[CI]Main2Main 0814

[LQDLove]

What this PR does / why we need it?

Upgrade baseline

• Update the verified vLLM main anchor from a311916a to 827a2af. The full upstream range is available in this comparison.
• Preserve the vLLM 0.27.1 compatibility lane while adapting the main lane to the new upstream contracts.
• The changes are organized in the same order as the changed files in this PR. Each item identifies the upstream change, the downstream adaptation, and why the adaptation is required.

Changes by file

1. .github/vllm-main-verified.commit

Change	Upstream change	Downstream adaptation	Why
Update anchor to 827a2af	Upgrade window a311916a...827a2af.	Set anchor.	Source of truth for main2main workflow.

2. vllm_ascend/patch/platform/patch_vision.py (deleted) / vllm_ascend/patch/platform/__init__.py / vllm_ascend/patch/__init__.py

Change	Upstream change	Downstream adaptation	Why
Remove the FusedInputNorm eps patch	—	Delete patch_vision.py, its platform registration, and its patch-manifest documentation.	The FusedInputNorm.forward eps=0.0 → eps=1e-5 workaround is no longer needed.

3. vllm_ascend/ops/triton/mamba/postprocess.py

Change	Upstream change	Downstream adaptation	Why
Adapt to TEMPORAL_TILES tiling	Upstream vLLM tiled the temporal state copy across TEMPORAL_TILES CTAs (3D grid) and passes the new constexpr from run_fused_postprocess/run_fused_postprocess_align.	Port _memcpy_u64_tiled and accept TEMPORAL_TILES (default 1 keeps the 2D-grid v0.27.1 contract); temporal copies are tiled, conv copies run on tile 0 only, and the num_accepted store is guarded with tile_idx == 0.	Without the new constexpr, Triton raised "Keyword argument TEMPORAL_TILES was specified but unrecognised" and the engine died.

4. vllm_ascend/worker/v2/input_batch.py

Change	Upstream change	Downstream adaptation	Why
Fix AscendInputBatch dataclass field ordering	Upstream added default-valued fields (e.g. max_query_len) at the end of InputBatch's field list.	Give seq_lens_np a None default.	Without it, AscendInputBatch.seq_lens_np (no default) followed a default-valued base field, so dataclass raised "non-default argument follows default argument" at import time, failing every test.
Accept max_query_len in AscendInputBatch.make_dummy	Upstream vLLM main (after #51256) added max_query_len to InputBatch.make_dummy.	Add max_query_len and forward it only on the main lane.	v0.27.1's make_dummy has no such parameter; forwarding unconditionally raises TypeError.

5. vllm_ascend/worker/v2/model_runner.py / vllm_ascend/worker/v2/aclgraph_utils.py

Change	Upstream change	Downstream adaptation	Why
Accept context_len in execute_model	Upstream vLLM main (after #51256) passes context_len to execute_model for dummy-run context setup.	Add context_len: int = 0 and forward it to super().execute_model only on the main lane.	v0.27.1's execute_model has no context_len parameter, so forwarding it unconditionally raises TypeError on that lane.
Accept varlen_decode in the graph manager factory	Upstream vLLM main (after #51256) passes varlen_decode to the graph manager factory and ModelCudaGraphManager.	Add varlen_decode to graph_manager_wrapper.factory and ModelAclGraphManager.__init__, forwarding only on the main lane.	v0.27.1 has no varlen_decode, so forwarding unconditionally raises TypeError ("unexpected keyword argument 'varlen_decode'").
Accept batch_req_state in prepare_inputs	Upstream vLLM main (after #51256) passes an extra batch_req_state arg to prepare_inputs.	Accept *args and parse batch_desc by lane.	v0.27.1 passes only (scheduler_output, batch_desc); the main lane passes (scheduler_output, batch_req_state, batch_desc), otherwise TypeError ("takes 3 positional arguments but 4 were given").

6. vllm_ascend/worker/v2/spec_decode/autoregressive/speculator.py

Change	Upstream change	Downstream adaptation	Why
Skip DSA in _ascend_update_seq_lens	DSA metadata (AscendDSAMetadata) owns per-step state in its builder.	Early-return for the DSA architecture, mirroring _update_decode_attn_metadata.	AscendDSAMetadata has no seq_lens/seq_len_list, so the unconditional update raised AttributeError (seen on v0.27.1 DeepSeek-V4 MTP).

7. tests/ut/patch/worker/test_patch_mamba_utils_source.py

Change	Upstream change	Downstream adaptation	Why
Expect _memcpy_u64_tiled in the postprocess module	postprocess.py now also defines the _memcpy_u64_tiled helper.	Update the source-level check to expect it and assert the kernel calls it.	Keep the source check in sync with the kernel structure.

8. tests/ut/spec_decode/test_extract_hidden_states_proposer.py / tests/ut/ops/test_routed_experts.py / tests/ut/patch/platform/test_patch_structured_output.py

Change	Upstream change	Downstream adaptation	Why
Adapt UT to upstream 2ac1f683 API changes	Upstream removed extract_hidden_states.PIN_MEMORY; RoutedExperts.expert_map now probes quant_method.moe_kernel; StructuredOutputManager.compile_grammar gained a stop_token_ids arg and _create_grammar reads sampling_params.all_stop_token_ids.	Drop the stale _no_pin_memory fixture; mock quant_method on the __new__ instance; add stop_token_ids/all_stop_token_ids to the fakes.	Keep UT passing against the new upstream APIs.

9. tests/ut/spec_decode/test_extract_hidden_states_proposer.py

Change	Upstream change	Downstream adaptation	Why
Force pin_memory=False for CPU-only tests	Ascend reports pin_memory as available, so CpuGpuBuffer defaults it to True.	Patch CpuGpuBuffer.__init__ to force pin_memory=False.	On a pure CPU test device, pin_memory=True trips torch_npu's RegisterPrivateUse1HooksInterface check.

10. vllm_ascend/patch/platform/patch_kv_cache_coordinator.py

Change	Upstream change	Downstream adaptation	Why
Accept num_prefill_lookahead in the KV cache coordinator	Upstream vLLM main (after #50062) added num_prefill_lookahead to get_kv_cache_coordinator and the coordinator constructors; the scheduler now passes it to KVCacheManager.	Add num_prefill_lookahead: int = 0 to the patched get_kv_cache_coordinator and AscendHybridKVCacheCoordinator.__init__; forward it to the original coordinator on the main lane only; set num_reprefillable_tokens = max(0, num_prefill_lookahead - 1).	The scheduler passes num_prefill_lookahead, so without the param the patched coordinator raises TypeError (44 CPU UTs + 3 LoRA e2e failing with "Engine core initialization failed"). v0.27.1 has no such parameter, so it is forwarded only on the main lane.

Does this PR introduce any user-facing change?

No.

How was this patch tested?

CI on the branch. See Buildkite workflow run for detailed results.

• vLLM version: v0.27.1
• vLLM main: vllm-project/vllm@58d3918

[gemini-code-assist]

Summary of Changes

Hello, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request performs a routine synchronization of the CI verification commit hash. By updating this reference, the project ensures that continuous integration processes remain consistent with the current state of the main branch.

Highlights

• CI Synchronization: Updated the vLLM main verified commit hash to align with the latest upstream HEAD.

New Features

🧠 You can now enable Memory (public preview) to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	/gemini review	Performs a code review for the current pull request in its current state.
Pull Request Summary	/gemini summary	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	/gemini help	Displays a list of available commands.

Customization

To customize the Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counterproductive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for GitHub and other Google products, sign up here.

Footnotes

1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[gemini-code-assist]

Code Review

This pull request updates the verified commit hash of the vLLM main branch in .github/vllm-main-verified.commit to a311916a291c1fed3dbfb72e60f74cd778c8419d. The reviewer pointed out that the PR title and description do not follow the repository's style guide and provided suggestions to format them correctly.

[gemini-code-assist]

The Pull Request title and description do not follow the repository's style guide. Please update them to adhere to the required format.

Suggested PR Title:

[CI][Misc] Update vllm-main-verified.commit to latest main HEAD

Suggested PR Summary:

### What this PR does / why we need it?

This PR updates the verified commit hash of the vLLM main branch in `.github/vllm-main-verified.commit` to the latest HEAD (`a311916a291c1fed3dbfb72e60f74cd778c8419d`). This ensures that CI workflows and downstream dependencies use the latest verified commit of vLLM.

### Does this PR introduce _any_ user-facing change?

No.

### How was this patch tested?

This is a configuration/commit hash update for CI. Testing is handled by the CI workflows that consume this commit hash.

References

1. The PR title and summary must follow the specified format in the Repository Style Guide. ^(link)

[github-actions]

👋 Hi! Thank you for contributing to the vLLM Ascend project. The following points will speed up your PR merge:‌‌

• A PR should do only one thing, smaller PRs enable faster reviews.
• Every PR should include unit tests and end-to-end tests ‌to ensure it works and is not broken by other future PRs.
• Write the commit message by fulfilling the PR description to help reviewer and future developers understand.

If CI fails, you can run linting and testing checks locally according Contributing and Testing.

[LQDLove]

/rerun
[Bot]: rerun completed.

Rerun:

• E2E

[LQDLove]

/rerun

[Bot]: rerun completed. No failed jobs found.