[CI] vllm main2main 0731 0351e9a

[zhangxinyuehfad]

What this PR does / why we need it?

Files	Upstream vLLM change	vllm-ascend adaptation
vllm_ascend/__init__.py	vllm#48841 — added from triton.experimental import gluon, from triton.experimental.gluon import language as gl and from triton.language.core import _aggregate to triton_utils/__init__.py, requiring Triton 3.6+ API absent from triton-ascend 3.2.1	Pre-register triton.experimental.gluon / .language as sys.modules stubs; stub triton.language.core._aggregate. Gated on os.getenv("VLLM_VERSION") != "0.26.0", mirroring vllm_version_is env var path.
tests/ut/patch/platform/test_patch_structured_output.py	vllm#49665 — changed VLLMValidationError base class from ValueError to VLLMClientError, breaking pytest.raises(ValueError) assertions	Updated 3 assertions from pytest.raises(ValueError, ...) to pytest.raises(VLLMValidationError, ...)
tests/e2e/pull_request/one_card/test_guided_decoding.py	vllm#49665 — same VLLMValidationError base-class change; on 0.26.0 the upstream validation may still raise ValueError	Version-gated assertion: pytest.raises(ValueError, ...) on 0.26.0 vs pytest.raises(VLLMValidationError, ...) on main
vllm_ascend/ops/vocab_parallel_embedding.py vllm_ascend/_310p/ops/vocab_parallel_embedding.py	vllm#49731 — added *, disable_tp: bool = False keyword arg to ParallelLMHead.__init__() and VocabParallelEmbedding.__init__()	Version-gated __init__ with vllm_version_is("0.26.0"): else branch accepts disable_tp kwarg and forwards it (Ascend manages TP via lmhead_tp_enable())
vllm_ascend/distributed/kv_transfer/kv_p2p/mooncake_connector.py vllm_ascend/distributed/kv_transfer/kv_p2p/mooncake_hybrid_connector.py	vllm#26593 — KVConnectorBase_V1.__init__() sets self._kv_transfer_config, required by requires_kv_delivery property	Version-gated __init__: else branch sets self._kv_transfer_config on vllm main
vllm_ascend/ops/fused_moe/fused_moe.py	vllm#50089 — added fused_output_is_reduced kwarg to _maybe_reduce_shared_expert_output() and output_is_reduced kwarg to _maybe_reduce_final_output()	Version-gated both methods: else branch accepts new kwargs but ignores them (Ascend handles reduction independently); trunc_size may now be None
vllm_ascend/patch/worker/patch_distributed.py	vllm#47288 — added use_all2all: bool = False param to GroupCoordinator.__init__()	Version-gated via _IS_VLLM_026 env var check (avoids vllm-ascend import cycle). Else branch stores self.use_all2all (unused, no all2all on Ascend)
vllm_ascend/distributed/device_communicators/npu_communicator.py	vllm#47288 — added use_all2all: bool = False param to DeviceCommunicatorBase.__init__()	Version-gated NPUCommunicator.__init__ with vllm_version_is("0.26.0"): else branch accepts and forwards use_all2all (NPU keeps the no-op _NpuAll2AllManager)
vllm_ascend/worker/npu_input_batch.py	vllm#48018 — added use_replayssm: bool kwarg (Mamba replay-SSM); vllm#40996 — added slot_mapping_modes kwarg (DCP hybrid attention)	Accepts both kwargs unconditionally; stores self.use_replayssm and self.slot_mapping_modes only on main (not vllm_version_is("0.26.0")) for interface alignment
vllm_ascend/worker/v2/spec_decode/eagle/aclgraph.py	vllm#49364 — renamed skip_attn parameter to full_cudagraph in cudagraph mode logic	Version-gated call: 0.26.0 keeps skip_attn=(cg_mode != PIECEWISE); main passes full_cudagraph=(cg_mode != PIECEWISE)
vllm_ascend/distributed/weight_transfer/npu_ipc_engine.py vllm_ascend/distributed/weight_transfer/__init__.py	vllm#48042 + vllm#48981 — replaced static IPCTrainerSendWeightsArgs/trainer_send_weights with a stateful IPCTrainerWeightTransferEngine driven by WeightTransferTrainerFactory.trainer_init(...).send_weights(); packed moved onto trainer init-info; per-round update_info slimmed	Whole module version-gated: 0.26.0 keeps the static NPUIPCWeightTransferEngine; main defines stateful NPUIPCTrainerWeightTransferEngine (_send_unpacked instance method, packed on NPUIPCWeightTransferInitInfo, no-arg npu_generate_uuid(), is_sender/_all_gather_and_merge_handles). register_engine() also registers the trainer engine in WeightTransferTrainerFactory on main only
examples/rl/rlhf_http_npu_ipc.py	vllm#48981 — ported RL examples to WeightTransferTrainerFactory.trainer_init(...) + engine.send_weights() with HTTPVLLMWeightSyncClient/ModuleSource	Version-gated example: 0.26.0 uses static NPUIPCWeightTransferEngine.trainer_send_weights; main uses the stateful trainer engine path
tests/e2e/pull_request/one_card/test_npu_ipc_weight_transfer.py	vllm#48981 — IPC weight-transfer E2E now exercised via the stateful trainer engine	Version-gated test: 0.26.0 keeps the manual _post(start/finish) + static trainer_send_weights; main drives WeightTransferTrainerFactory.trainer_init(...).send_weights() (engine owns the lifecycle)
tests/ut/distributed/weight_transfer/test_npu_ipc_engine.py	vllm#48981 — trainer-side IPC APIs changed from static methods to instance methods on the stateful engine	Version-gated tests (IS_VLLM_026): _send_unpacked, packed placement, npu_generate_uuid arity, and start/finish_weight_update no-op behavior asserted per version
vllm_ascend/worker/v2/spec_decode/dflash/speculator.py	vllm#50000 — _prepare_dflash_inputs_kernel gained temperature/seeds sampling params for probabilistic draft sampling	Version-gated _prepare_dflash_inputs_kernel_ascend: 0.26.0 branch keeps the old signature; main branch adds the four temperature/seeds pointers and the corresponding stores to stay aligned with upstream while keeping Ascend's own kernel
.github/vllm-main-verified.commit	—	Updated verified main commit hash to 0351e9aa1fdf1a51329d1906881528dfe61fc88e

Does this PR introduce any user-facing change?

How was this patch tested?

• vLLM version: v0.26.0
• vLLM main: vllm-project/vllm@d02df74

[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 PR synchronizes vllm-ascend with the latest vLLM upstream changes while ensuring backward compatibility with vLLM 0.26.0. It introduces version-gated logic across core components, including distributed communicators, weight transfer engines, and custom ops, to handle upstream API evolutions. Additionally, it updates verified commit hashes and refactors the weight transfer engine to support stateful trainer-side operations.

Highlights

• Triton API Compatibility: Added stubs for triton.experimental.gluon and triton.language.core._aggregate to support vLLM main branch requirements on Triton-Ascend 3.2.1.
• Version-Gated API Adaptations: Implemented extensive version-gating (vLLM 0.26.0 vs. main) across ops, distributed, and worker modules to accommodate upstream API changes like disable_tp, use_all2all, and VLLMValidationError.
• Weight Transfer Engine Refactor: Migrated NPU IPC weight transfer to a stateful NPUIPCTrainerWeightTransferEngine for vLLM main, while maintaining the legacy static path for 0.26.0.
• Kernel and Test Updates: Synchronized NPU-specific kernels and E2E/UT test suites with upstream changes, including updated sampling parameters and exception handling.

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. ↩

[mergify]

⚠️ The sha of the head commit of this PR conflicts with #11090. Mergify cannot evaluate rules on this PR. Once #11090 is merged or closed, Mergify will resume processing this PR. ⚠️

[gemini-code-assist]

Code Review

Suggested PR Title:

[Ops][Misc] Add compatibility support for vLLM main branch (post-0.26.0)

Suggested PR Summary:

### What this PR does / why we need it?
This pull request introduces comprehensive compatibility updates to ensure the `vllm-ascend` plugin works seamlessly with both vLLM `0.26.0` and the upstream `main` branch (post-`0.26.0`). It addresses various API, interface, and signature changes introduced in the upstream vLLM repository, including:
- Refactoring the NPU IPC weight transfer engine to support both the static `0.26.0` path and the stateful `IPCTrainerWeightTransferEngine` on `main`.
- Aligning signatures for `ParallelLMHead`, `NPUCommunicator`, `GroupCoordinator`, `MooncakeConnector`, `fused_moe` operators, and `InputBatch` to handle new arguments (such as `disable_tp`, `use_all2all`, `use_replayssm`, and `slot_mapping_modes`).
- Stubbing missing Triton modules (`triton.experimental.gluon`) and core functions (`_aggregate`) to prevent import failures on newer vLLM versions when using `triton-ascend 3.2.1`.
- Updating Triton kernels and Speculator/Eagle execution paths to match the updated upstream signatures (e.g., adding temperature/seeds parameters to `_prepare_dflash_inputs_kernel_ascend`).

Additionally, three critical issues were identified in the review:
1. In `examples/rl/rlhf_http_npu_ipc.py`, the local variable `client` shadows the global OpenAI `client`, which will cause an `AttributeError` during generation. It should be renamed to `sync_client`.
2. In `vllm_ascend/distributed/weight_transfer/npu_ipc_engine.py`, `device_index` is retrieved but not passed to `npu_generate_uuid()`, which can lead to generating the wrong physical NPU UUID.
3. The corresponding unit test in `tests/ut/distributed/weight_transfer/test_npu_ipc_engine.py` needs to be updated to assert that `npu_generate_uuid` is called with `device_index` on both branches.

### Does this PR introduce _any_ user-facing change?
No user-facing changes are introduced. These are internal compatibility and alignment updates to support different vLLM versions.

### How was this patch tested?
The changes are covered by existing and updated end-to-end and unit tests, which have been adapted to handle version-specific behaviors.

[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.

[wxsIcey]

why change the ops? #48042 is not associate with it.

[zhangxinyuehfad]

The kernel change is required because upstream _prepare_dflash_inputs_kernel gained 4 temperature/seeds params in #50000 (Kimi K3); since the Ascend kernel is monkey-patched, the signature must match the upstream call site or Triton fails at launch. You're right that #48042 is unrelated — that was a typo, the comment/PR description will be corrected to #50000.

[github-actions]

This pull request has conflicts, please resolve those before we can evaluate the pull request.

[github-actions]

This pull request has conflicts, please resolve those before we can evaluate the pull request.