Skip to content

[XPU] [torch.compile] Skipping CUDA graph memory estimation to avoid startup errors. #39977

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
jikunshang merged 2 commits into from
Apr 20, 2026
+6 −2
Merged

[XPU] [torch.compile] Skipping CUDA graph memory estimation to avoid startup errors. #39977

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
fc95df8
[XPU] skip cudagraph memory estimate on xpu
chaojun-zhang Apr 16, 2026
03410e0
Merge branch 'main' into fix_start_error
jikunshang Apr 18, 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
8 changes: 6 additions & 2 deletions vllm/v1/worker/gpu_worker.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -374,10 +374,14 @@ def determine_available_memory(self) -> int:
)

# Profile CUDA graph memory if graphs will be captured.
# Skip on ROCm/HIP as graph pool handles and mem_get_info behave
# Skip on ROCm/HIP/XPU as graph pool handles and mem_get_info behave
# differently and can produce incorrect/negative estimates.
cudagraph_memory_estimate = 0
if not self.model_config.enforce_eager and not current_platform.is_rocm():
if (
not current_platform.is_rocm()
and self.vllm_config.compilation_config.cudagraph_mode
!= CUDAGraphMode.NONE
):
cudagraph_memory_estimate = self.model_runner.profile_cudagraph_memory()

# Use the pre-cudagraph torch peak to avoid double-counting.
Expand Down
Loading