Skip to content

Fix FSDP2 model key miss match when sync LoRA model to vllm server #4603

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

Open
wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Fix FSDP2 model key miss match when sync LoRA model to vllm server #4603

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
f5a2b51
Skip lora adapters and remove the FSDPPEFT model prefix when sync to …
Xiao-Chenguang Nov 29, 2025
3e07550
Apply FSDP2+LoRA+vLLM fix to dpo and rloo as well
Xiao-Chenguang Nov 29, 2025
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
5 changes: 5 additions & 0 deletions trl/experimental/online_dpo/online_dpo_trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -933,6 +933,11 @@ def _sync_fsdp1_params_to_vllm(self, module: nn.Module, prefix: str = "", visite
def _sync_fsdp2_params_to_vllm(self, module: nn.Module):
# For FSDP2, module.state_dict() already covers all parameters, so no need for recursion
for name, param in module.state_dict().items():
# Fix FSDP2 + LoRA parameter naming issue with vLLM
if 'lora_' in name:
continue # skip LoRA parameters
name = name.removeprefix("base_model.model.").replace(".base_layer", "")

if param.is_cpu:
param = param.to(torch.device("cuda"))
param = param.full_tensor()
Expand Down
5 changes: 5 additions & 0 deletions trl/trainer/grpo_trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -968,6 +968,11 @@ def _sync_fsdp1_params_to_vllm(self, module: nn.Module, prefix: str = "", visite
def _sync_fsdp2_params_to_vllm(self, module: nn.Module):
# For FSDP2, module.state_dict() already covers all parameters, so no need for recursion
for name, param in module.state_dict().items():
# Fix FSDP2 + LoRA parameter naming issue with vLLM
if 'lora_' in name:
continue # skip LoRA parameters
name = name.removeprefix("base_model.model.").replace(".base_layer", "")

if param.is_cpu:
param = param.to(torch.device("cuda"))
param = param.full_tensor()
Expand Down
5 changes: 5 additions & 0 deletions trl/trainer/rloo_trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -808,6 +808,11 @@ def _sync_fsdp1_params_to_vllm(self, module: nn.Module, prefix: str = "", visite
def _sync_fsdp2_params_to_vllm(self, module: nn.Module):
# For FSDP2, module.state_dict() already covers all parameters, so no need for recursion
for name, param in module.state_dict().items():
# Fix FSDP2 + LoRA parameter naming issue with vLLM
if 'lora_' in name:
continue # skip LoRA parameters
name = name.removeprefix("base_model.model.").replace(".base_layer", "")

if param.is_cpu:
param = param.to(torch.device("cuda"))
param = param.full_tensor()
Expand Down