Automated PR - 2026-03-30

This commit is contained in:
github-actions[bot]
2026-03-30 17:59:34 +00:00
parent ae855f8538
commit f4d0c1ec0e
48 changed files with 8429 additions and 6644 deletions
@@ -133,6 +133,7 @@ class OptimizationConfig(ConfigBaseModel):
"cosine",
"cosine_with_restarts",
"polynomial",
"step",
] = Field(
default="linear",
description="Type of scheduler to use for training",
@@ -398,6 +399,21 @@ class CheckpointsConfig(ConfigBaseModel):
description="Precision to use when saving checkpoint weights. Options: 'bfloat16' or 'float32'.",
)
no_resume: bool = Field(
default=False,
description="When True, ignore any saved training state and start from step 0. "
"Model weights from load_checkpoint are still loaded, but optimizer/scheduler "
"state and step counter are reset.",
)
save_training_state: Literal["full", "minimal", "off"] = Field(
default="minimal",
description="Save training state alongside checkpoints for resume. "
"'full': optimizer + scheduler + RNG + step (~800MB for LoRA, much larger for full fine-tuning). "
"'minimal': scheduler + RNG + step only (~few KB, sufficient for LoRA). "
"'off': nothing saved, resume not possible.",
)
class HubConfig(ConfigBaseModel):
"""Configuration for Hugging Face Hub integration"""