3b16e317b8
The training UI's 'Standard' option sends the string "true". The worker converted "false"/"none"/"" to False but passed "true" through as a string. unsloth's _configure_gradient_checkpointing only unpatches the smart offloader on the (True, False) boolean branch; the string "true" matches neither that nor "unsloth", so it returns without unpatching and the offloader stays active. On memory-constrained / unified-memory GPUs that offload then OOM-crashes even though the user asked for standard checkpointing (reported on gfx1201 R9700, applies to Strix Halo too). Normalize "true"/"1"/"yes" -> True so unsloth gets a real bool and unpatches, matching the existing normalization in trainer.py. "unsloth" and "mlx" stay strings; booleans pass through unchanged. |
||
|---|---|---|
| .. | ||
| data_recipe | ||
| export | ||
| inference | ||
| rag | ||
| training | ||
| __init__.py | ||
| _torchao_stub.py | ||
| import_guards.py | ||
| tool_healing.py | ||