Skip to content

[3/3] Optimize: Cache CUDA device to reduce redundant calls during tensor l… #8996

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
zhyncs merged 4 commits into from
Aug 10, 2025
Merged

[3/3] Optimize: Cache CUDA device to reduce redundant calls during tensor l… #8996

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
e8ac749
Optimize: Cache CUDA device to reduce redundant calls during tensor l…
GeLee-Q Aug 9, 2025
d5984ff
rm monkey_patch_torch_reductions in _unwarp_tensor
GeLee-Q Aug 9, 2025
367c865
fix: correct argument passing for _unwarp_tensor
GeLee-Q Aug 9, 2025
966d634
Merge branch 'main' into reduce_cuda_device_calls
hebiao064 Aug 10, 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
11 changes: 7 additions & 4 deletions python/sglang/srt/model_executor/model_runner.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -895,8 +895,12 @@ def update_weights_from_tensor(
named_tensors: List[Tuple[str, Union[torch.Tensor, "LocalSerializedTensor"]]],
load_format: Optional[str] = None,
):
monkey_patch_torch_reductions()
# We need to get device after patch otherwise the device would be wrong
infered_device = torch.cuda.current_device()

named_tensors = [
(name, _unwrap_tensor(tensor, tp_rank=self.tp_rank))
(name, _unwrap_tensor(tensor, tp_rank=self.tp_rank, device=infered_device))
for name, tensor in named_tensors
]
if load_format == "direct":
Expand Down Expand Up @@ -1809,11 +1813,10 @@ def _model_load_weights_direct(model, named_tensors: List[Tuple[str, torch.Tenso
default_weight_loader(params_dict[name], tensor)


def _unwrap_tensor(tensor, tp_rank):
def _unwrap_tensor(tensor, tp_rank, device):
if isinstance(tensor, LocalSerializedTensor):
monkey_patch_torch_reductions()
tensor = tensor.get(tp_rank)
return tensor.to(torch.cuda.current_device())
return tensor.to(device)


@dataclass
Expand Down
Loading