[Community][Misc] Refresh maintainer team

[wangxiyuan]

docs: add maintainer emeritus mechanism and retire ganyi1996ppo, whx-sjtu, MengqingCao

• Add Maintainer Emeritus section in governance.md describing retirement process, retained permissions, and welcome-back policy

• Move ganyi1996ppo, whx-sjtu, MengqingCao from active Committers to Emeritus Committers in contributors.md

---

• vLLM version: v0.26.0
• vLLM main: vllm-project/vllm@0351e9a

[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 updates the vLLM Ascend community documentation to reflect changes in the maintainer team. It introduces an 'Emeritus' status for maintainers who are no longer active, formalizing the transition process in the governance policy and updating the contributors list accordingly.

Highlights

• Maintainer List Update: Updated the maintainer list by adding new members and moving several individuals to the newly created Emeritus Committers section.
• Governance Documentation: Added a new section to the governance documentation defining the 'Maintainer Emeritus' status and the process for transitioning to it.

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 community documentation by adding new active committers, creating an "Emeritus Committers" section, and defining the policy for transitioning active maintainers to emeritus status in the governance guidelines. The review feedback correctly points out that the PR title and description do not follow the repository's Pull Request Summary Style Guide and provides a compliant template.

[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.

[wangxiyuan]

@Tflowers-0129

• ✅ Review Quality‌:

He has reviewed 102 distinct PRs and submitted at least 118 review records in vLLM Ascend in 2026. His reviews cover Ascend 310P/A5 adaptation, model support, quantization, ACLGraph, linear attention, speculative decoding, LoRA, CI, main2main upgrades, and documentation.

His reviews consistently focus on code correctness, numerical accuracy, hardware alignment constraints, boundary conditions, test coverage, regression protection, performance impact, and consistency with upstream vLLM semantics. He has demonstrated a deep understanding of the vLLM and vLLM Ascend codebases, Ascend NPU operator behavior, distributed inference, graph execution, and CI workflows, and has provided high-quality and actionable feedback to contributors.

• ✅ Sustained Contributions‌:

He has made sustained contributions to vLLM Ascend throughout 2026, with more than 40 merged PRs. His main contribution area is the end-to-end enablement and continuous evolution of the Ascend 310P/300I DUO backend, including architecture decoupling, model adaptation, quantization, graph mode, performance optimization, accuracy fixes, CI coverage, and deployment documentation.

310P architecture and basic model support:

• [Feature]: Support 310P device run qwen2.5/3 dense and qwen2.5vl models #5776 introduced the initial 310P support for Qwen2.5/Qwen3 dense models and Qwen2.5-VL.
• [Refact.]: refactoring for 310p kvcache and some ops class #6117 refactored the 310P KV cache, attention, model-runner, normalization, and activation implementations, reducing coupling between the 310P backend and the mainline A2/A3 implementation.
• RFC [RFC]: Support Qwen3 for 300I DUO #6600 defined the architecture of the dedicated _310p backend and proposed routing platform-specific operators and attention implementations through the platform layer.

Quantization and weight-layout optimization:

• [Feat.]: support 310p w8a8 #6454 added static W8A8 quantization support for 310P, including platform-specific quantization configuration and linear methods.
• [Feat.][310P]: weightNZ feature with quant or unquant. #6705 introduced FRACTAL_NZ weight-format support for both quantized and unquantized linear layers.
• [300I][Bugfix] fix unquant model weight nd2nz error #6851 fixed incorrect ND-to-NZ conversion for unquantized models and centralized the conversion logic to avoid duplicated platform-specific implementations.

ACLGraph and speculative-decoding support:

• [300I] support decode-only aclgraph mode #6849 added decode-only ACLGraph support for 310P.
• [310P][Bugfix]: fix ngram graph replay accuracy error #7134 fixed an accuracy issue caused by incorrect graph capture and replay state selection when ACLGraph was combined with n-gram speculative decoding.
• [Performance][ops][310p] Support Qwen3.5-Dense ACLGraph with fused operators #9104 integrated fused causal-conv1d and recurrent-gated-delta-rule operators to enable graph capture for Qwen3.5/Qwen3.6 Dense models.
• [Feature][310p] support compressed mask for 310P flashattention and pagedattention-split-fuse #10181 added compressed-mask support for 310P FlashAttention and PagedAttention split-fuse paths.

Qwen3.5 and linear-attention enablement:

• [310P]fused recurrent gated delta rule pytorch core and ut #7398, [310p]: add rmsnorm gated fallback and unit test #7424, [310p] fused gdn gating ut #7430, and [310P]: add torch chunk gated delta rule and 910b parity ut #7594 implemented the 310P PyTorch fallback components for recurrent gated delta rule, RMSNormGated, GDN gating, and chunk gated delta rule.

• [Feature][310P] support shared experts path in fused MoE for qwen3.5 #7674 enabled the Qwen3.5 shared-experts execution path on 310P.

• [BugFix][310p] Align 310P GDN state semantics with vLLM #7902 aligned the GDN state layout with upstream vLLM semantics, preparing the backend for fused operator integration.

• [Feature][310p] support Prefix Mamba Cache on 310p #9514 added Prefix Mamba Cache support for 310P.

• ✅ Quality Contribution‌:

He has fixed multiple important accuracy, stability, compatibility, and availability issues affecting real 310P deployments.

Accuracy and execution stability:

• [BugFix][310p] Fix post-sampling not working in graph mode on 310P #8017 fixed a graph-mode post-sampling hang caused by the unsupported behavior of q.exponential_() and the underlying Add kernel.
• [BugFix][310p]Fix Triton kernel in block_table crash that caused failures #8085 prevented 310P from incorrectly entering unsupported Triton block-table kernels.
• [BugFix][310p] Fixing the aclgraph error caused by blocktable #8948 moved the block-table slot-mapping calculation to the CPU to avoid unsupported device-side arithmetic during ACLGraph execution.
• [BugFix][Ops][310p]:fix the accuracy issue caused by MoEGatingTopkSoftmax #11391 identified the root cause of Qwen3.5-MoE accuracy degradation: UB overflow in MoEGatingTopkSoftmax when the token dimension was 2048, which could produce INF values.
• [BugFix][310p] Qwen3.5 shared-experts linear failuer when out_dim=1 #11714 and its release backport [BugFix][310P] Fix shared experts linear when out_dim=1 #11730 fixed shared-expert linear failures when out_dim=1, keeping degenerate weights in ND format instead of converting them to unsupported FRACTAL_NZ shapes.

CI and environment reliability:

• [CI]: add production safeguards for 300I #6343 enabled per-PR 310P E2E safeguards through the file tracker.

• [BugFix][310p] Fix torch-npu cannot import error #9249 fixed torch_npu import failures by correcting the uninstall order of triton-ascend and triton.

• [BugFix][CI][310p] Fix CI error for 310p caused by DSV4 #9402 fixed a 310P online-CI failure introduced by DeepSeek V4 block-table changes.

• [CI] Upgrade 310P CANN image to 9.1.0-beta.1 #9710 upgraded all relevant 310P CI, wheel, and Docker images to CANN 9.1.0-beta.1.

• [CI][310p]: Optimize 310P E2E test duration #11528 reduced 310P E2E execution time by removing redundant coverage while retaining the necessary validation.

• ✅ Community Involvement‌:

He has taken sustained ownership of the 310P/300I DUO technical roadmap rather than contributing only isolated fixes.

He authored RFC #6600 to establish a decoupled 310P backend architecture. He subsequently authored RFC #9702 to complete compressed attention masks, Prefix Mamba Cache, MTP, and function-calling support, and RFC #9711 to improve Qwen3.5 performance through fused chunk-gated-delta-rule, recurrent-gated-delta-rule, grouped-matmul, and split-fuse graph optimizations.

He also improved usability through #7065 and #8640 by documenting 310P-specific configuration requirements and providing a detailed deployment guide, reducing the cost for users to install, configure, and run vLLM Ascend on Atlas inference products.

Ongoing open contributions, recorded separately from merged PRs:

• [Performance][310p] overlap shared expert compute #9107 — overlap shared-expert computation on 310P.
• [Misc][310p] Fix CPU binding and NPU memory usage issues during server start #9599 — fix CPU binding and NPU memory usage during server startup.
• [Feature][Gemma4][A5]: Support Gemma4 for A5 Device #10364 — add Gemma4 support for A5 devices.
• [Performance][310p] MTP optimization #11662 — optimize MTP performance on 310P.
• [Performance][310p]support multi-stream-shared-experts #12388 — enable multi-stream shared experts on 310P.

Closed or reworked but unmerged development PRs, retained as contribution history:

#5686, #5773, #5774, #6706, #6843, #7113, #7212, #7238, #7346, #7397, #7516, #7933, #7972, #7992, #7999, #8295, #8672, #9084, and #9106.

These PRs include early iterations of the 310P basic-model adaptation, Qwen3.5 and linear-attention exploration, graph-mode validation, dependency cleanup, operator compilation fixes, and fused GDN integration. They should be recorded as engineering and review history, but should not be counted as merged contributions.

Overall, his sustained work on architecture, operator adaptation, graph execution, quantization, model enablement, accuracy, CI, documentation, and technical planning is the main reason why I nominated him.

[wangxiyuan]

@kunpengW-code

• ✅ Review Quality:
  Since February 2025, he has completed 50+ reviews, including discussion_r3490493840, discussion_r3511749400, discussion_r3511522126, discussion_r2945417799 with high quality.

• ✅ Sustained Contributions:
  Since February 2025, he has had 39 merged PRs, with continuous activity across multiple major release cycles (v0.18.0, v0.20.2rc1, v0.21.0rc, v0.23.0, v0.24.0rc).

• ✅ Quality Contributions:

  Quantization Feature Owner:
  He is one of the feature owners and main contributors of quantization. He implemented support for multiple quantization formats in MOE W4A8 quantization (#2392, #3011), is the co-author of MLA kvcache int8 quantization (#7222), led the completion of e2e test cases for quantization features and the removal of redundant quantization cases (#8666, #8806, #9160), and completed MC2 dispatch MXFP4 quantization support (#9365).

  Model Runner V1 Refactoring:
  He has been deeply involved in the model runner v1 refactoring. He removed unnecessary attributes from set_ascend_forward_context (#5204), and aligned the structure of two core methods excute_model and _dymmy_run in NPUModelRunner with GPUModelRunner to improve maintainability (#6043).

  Prefill Performance Optimization:
  He participated in the prefill optimization for large EP and completed the graph mode adaptation for prefill optimization (#2090), and is the co-author of the shared expert DP feature (#2198).

  Some Important Bugfixes:
  Fixed the MHA model graph mode error (#5397), fixed the PCP/DCP accuracy issue (#6491), and fixed the dsv3.2 block_size setting failure error (#7630).

• ✅ Community Involvement:
  He led the design and implementation of the following RFCs:
  [RFC]: Support MOE w4a8 per-channel quantization #3214 -- Support W4A8 per-channel granularity quantization to further improve MOE model performance.
  [RFC]: vLLM-Ascend Operator Direct Tuning #4298 -- vllm-ascend supports direct operator invocation. Operators not supported by PTA can be first supported in vllm-ascend through the EXEC_NPU_CMD macro.

  He has completed 12+ bugfix, covering quantization, graph mode, kvcache, feature stacking, etc.

  Since March 2026, he has taken part in issue triage and user support, responding to 140+ issues.

[wangxiyuan]

@HF-001

He is nominated primarily for his professional expertise, key contributions to KV cache offloading, speculative inference and asynchronous scheduling, as well as sustained substantial improvements around these features.
Since October 2025, @HF-001 has started contributing with the support of KV cache offloading. He has delivered around 17 commits and bug fixes to improve vLLM. Besides, he helped more than ten users resolve issues related to KV cache offloading and speculative inference.
I believe his participation will further strengthen the support of KV cache offloading, speculative decoding and asynchronous scheduling on vLLM Ascend.

• ✅ Review Quality‌:

https://github.com/vllm-project/vllm-ascend/pulls?q=commenter%3AHF-001

• ✅ Sustained and Major Contributions:
  My main contributions are in kvcache management, speculative inference, and asynchronous scheduling. kvcache management includes: native version kvcache offloading, batch kvcache offloading, simple version kvcache offloading, native multi-level caching, etc. Speculative inference includes: unified parallel speculative inference (PARD \ p-eagle), draft_madel, ngram_npu, dynamic speculative inference, etc. Asynchronous scheduling includes: zero consumable asynchronous scheduling+speculative inference, ngram_npu+asynchronous scheduling, zero-bubble asynchronous scheduling+pcp/dcp/hybrid model, etc. In addition, there are fixes for dsv4 prefix cache hit rate issues, dflash error issues, etc.
  https://github.com/vllm-project/vllm-ascend/pulls?q=is%3Apr+is%3Aclosed+author%3AHF-001+
  [Feature] change ngram spec deode to triton op #11925
  [BugFix] fix dflash spec decode rope and graph error #12940
  [Feature] SD spec decode op #12368
• ✅ Community Involvement‌:
  https://github.com/vllm-project/vllm-ascend/issues?q=is%3Aissue%20commenter%3AHF-001

[wangxiyuan]

@ZT-AIA

✅ Review Quality

He has completed 68 PR reviews (62 closed + 6 open) since Apr 2026, with a focus on custom-op correctness, CANN/HDK compatibility, CI stability, and bugfix validation across the project.

✅ Sustained Contributions

He has contributed 31 merged PRs (all approved) and co-authored 1 large onboarding PR (#9228), covering AscendC custom operators, DeepSeek V4 custom ops, CANN/HDK version adaptation, precision bugfix, and custom-op CI infrastructure.

✅ Quality Contributions

• AscendC Custom Operators (KV-cache): As the owner of the store_kv_block / storekv_aicpu operator family, he delivered the initial implementation and continues to fix metadata and data slicing issues, which are critical to KV-cache correctness on Ascend:

  • [Feature] add ascendc ops store_kv_block #10292
  • [BugFix] Added the store_kv_block_metadata ascendC operator #11865
  • [BugFix] Fix the issue of incorrect data slicing in storekv_aicpu #12124
  • [CI] del atb_npu_reshape_and_cache ops #11900

• DeepSeek V4 Custom Op Suite: As co-author of the DeepSeek V4 custom op onboarding PR, he helped land the fused MoE/GMM + SwiGLU + quant kernel suite that DeepSeek V4 inference depends on:

  • [Ops][Feature] Add DeepSeek V4 custom operators #9228 (co-authored; ports dequant_swiglu_quant, grouped_matmul_swiglu_quant_weight_nz_tensor_list, grouped_matmul_swiglu_quant_v2)

• Precision Bugfix: He fixed a key precision regression in DeepSeek V4 caused by the dequant_swiglu_quant custom op, restoring numerical correctness for production W8A8 inference:

  • [Feature] Fix the precision issue of dsV4 caused by dequant_swiglu_quant #9600

• CANN / HDK Version Adaptation: He proactively fixed compatibility issues across CANN and HDK revisions, ensuring vllm-ascend builds and runs on the latest SDKs:

  • [BugFix] fix for new verision of cann #11445
  • [BugFix] Fixing compilation errors in the new version of cann vLLM Ascend #11684
  • [BugFix] A2G3 adapted for the CH version of HDK #11504
  • A2G3 adapted for the CH version of HDK #11382

• Custom-op CI Infrastructure: He owns the nightly custom-op CI stability, delivering multiple rounds of fixes and optimizations that keep the custom-op test pipeline green:

  • [CI][Misc] Add timeout check for custom op CI and optimize test parameters #8755
  • [CI] repair custom ops ci #9465
  • [CI] repari nightly custom op #9670
  • [CI] repair nightly custom op ci #9827

✅ Community Involvement

He actively helps users in 143 issues (124 open + 19 closed), tracking and following up on problem reports, answering questions and guiding users through troubleshooting across custom-op, CI, and SDK compatibility issues.

[wangxiyuan]

@weiguihua2

• ✅ Review Quality:
  He has reviewed 66+ PRs since June 2025, covering PCP/DCP, graph mode, CI, distributed serving, and model-related changes.

His review participation is closely connected to the PCP/DCP and long-context feature area, where he has continuously helped validate feature compatibility, stability fixes, and regression handling.

• ✅ Sustained Contributions:
  He has 58+ merged PRs since June 2025, with continuous contributions across multiple release cycles.

• ✅ Quality Contributions:

Context Parallelism — Core Contributor:
He implemented DeepSeek-V3.2 PCP support (#6733), then extended it to work with MTP and chunked prefill (#6917). He also enabled cross-machine prefill/decode disaggregation (#5008) and added ACLGraph support for DCP/PCP, including MLA attention v1 (#3731).

Model Compatibility & Distributed Inference Stability:
He extended Qwen3.5 PCP support with FlashComm1 (#10596), fixed hybrid block-table out-of-range failures in PCP/DCP scenarios (#9724), and resolved a DCP+DP service-stall issue caused by inconsistent MTP execution across DP ranks (#12036).

Long-Context Enablement, Testing & Documentation:
He added a multi-node long-sequence DeepSeek tutorial for context parallelism (#5364) and PCP accuracy E2E coverage (#5141). These contributions make distributed long-context features easier to validate and use.

• ✅ Community Involvement:
  He has participated in 125+ issue discussions. He actively follows up on PCP/DCP compatibility, CI failures, long-context deployment, and distributed-service stability issues.

[github-actions]

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