Skip to content

[vLLM][moe] Keep batched moe reference activation eager, guard CUDA stream-capture checks, and update skiplist#7440

Draft
sarah12121212 wants to merge 4 commits into
mainfrom
moe2-batched-moe2
Draft

[vLLM][moe] Keep batched moe reference activation eager, guard CUDA stream-capture checks, and update skiplist#7440
sarah12121212 wants to merge 4 commits into
mainfrom
moe2-batched-moe2

Conversation

@sarah12121212

@sarah12121212 sarah12121212 commented Jul 14, 2026

Copy link
Copy Markdown
Contributor

This updates the vLLM patch for batched MoE XPU coverage.

The patch:

  • builds the MoE torch reference activation with compile_native=False
  • reuses that eager activation in the expert loop
  • guards torch.cuda.is_current_stream_capturing() behind current_platform.is_cuda()
  • moves tests/kernels/moe/test_batched_moe.py::test_fused_moe_batched_experts in the default and Xe2 MoE skiplists to the WorkspaceManager not initialized error section.

CI Validation:

PVC:
BMG:

@sarah12121212 sarah12121212 changed the title [vLLM][moe] Keep batched moe reference activation eager and guard CUDA stream-capture checks [vLLM][moe] Keep batched moe reference activation eager, guard CUDA stream-capture checks, and update skiplist Jul 14, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant