Skip to content

[Bug fix] Avoid OOM when casting fp32 on NPU for GRPO with vLLM colocate #9335

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
hjh0119 merged 7 commits into from
May 19, 2026
+10 −0
Merged

[Bug fix] Avoid OOM when casting fp32 on NPU for GRPO with vLLM colocate #9335

Changes from 3 commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
dc5332b
avoid OOM when casting fp32 on NPU for GRPO with vLLM colocate
ys2025-AI May 13, 2026
9bfc878
Merge pull request #1 from ys2025-AI/fsdp2_npu
ys2025-AI May 13, 2026
b466051
Merge branch 'modelscope:main' into main
ys2025-AI May 13, 2026
b48494a
Update fsdp.py
ys2025-AI May 15, 2026
1f6ec9c
Merge branch 'modelscope:main' into main
ys2025-AI May 19, 2026
3e32487
Update fsdp.py
ys2025-AI May 19, 2026
0c4cbe2
Update fsdp.py
ys2025-AI May 19, 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
9 changes: 9 additions & 0 deletions swift/model/npu_patch/fsdp.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -26,7 +26,16 @@ def _cast_module_to_fp32_for_npu_if_needed(module: torch.nn.Module, accelerator:
# entering that path with bf16/fp16 parameters can fail before mixed
# precision policy has a chance to manage runtime compute dtype. Cast early
# while parameters are still on CPU or meta, so only dtype changes here.

# GRPO with vLLM colocate mode may preload the model onto NPU before
# Accelerator.prepare() is called. In that case, casting fp32 on NPU
# would temporarily duplicate the full model (bf16 + fp32), causing OOM.
# We move the model back to CPU first to free NPU memory, then cast.
try:
import torch_npu
module = module.cpu()
torch_npu.npu.synchronize()
torch_npu.npu.empty_cache()
Copy link
Copy Markdown
Contributor

@gemini-code-assist gemini-code-assist Bot May 13, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

medium

It is recommended to only move the module to CPU and clear the NPU cache if the parameters are currently residing on the NPU. This avoids unnecessary synchronization and cache clearing overhead for standard training paths (like SFT or LoRA) where the model is already on the CPU or meta device. It also ensures better compatibility with meta-device initialization, as calling .cpu() on meta-parameters might lead to unexpected behavior depending on the PyTorch version.

Suggested change
import torch_npu
module = module.cpu()
torch_npu.npu.synchronize()
torch_npu.npu.empty_cache()
if param.device.type == 'npu':
import torch_npu
module = module.cpu()
torch_npu.npu.synchronize()
torch_npu.npu.empty_cache()

return module.to(torch.float32)
except Exception as exc:
raise NPUCastError(f'Failed to cast {module.__class__.__name__} to fp32.') from exc
Expand Down
Loading