Skip to content

Cherry-pick PR #297 to v0.3.0-dev: Add thead vendor attention backend - #360

Merged
cyber-pioneer merged 1 commit into
v0.3.0-devfrom
cherry-pick-297-to-v0.3.0
Aug 21, 2026
Merged

Cherry-pick PR #297 to v0.3.0-dev: Add thead vendor attention backend#360
cyber-pioneer merged 1 commit into
v0.3.0-devfrom
cherry-pick-297-to-v0.3.0

Conversation

@physics31415926

@physics31415926 physics31415926 commented Aug 10, 2026

Copy link
Copy Markdown
Collaborator

This PR cherry-picks #297 (add thead vendor attention backend) to the v0.3.0-dev branch.

Test Results

Environment: Host 810e, vLLM 0.24.0 + vllm-plugin-FL PR#360

Model: Qwen3.6-35B-A3B (MoE, TP=1, BF16)

✅ CUDA Graph Mode (Default)

  • Status: PASS
  • Config: max_model_len=2048, max_num_seqs=4, no enforce_eager
  • Results:
    • Inference completed successfully without hang or crash
    • Output quality: 2/3 prompts generated correct responses (first prompt showed warmup artifacts, expected behavior)
    • KV cache: 379,699 tokens, max concurrency: 185.40x
    • Compilation time: 55.32s (torch.compile), total init: 379.45s

✅ Eager Mode

  • Status: PASS
  • Config: enforce_eager=True, max_model_len=2048
  • Results: All outputs semantically correct

Conclusion: PR#360 works correctly on thead hardware with both eager and CUDA graph modes.

<!-- One of [Core | Vendor | OP | Tools | Others] -->
vendor
<!-- One of [User Experience | New Features | Bug Fixes | Improvements |
Performance | Breaking Change | Deprecations | Test Case | Docs |
Others] -->
User Experience
<!-- Describe what this PR does and why. -->
add thead vendor attention backend
<!-- Link any related issues: Fixes #issue, Closes #issue, or Related to

<!-- List the key changes made in this PR. -->
-
add thead vendor attention backend
<!-- How has this change been tested? Include test commands, hardware
used, etc. -->
-
verify qwen3.6-35B use vendor attention backend on PPU without accuracy
and perf
verify minicpm5 use vendor attention backend on PPU with accuracy and
perf
- [ ] I have run the existing tests and they pass
- [ ] I have added tests for my changes (if applicable)
- [ ] I have updated the documentation (if applicable)

(cherry picked from commit d1327ae)
@github-actions github-actions Bot added the core label Aug 10, 2026
@physics31415926
physics31415926 requested review from ceci3 and cyber-pioneer and removed request for ceci3 August 10, 2026 08:41
@physics31415926
physics31415926 marked this pull request as draft August 10, 2026 08:48
@physics31415926
physics31415926 marked this pull request as ready for review August 20, 2026 07:23
@cyber-pioneer
cyber-pioneer merged commit 5b109ed into v0.3.0-dev Aug 21, 2026
3 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants