Skip to content

Fix PEFT forward dtype cast corrupting 4-bit quantized model inputs #482

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
danielhanchen merged 1 commit into from
Feb 10, 2026
Merged

Fix PEFT forward dtype cast corrupting 4-bit quantized model inputs #482

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
3 changes: 3 additions & 0 deletions unsloth_zoo/compiler.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2395,6 +2395,8 @@ def patch_lora_forwards(torch_compile_options):
# When autocast is disabled and base_layer has float32 weights,
# cast x to match the weight dtype to prevent dtype mismatch.
# For 8-bit layers, the base_layer call was already replaced above.
# For 4-bit layers, weight.dtype is uint8 (packed quantized bytes),
# so we must skip the cast to avoid corrupting input values.
_base_layer_call = "result = self.base_layer(x, *args, **kwargs)"
_m = re.search(r'^( *)' + re.escape(_base_layer_call), source, re.MULTILINE)
if _m:
Expand All @@ -2403,6 +2405,7 @@ def patch_lora_forwards(torch_compile_options):
_base_layer_call,
f"if not torch.is_autocast_enabled() and hasattr(self.base_layer, 'weight') "
f"and self.base_layer.weight is not None "
f"and not hasattr(self.base_layer.weight, 'quant_state') "
f"and x.dtype != self.base_layer.weight.dtype:\n"
f"{_ind} x = x.to(self.base_layer.weight.dtype)\n"
f"{_ind}{_base_layer_call}",
Expand Down