizer:
             raise ValueError("Tokenizer model operation must be performed before processor model operation")
         module.processor = Gemma3Processor(image_processor=image_processor, tokenizer=module.tokenizer.tokenizer)

packages/ltx-core/src/ltx_core/text_encoders/gemma/encoders/encoder_configurator.py

@@ -1,4 +1,5 @@
 import torch
+import transformers
 from transformers import Gemma3Config
 from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS
 from transformers.models.gemma3 import Gemma3ForConditionalGeneration
@@ -19,6 +20,8 @@ from ltx_core.text_encoders.gemma.feature_extractor import (
     FeatureExtractorV2,
 )
 
+_TRANSFORMERS_V5: bool = int(transformers.__version__.split(".", 1)[0]) >= 5
+
 
 class GemmaTextEncoderConfigurator(ModelConfigurator[GemmaTextEncoder]):
     @classmethod
@@ -100,14 +103,30 @@ def _create_feature_extractor(transformer_config: dict) -> torch.nn.Module:
 
 # --- Split SDOps: Gemma LLM keys vs Embeddings Processor keys ---
 
-GEMMA_LLM_KEY_OPS = (
+
+def _build_gemma_llm_key_ops(*, transformers_v5: bool) -> SDOps:
+    """Build the checkpoint-key remapping for the Gemma multimodal encoder.
+    The vision-tower mapping differs between transformers <5 and >=5 because
+    upstream PR https://github.com/huggingface/transformers/pull/39847 flattened
+    ``Gemma3ForConditionalGeneration.model.vision_tower.vision_model`` into
+    ``model.vision_tower``. Checkpoints continue to ship the legacy
+    ``vision_tower.vision_model.*`` prefix, so we strip the inner ``vision_model.``
+    when targeting v5 and pass it through unchanged for v4.
+    """
+    base = (
         SDOps("GEMMA_LLM_KEY_OPS")
         # 1. Map language model layers (note the double .model prefix)
         .with_matching(prefix="language_model.model.")
         .with_replacement("language_model.model.", "model.model.language_model.")
-    # 2. Map the Vision Tower
+        # 2. Map the Vision Tower (version-dependent — see docstring)
         .with_matching(prefix="vision_tower.")
-    .with_replacement("vision_tower.", "model.model.vision_tower.")
+    )
+    if transformers_v5:
+        base = base.with_replacement("vision_tower.vision_model.", "model.model.vision_tower.")
+    else:
+        base = base.with_replacement("vision_tower.", "model.model.vision_tower.")
+    return (
+        base
         # 3. Map the Multi-Modal Projector
         .with_matching(prefix="multi_modal_projector.")
         .with_replacement("multi_modal_projector.", "model.model.multi_modal_projector.")
@@ -119,7 +138,10 @@ GEMMA_LLM_KEY_OPS = (
             ],
             key_prefix="model.model.language_model.embed_tokens.weight",
         )
-)
+    )
+
+
+GEMMA_LLM_KEY_OPS = _build_gemma_llm_key_ops(transformers_v5=_TRANSFORMERS_V5)
 
 EMBEDDINGS_PROCESSOR_KEY_OPS = (
     SDOps("EMBEDDINGS_PROCESSOR_KEY_OPS")
@@ -152,24 +174,85 @@ VIDEO_ONLY_EMBEDDINGS_PROCESSOR_KEY_OPS = (
 )
 
 
+def _resolve_local_base_freq(config: object) -> float:
+    rope_parameters = getattr(config, "rope_parameters", None)
+    if isinstance(rope_parameters, dict) and "sliding_attention" in rope_parameters:
+        sliding = rope_parameters["sliding_attention"]
+        if isinstance(sliding, dict) and "rope_theta" in sliding:
+            return float(sliding["rope_theta"])
+    if hasattr(config, "rope_local_base_freq"):
+        return float(config.rope_local_base_freq)
+    raise AttributeError(
+        "Gemma text_config exposes neither rope_local_base_freq nor rope_parameters['sliding_attention']['rope_theta']"
+    )
+
+
+def _resolve_full_rope_type(config: object) -> str:
+    rope_parameters = getattr(config, "rope_parameters", None)
+    if isinstance(rope_parameters, dict) and "full_attention" in rope_parameters:
+        full = rope_parameters["full_attention"]
+        if isinstance(full, dict) and "rope_type" in full:
+            return str(full["rope_type"])
+    rope_scaling = getattr(config, "rope_scaling", None)
+    if rope_scaling is not None:
+        if isinstance(rope_scaling, dict):
+            if "rope_type" in rope_scaling:
+                return str(rope_scaling["rope_type"])
+        elif hasattr(rope_scaling, "rope_type"):
+            return str(rope_scaling.rope_type)
+    raise AttributeError(
+        "Gemma text_config exposes neither rope_scaling.rope_type nor rope_parameters['full_attention']['rope_type']"
+    )
+
+
+def _populate_rotary_v4(l_model: torch.nn.Module, config: object) -> None:
+    """transformers <5 layout: separate ``rotary_emb_local`` + ``rotary_emb``."""
+    dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
+    base = _resolve_local_base_freq(config)
+    local_inv = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.int64).to(dtype=torch.float) / dim))
+    full_inv, _ = ROPE_INIT_FUNCTIONS[_resolve_full_rope_type(config)](config)
+    l_model.rotary_emb_local.register_buffer("inv_freq", local_inv)
+    l_model.rotary_emb.register_buffer("inv_freq", full_inv)
+
+
+def _populate_rotary_v5(l_model: torch.nn.Module, config: object) -> None:
+    """transformers >=5 layout: single ``rotary_emb`` with per-layer-type buffers.
+    Mirrors ``Gemma3PreTrainedModel._init_weights`` for ``Gemma3RotaryEmbedding``
+    so meta-built models reach the same numerical state as a from_pretrained load.
+    """
+    rope_emb = l_model.rotary_emb
+    for layer_type in dict.fromkeys(config.layer_types):
+        rope_params = config.rope_parameters[layer_type]
+        if rope_params is None:
+            continue
+        rope_type = rope_params["rope_type"]
+        if rope_type == "default":
+            inv_freq, attn_scaling = rope_emb.compute_default_rope_parameters(config, layer_type=layer_type)
+        else:
+            inv_freq, attn_scaling = ROPE_INIT_FUNCTIONS[rope_type](config, layer_type=layer_type)
+        rope_emb.register_buffer(f"{layer_type}_inv_freq", inv_freq, persistent=False)
+        rope_emb.register_buffer(f"{layer_type}_original_inv_freq", inv_freq.clone(), persistent=False)
+        setattr(rope_emb, f"{layer_type}_attention_scaling", attn_scaling)
+
+
 def create_and_populate(module: GemmaTextEncoder) -> GemmaTextEncoder:
     model = module.model
-    v_model = model.model.vision_tower.vision_model
+    v_tower = model.model.vision_tower
+    v_model = v_tower.vision_model if hasattr(v_tower, "vision_model") else v_tower
     l_model = model.model.language_model
 
     config = model.config.text_config
-    dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
+
-    base = config.rope_local_base_freq
+    if hasattr(l_model, "rotary_emb_local"):
-    local_rope_freqs = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.int64).to(dtype=torch.float) / dim))
+        _populate_rotary_v4(l_model, config)
-    inv_freqs, _ = ROPE_INIT_FUNCTIONS[config.rope_scaling["rope_type"]](config)
+    else:
+        _populate_rotary_v5(l_model, config)
 
     positions_length = len(v_model.embeddings.position_ids[0])
     position_ids = torch.arange(positions_length, dtype=torch.long, device="cpu").unsqueeze(0)
     v_model.embeddings.register_buffer("position_ids", position_ids)
-    embed_scale = torch.tensor(model.config.text_config.hidden_size**0.5, device="cpu")
+    embed_scale = torch.tensor(config.hidden_size**0.5, device="cpu")
     l_model.embed_tokens.register_buffer("embed_scale", embed_scale)
-    l_model.rotary_emb_local.register_buffer("inv_freq", local_rope_freqs)
-    l_model.rotary_emb.register_buffer("inv_freq", inv_freqs)
 
     return module
 

packages/ltx-core/src/ltx_core/text_encoders/gemma/feature_extractor.py

@@ -11,37 +11,25 @@ from torch import nn
 
 def _norm_and_concat_padded_batch(
     encoded_text: torch.Tensor,
-    sequence_lengths: torch.Tensor,
+    attention_mask: torch.Tensor,
-    padding_side: str = "right",
 ) -> torch.Tensor:
     """Normalize and flatten multi-layer hidden states, respecting padding.
     Performs per-batch, per-layer normalization using masked mean and range,
-    then concatenates across the layer dimension.
+    then concatenates across the layer dimension. Padding-side agnostic: the
+    binary ``attention_mask`` already encodes which positions are valid.
     Args:
         encoded_text: Hidden states of shape [batch, seq_len, hidden_dim, num_layers].
-        sequence_lengths: Number of valid (non-padded) tokens per batch item.
+        attention_mask: Binary mask of shape [batch, seq_len], 1 for valid tokens, 0 for padding.
-        padding_side: Whether padding is on "left" or "right".
     Returns:
         Normalized tensor of shape [batch, seq_len, hidden_dim * num_layers],
         with padded positions zeroed out.
     """
-    b, t, d, l = encoded_text.shape  # noqa: E741
+    b, _, d, l = encoded_text.shape  # noqa: E741
-    device = encoded_text.device
-
-    token_indices = torch.arange(t, device=device)[None, :]
-
-    if padding_side == "right":
-        mask = token_indices < sequence_lengths[:, None]
-    elif padding_side == "left":
-        start_indices = t - sequence_lengths[:, None]
-        mask = token_indices >= start_indices
-    else:
-        raise ValueError(f"padding_side must be 'left' or 'right', got {padding_side}")
-
-    mask = rearrange(mask, "b t -> b t 1 1")
-
     eps = 1e-6
 
+    sequence_lengths = attention_mask.sum(dim=-1)
+    mask = rearrange(attention_mask.bool(), "b t -> b t 1 1")
+
     masked = encoded_text.masked_fill(~mask, 0.0)
     denom = (sequence_lengths * d).view(b, 1, 1, 1)
     mean = masked.sum(dim=(1, 2), keepdim=True) / (denom + eps)
@@ -51,12 +39,10 @@ def _norm_and_concat_padded_batch(
     range_ = x_max - x_min
 
     normed = 8 * (encoded_text - mean) / (range_ + eps)
-    normed = normed.reshape(b, t, -1)
+    normed = normed.reshape(b, -1, d * l)
 
     mask_flattened = rearrange(mask, "b t 1 1 -> b t 1").expand(-1, -1, d * l)
-    normed = normed.masked_fill(~mask_flattened, 0.0)
+    return normed.masked_fill(~mask_flattened, 0.0)
-
-    return normed
 
 
 def norm_and_concat_per_token_rms(
@@ -97,12 +83,14 @@ class FeatureExtractorV1(nn.Module):
         self.is_av = is_av
 
     def forward(
-        self, hidden_states: torch.Tensor, attention_mask: torch.Tensor, padding_side: str = "left"
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor,
+        padding_side: str = "left",  # noqa: ARG002 — kept for API stability; norm is layout-agnostic
     ) -> tuple[torch.Tensor, torch.Tensor | None]:
         encoded = torch.stack(hidden_states, dim=-1) if isinstance(hidden_states, (list, tuple)) else hidden_states
         dtype = encoded.dtype
-        sequence_lengths = attention_mask.sum(dim=-1)
+        normed = _norm_and_concat_padded_batch(encoded, attention_mask)
-        normed = _norm_and_concat_padded_batch(encoded, sequence_lengths, padding_side)
         features = self.aggregate_embed(normed.to(dtype))
         if self.is_av:
             return features, features

packages/ltx-core/src/ltx_core/text_encoders/gemma/tokenizer.py

@@ -1,6 +1,13 @@
+from enum import Enum
+
 from transformers import AutoTokenizer
 
 
+class PaddingSide(str, Enum):
+    LEFT = "left"
+    RIGHT = "right"
+
+
 class LTXVGemmaTokenizer:
     """
     Tokenizer wrapper for Gemma models compatible with LTXV processes.
@@ -8,18 +15,18 @@ class LTXVGemmaTokenizer:
     ensuring correct settings and output formatting for downstream consumption.
     """
 
-    def __init__(self, tokenizer_path: str, max_length: int = 256):
+    def __init__(self, tokenizer_path: str, max_length: int = 256, padding_side: PaddingSide = PaddingSide.LEFT):
         """
         Initialize the tokenizer.
         Args:
             tokenizer_path (str): Path to the pretrained tokenizer files or model directory.
             max_length (int, optional): Max sequence length for encoding. Defaults to 256.
+            padding_side (PaddingSide, optional): Side to pad on. Defaults to ``PaddingSide.LEFT``.
         """
         self.tokenizer = AutoTokenizer.from_pretrained(
             tokenizer_path, local_files_only=True, model_max_length=max_length
         )
-        # Gemma expects left padding for chat-style prompts; for plain text it doesn't matter much.
+        self.tokenizer.padding_side = padding_side.value
-        self.tokenizer.padding_side = "left"
         if self.tokenizer.pad_token is None:
             self.tokenizer.pad_token = self.tokenizer.eos_token
 

packages/ltx-core/src/ltx_core/tiling.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 import itertools
+import math
 from dataclasses import dataclass, replace
 from typing import Callable, NamedTuple
 
@@ -462,3 +463,16 @@ class TileCountConfig:
     frames: DimensionTilingConfig = DimensionTilingConfig(num_tiles=1, overlap=0)
     height: DimensionTilingConfig = DimensionTilingConfig(num_tiles=1, overlap=0)
     width: DimensionTilingConfig = DimensionTilingConfig(num_tiles=1, overlap=0)
+
+
+def balanced_tile_split(num_tiles: int) -> tuple[int, int]:
+    """Factor ``num_tiles`` into ``(small, large)`` as square as possible.
+    ``small`` is the largest divisor not exceeding the square root, so
+    ``small * large == num_tiles`` and ``small <= large``. E.g. 2 -> (1, 2),
+    4 -> (2, 2), 8 -> (2, 4), 16 -> (4, 4). The caller decides which tiled
+    dimension gets which factor.
+    """
+    if num_tiles < 1:
+        raise ValueError(f"num_tiles must be >= 1, got {num_tiles}")
+    small = next(d for d in range(math.isqrt(num_tiles), 0, -1) if num_tiles % d == 0)
+    return small, num_tiles // small

packages/ltx-core/src/ltx_core/tools.py

@@ -75,6 +75,9 @@ class LatentTools(Protocol):
         clean_latent = latent_state.clean_latent[:, :num_tokens]
         denoise_mask = torch.ones_like(latent_state.denoise_mask)[:, :num_tokens]
         positions = latent_state.positions[:, :, :num_tokens]
+        cond_channels = (
+            latent_state.cond_channels[:, :num_tokens] if latent_state.cond_channels is not None else None
+        )
 
         return LatentState(
             latent=latent,
@@ -82,6 +85,7 @@ class LatentTools(Protocol):
             positions=positions,
             clean_latent=clean_latent,
             attention_mask=None,
+            cond_channels=cond_channels,
         )
 
 
@@ -138,7 +142,7 @@ class VideoLatentTools(LatentTools):
             LatentState(
                 latent=initial_latent,
                 denoise_mask=denoise_mask,
-                positions=positions.to(dtype),
+                positions=positions,
                 clean_latent=clean_latent,
             )
         )

packages/ltx-core/src/ltx_core/types.py

@@ -20,15 +20,17 @@ class SpatioTemporalScaleFactors(NamedTuple):
     """
     Describes the spatiotemporal downscaling between decoded video space and
     the corresponding VAE latent grid.
+    Field order matches the (frame/time, height, width) axis layout used by
+    latent tensors and meshgrid coordinates elsewhere in the codebase.
     """
 
     time: int
-    width: int
     height: int
+    width: int
 
     @classmethod
     def default(cls) -> "SpatioTemporalScaleFactors":
-        return cls(time=8, width=32, height=32)
+        return cls(time=8, height=32, width=32)
 
 
 VIDEO_SCALE_FACTORS = SpatioTemporalScaleFactors.default()
@@ -74,9 +76,9 @@ class VideoLatentShape(NamedTuple):
         latent_channels: int = 128,
         scale_factors: SpatioTemporalScaleFactors = VIDEO_SCALE_FACTORS,
     ) -> "VideoLatentShape":
-        frames = (shape.frames - 1) // scale_factors[0] + 1
+        frames = (shape.frames - 1) // scale_factors.time + 1
-        height = shape.height // scale_factors[1]
+        height = shape.height // scale_factors.height
-        width = shape.width // scale_factors[2]
+        width = shape.width // scale_factors.width
 
         return VideoLatentShape(
             batch=shape.batch,
@@ -191,6 +193,11 @@ class LatentState:
         clean_latent: Initial state of the latent before denoising, may include conditioning latents.
         attention_mask: Optional 2D self-attention mask of shape (B, T, T). Values in [0, 1] where 1 = full attention,
             0 = no attention. None means full attention everywhere. Built incrementally by conditioning items.
+        cond_channels: Optional in-context conditioning channels in patchified token space, shape (B, T, C_cond),
+            aligned token-for-token with ``latent``. These extra per-token feature channels (e.g. SCAIL-2 mask
+            channels) are concatenated onto the latent right before the model's first projection; they are never
+            noised or denoised. ``None`` means no conditioning channels (standard models). When present, the token
+            length T must always match ``latent``; token-appending conditioning items extend it with zeros.
     """
 
     latent: torch.Tensor
@@ -198,6 +205,7 @@ class LatentState:
     positions: torch.Tensor
     clean_latent: torch.Tensor
     attention_mask: torch.Tensor | None = None
+    cond_channels: torch.Tensor | None = None
 
     def clone(self) -> "LatentState":
         return LatentState(
@@ -206,4 +214,5 @@ class LatentState:
             positions=self.positions.clone(),
             clean_latent=self.clean_latent.clone(),
             attention_mask=self.attention_mask.clone() if self.attention_mask is not None else None,
+            cond_channels=self.cond_channels.clone() if self.cond_channels is not None else None,
         )