Skip to content

fix: model dtype is not same as lora dtype in FSDP train #183

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
tastelikefeet merged 2 commits into from
Apr 24, 2026
+17 −0
Merged

fix: model dtype is not same as lora dtype in FSDP train #183

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
c433015
fix: model dtype not same as lora dtype in FSDP train
0hujun Apr 23, 2026
c120c06
fix: model dtype is not same as lora dtype in FSDP train
0hujun Apr 23, 2026
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
17 changes: 17 additions & 0 deletions src/twinkle/model/transformers/transformers.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -963,6 +963,22 @@ def _load_optimizer(self, checkpoint_dir, **kwargs):
state_dict = torch.load(scheduler_path, map_location='cpu')
optimizer_config.lr_scheduler.load_state_dict(state_dict)

def _ensure_lora_dtype(self, model):
"""Force LoRA parameters to use the same dtype as base model for FSDP2 compatibility."""
base_dtype = None
for param in model.parameters():
if param.dtype in (torch.float16, torch.bfloat16, torch.float32):
base_dtype = param.dtype
break
if base_dtype is None:
return

# Convert all LoRA parameters to the base model dtype
with torch.no_grad():
for name, param in model.named_parameters():
if 'lora_' in name.lower() and param.dtype != base_dtype:
param.data = param.data.to(base_dtype)
Comment thread
0hujun marked this conversation as resolved.

@remote_function(collect='first')
def get_state_dict(self, **kwargs):
return self._get_trainable_parameters(kwargs.pop('adapter_name', self._get_default_group()))
Expand Down Expand Up @@ -1019,6 +1035,7 @@ def _patch_adapter(self, adapter_name: str, config_or_dir: Union[PeftConfig, str
else:
unwrapped_model.add_adapter(adapter_name, config)

self._ensure_lora_dtype(self.model)
self.optimizer_group[adapter_name] = self._construct_default_optimizer_group()
self.optimizer_group[adapter_name].adapter_name = adapter_name
self.optimizer_group[adapter_name].adapter_config = config
Expand Down
Loading