Skip to content

Fix LoRA forward returning wrong dtype when autocast is disabled #564

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
Mar 27, 2026
Merged

Fix LoRA forward returning wrong dtype when autocast is disabled #564

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
10 changes: 9 additions & 1 deletion unsloth_zoo/compiler.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2491,7 +2491,15 @@ def patch_lora_forwards(torch_compile_options):
if (old1 not in source and add not in source) and (old2 not in source):
pass
else:
replace = "return lora_forward(result, lora_A, lora_B, dropout, x, scaling)"
# Linear/GPTQ/LoraParallel reassign result to float32 before the
# loop, so they save the original dtype in torch_result_dtype.
# Linear4bit/Linear8bitLt only cast x, leaving result untouched,
# so result.dtype is still the base-layer dtype at return time.
if re.search(r"\btorch_result_dtype\s*=\s*result\.dtype\b", source):
dtype_cast = "torch_result_dtype"
else:
dtype_cast = "result.dtype"
replace = f"return lora_forward(result, lora_A, lora_B, dropout, x, scaling).to({dtype_cast})"
source = source.replace(old1, replace)
source = source.replace(old2, replace)
pass
Expand Down
Loading