Automated PR - 2026-01-12

packages/ltx-trainer/configs/ltx2_av_lora.yaml

@@ -253,6 +253,10 @@ checkpoints:
   # Set to -1 to keep all checkpoints
   keep_last_n: -1
 
+  # Precision to use when saving checkpoint weights
+  # Options: "bfloat16" (default, smaller files) or "float32" (full precision)
+  precision: "bfloat16"
+
 # -----------------------------------------------------------------------------
 # Flow Matching Configuration
 # -----------------------------------------------------------------------------

packages/ltx-trainer/configs/ltx2_v2v_ic_lora.yaml

@@ -263,6 +263,10 @@ checkpoints:
   # Set to -1 to keep all checkpoints
   keep_last_n: 3
 
+  # Precision to use when saving checkpoint weights
+  # Options: "bfloat16" (default, smaller files) or "float32" (full precision)
+  precision: "bfloat16"
+
 # -----------------------------------------------------------------------------
 # Flow Matching Configuration
 # -----------------------------------------------------------------------------

packages/ltx-trainer/docs/configuration-reference.md

@@ -292,8 +292,9 @@ Model checkpointing configuration.
 
 ```yaml
 checkpoints:
-  interval: 250   # Steps between checkpoint saves (null = disabled)
-  keep_last_n: 3  # Number of recent checkpoints to retain
+  interval: 250       # Steps between checkpoint saves (null = disabled)
+  keep_last_n: 3      # Number of recent checkpoints to retain
+  precision: bfloat16 # Precision for saved weights (bfloat16 or float32)
 ```
 
 **Key parameters:**
@@ -302,6 +303,7 @@ checkpoints:
 |---------------|------------------------------------------------------------------------|
 | `interval`    | Steps between intermediate checkpoint saves (set to `null` to disable) |
 | `keep_last_n` | Number of most recent checkpoints to keep (-1 = keep all)              |
+| `precision`   | Precision for saved checkpoint weights: `"bfloat16"` (default) or `"float32"` |
 
 ### HubConfig
 

packages/ltx-trainer/src/ltx_trainer/config.py

@@ -350,6 +350,11 @@ class CheckpointsConfig(ConfigBaseModel):
         ge=-1,
     )
 
+    precision: Literal["bfloat16", "float32"] = Field(
+        default="bfloat16",
+        description="Precision to use when saving checkpoint weights. Options: 'bfloat16' or 'float32'.",
+    )
+
 
 class HubConfig(ConfigBaseModel):
     """Configuration for Hugging Face Hub integration"""

packages/ltx-trainer/src/ltx_trainer/trainer.py

@@ -873,6 +873,9 @@ class LtxvTrainer:
 
         save_dir.mkdir(exist_ok=True, parents=True)
 
+        # Determine save precision
+        save_dtype = torch.bfloat16 if self._config.checkpoints.precision == "bfloat16" else torch.float32
+
         # For LoRA: extract only adapter weights; for full: use as-is
         if is_lora:
             unwrapped = self._accelerator.unwrap_model(self._transformer, keep_torch_compile=False)
@@ -885,9 +888,15 @@ class LtxvTrainer:
             # Convert to ComfyUI-compatible format (add "diffusion_model." prefix)
             state_dict = {f"diffusion_model.{k}": v for k, v in state_dict.items()}
 
+            # Cast to configured precision
+            state_dict = {k: v.to(save_dtype) if isinstance(v, Tensor) else v for k, v in state_dict.items()}
+
             # Save to disk
             save_file(state_dict, saved_weights_path)
         else:
+            # Cast to configured precision
+            full_state_dict = {k: v.to(save_dtype) if isinstance(v, Tensor) else v for k, v in full_state_dict.items()}
+
             # Save to disk
             self._accelerator.save(full_state_dict, saved_weights_path)