[0.9.1][Feature]Moe alltoallv communication optimization for unquantized RL training sence & alltoallv support dpo #1547

weijinqian0 · 2025-07-01T05:53:06Z

[Feature]Moe alltoallv communication optimization for unquantized RL training sence & alltoallv support dpo

Introduction

This PR introduces two key optimizations for MoE model performance:

Efficient Token Dispatcher:
- Implements an optimized alltoallv_seq token dispatcher (adopted from NVIDIA Megatron and Ascend MindSpeed)
- Significantly more efficient than current alltoall implementation when using token_permute/unpermute fusion
- Enable with: VLLM_ASCEND_ENABLE_MOE_ALL2ALL_SEQ=1
DBO Support for alltoallv_seq:
- Builds upon the alltoallv_seq dispatcher to support DBO (Dual Batch Overlap)
- Enables overlapping of alltoallv communication during the prefilling stage
- Enable with both:
  - VLLM_ASCEND_ENABLE_MOE_ALL2ALL_SEQ=1
  - VLLM_ASCEND_ENABLE_DBO=1

Performance Improvements

Testing on Qwen3-30B-A3B shows nearly 2x throughput improvement compared to the original alltoall implementation.

…training sence & alltoallv support dpo Signed-off-by: weijinqian_v1 <[email protected]>

github-actions · 2025-07-03T01:29:02Z

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

Signed-off-by: weijinqian_v1 <[email protected]>

github-actions · 2025-07-08T10:50:24Z

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

Signed-off-by: weijinqian_v1 <[email protected]>

Signed-off-by: duyangkai <[email protected]>

…zed RL training sence & alltoallv support dpo (vllm-project#1547) Signed-off-by: weijinqian_v1 <[email protected]>

…ed RL training sence (#2088) It comes from 0.9.1dev [0.9.1][Feature]Moe alltoallv communication optimization for unquantized RL training sence & alltoallv support dpo (#1547) - vLLM version: v0.10.0 - vLLM main: vllm-project/vllm@97608dc --------- Signed-off-by: weijinqian_v1 <[email protected]> Signed-off-by: whx-sjtu <[email protected]> Signed-off-by: curryliu <[email protected]> Signed-off-by: wangli <[email protected]> Signed-off-by: ChenTaoyu-SJTU <[email protected]> Signed-off-by: taoxudonghaha <[email protected]> Signed-off-by: shen-shanshan <[email protected]> Signed-off-by: Shanshan Shen <[email protected]> Signed-off-by: leo-pony <[email protected]> Signed-off-by: wangxiyuan <[email protected]> Signed-off-by: MengqingCao <[email protected]> Co-authored-by: weijinqian_v1 <[email protected]> Co-authored-by: whx <[email protected]> Co-authored-by: curryliu <[email protected]> Co-authored-by: Li Wang <[email protected]> Co-authored-by: TaoYu Chen <[email protected]> Co-authored-by: taoxudonghaha <[email protected]> Co-authored-by: Shanshan Shen <[email protected]> Co-authored-by: leo-pony <[email protected]> Co-authored-by: wangxiyuan <[email protected]> Co-authored-by: Mengqing Cao <[email protected]>

…ed RL training sence (vllm-project#2088) It comes from 0.9.1dev [0.9.1][Feature]Moe alltoallv communication optimization for unquantized RL training sence & alltoallv support dpo (vllm-project#1547) - vLLM version: v0.10.0 - vLLM main: vllm-project/vllm@97608dc --------- Signed-off-by: weijinqian_v1 <[email protected]> Signed-off-by: whx-sjtu <[email protected]> Signed-off-by: curryliu <[email protected]> Signed-off-by: wangli <[email protected]> Signed-off-by: ChenTaoyu-SJTU <[email protected]> Signed-off-by: taoxudonghaha <[email protected]> Signed-off-by: shen-shanshan <[email protected]> Signed-off-by: Shanshan Shen <[email protected]> Signed-off-by: leo-pony <[email protected]> Signed-off-by: wangxiyuan <[email protected]> Signed-off-by: MengqingCao <[email protected]> Co-authored-by: weijinqian_v1 <[email protected]> Co-authored-by: whx <[email protected]> Co-authored-by: curryliu <[email protected]> Co-authored-by: Li Wang <[email protected]> Co-authored-by: TaoYu Chen <[email protected]> Co-authored-by: taoxudonghaha <[email protected]> Co-authored-by: Shanshan Shen <[email protected]> Co-authored-by: leo-pony <[email protected]> Co-authored-by: wangxiyuan <[email protected]> Co-authored-by: Mengqing Cao <[email protected]>

…nalinaly

…nalinaly (#3406) I'd like to nominate 4 new maintainers for vllm-ascend: ---- Yizhou Liu [@yiz-liu](https://github.com/yiz-liu) ---- **Review Quality‌**: He has completed [40+ reviews](https://github.com/vllm-project/vllm-ascend/pulls?q=is%3Apr+commenter%3Ayiz-liu) and provided solutions or guides for [10+ issues](https://github.com/vllm-project/vllm-ascend/issues?q=is%3Aissue%20commenter%3Ayiz-liu), which includes many quality review like [#issue-3428408401](#3002 (comment)), [#discussion_r2224572309](#1803 (comment)), [#issuecomment-2982470226](#1261 (comment)), [#issuecomment-2903621197](#836 (comment)), [#issuecomment-2857678691](#778 (comment)). **Sustained and High-Quality Contributions:** He has contributed more than [30+ commits](https://github.com/vllm-project/vllm-ascend/commits?author=yiz-liu) since Mar.2025, especially, aclgraph, DP, and EP related contributions are the main reason why I nominated him. As the owner of aclgraph support, he continuously improves aclgraph stability and performance as well as fixes key bugs. he laid the groundwork for EP-related functionality and delivered multiple foundational improvements **Community involvement:** He has a very good habit of logging issues：#1649 and is also very active and involved in [many issues](https://github.com/vllm-project/vllm-ascend/issues?q=is%3Aissue%20state%3Aopen%20commenter%3Ayiz-liu%20-author%3Ayiz-liu) to help users resolve issues. ---- Peng Yu [@paulyu12](https://github.com/paulyu12) --- The main reasons for his nomination are his expertise and key contributions to the LORA and sustained and major contributions (initial support/doc/bugfix) around Lora. **Sustained and Major Contributions:** @paulyu12 starts his contribution with [Lora and Mulit-Lora support](697908f) since Apr 2025, he contributed about [10+ commits and bugfixes](697908f) on vllm-ascend. **Review Quality‌ and Community Involvement‌:** He also helped more than 10+ users address [Lora related issues](https://github.com/vllm-project/vllm-ascend/pulls?q=is%3Apr+commenter%3Apaulyu12+-author%3Apaulyu12+is%3Aclosed). I believe his addition will further improve vLLM Ascend Lora support. ---- Jinqian Wei [@weijinqian0](https://github.com/weijinqian0) --- The main reasons for his nomination are his key contributions to the RL scene and the high quality of his code reviews. **Review Quality‌:** He has completed [60+ reviews](https://github.com/vllm-project/vllm-ascend/pulls?q=is%3Apr+commenter%3Aweijinqian0+is%3Aopen+-author%3Aweijinqian0) since June. 2025, include [#comment-3284055430](#2791 (comment)), [discussion_r2332166704](#2817 (comment)), [discussion_r2343289692](#2846 (comment)) high quality review. **Sustained and Quality Contributions:** He has Deep understanding of ‌vLLM‌ and ‌vLLM Ascend‌ codebases and solid contributions in RL scene (about [10+ PR merged](https://github.com/vllm-project/vllm-ascend/pulls?q=is%3Apr+author%3Aweijinqian0+is%3Amerged+) and 10+ PRs merged as co-author. - Code Refactor: As a co-author, he participated in the refactoring of the MOE module #2150 #2706 #2867 - Performance Enhancement for RL: Participated as a co-author in the design and development of the solution, contributing to the planning of core capabilities. #1547 #2120 and so on. So I think he's a great addition to the vLLM Ascend Maintainer team. ---- Chuanyu Qin [@nalinaly](https://github.com/nalinaly) --- The main reason I nominated Qinchuanyu is because he is the initial designer of aclgraph and torch-npu, two key components of vllm-ascend. Considering aclgraph will eventually become the main path for vllm-ascend's graph model, I propose to nominate him. **Sustained and Major Contributions:** In fact, chuanyu actively helped the users/developers of vllm-ascend since Mar 2025 ([vllm-discuss#162](https://discuss.vllm.ai/t/can-ascend-officially-draft-a-documentation-on-the-vllm-ascend-adaptation-for-graph-mode/162/5)), and also helped early users of vllm-ascend understand aclgraph. He provided lots of help in the process of integrating aclgraph with vllm-ascend. **Community Involvement‌:** As speaker, he also presents help users understand aclgraph and torch_npu [《The design philosophy of torch_npu and the high performance principle of aclGraph》](https://github.com/PyTorch-China/pytorch-meetup/blob/main/beijing-2025/%E3%80%905%E3%80%91torch_npu%20%E7%9A%84%E8%AE%BE%E8%AE%A1%E5%93%B2%E5%AD%A6%E4%B8%8E%20aclGraph%20%E9%AB%98%E6%80%A7%E8%83%BD%E5%8E%9F%E7%90%86-%E7%A7%A6%E4%BC%A0%E7%91%9C-0920.pdf) ---- They have activate contribution to vllm-ascend or have rich experience for ascend AI. Welcome! - vLLM version: v0.11.0rc3 - vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0 Signed-off-by: wangxiyuan <[email protected]>

…ed RL training sence (vllm-project#2088) It comes from 0.9.1dev [0.9.1][Feature]Moe alltoallv communication optimization for unquantized RL training sence & alltoallv support dpo (vllm-project#1547) - vLLM version: v0.10.0 - vLLM main: vllm-project/vllm@97608dc --------- Signed-off-by: weijinqian_v1 <[email protected]> Signed-off-by: whx-sjtu <[email protected]> Signed-off-by: curryliu <[email protected]> Signed-off-by: wangli <[email protected]> Signed-off-by: ChenTaoyu-SJTU <[email protected]> Signed-off-by: taoxudonghaha <[email protected]> Signed-off-by: shen-shanshan <[email protected]> Signed-off-by: Shanshan Shen <[email protected]> Signed-off-by: leo-pony <[email protected]> Signed-off-by: wangxiyuan <[email protected]> Signed-off-by: MengqingCao <[email protected]> Co-authored-by: weijinqian_v1 <[email protected]> Co-authored-by: whx <[email protected]> Co-authored-by: curryliu <[email protected]> Co-authored-by: Li Wang <[email protected]> Co-authored-by: TaoYu Chen <[email protected]> Co-authored-by: taoxudonghaha <[email protected]> Co-authored-by: Shanshan Shen <[email protected]> Co-authored-by: leo-pony <[email protected]> Co-authored-by: wangxiyuan <[email protected]> Co-authored-by: Mengqing Cao <[email protected]>

…nalinaly

…nalinaly (vllm-project#3406) I'd like to nominate 4 new maintainers for vllm-ascend: ---- Yizhou Liu [@yiz-liu](https://github.com/yiz-liu) ---- **Review Quality‌**: He has completed [40+ reviews](https://github.com/vllm-project/vllm-ascend/pulls?q=is%3Apr+commenter%3Ayiz-liu) and provided solutions or guides for [10+ issues](https://github.com/vllm-project/vllm-ascend/issues?q=is%3Aissue%20commenter%3Ayiz-liu), which includes many quality review like [#issue-3428408401](vllm-project#3002 (comment)), [#discussion_r2224572309](vllm-project#1803 (comment)), [#issuecomment-2982470226](vllm-project#1261 (comment)), [#issuecomment-2903621197](vllm-project#836 (comment)), [#issuecomment-2857678691](vllm-project#778 (comment)). **Sustained and High-Quality Contributions:** He has contributed more than [30+ commits](https://github.com/vllm-project/vllm-ascend/commits?author=yiz-liu) since Mar.2025, especially, aclgraph, DP, and EP related contributions are the main reason why I nominated him. As the owner of aclgraph support, he continuously improves aclgraph stability and performance as well as fixes key bugs. he laid the groundwork for EP-related functionality and delivered multiple foundational improvements **Community involvement:** He has a very good habit of logging issues：vllm-project#1649 and is also very active and involved in [many issues](https://github.com/vllm-project/vllm-ascend/issues?q=is%3Aissue%20state%3Aopen%20commenter%3Ayiz-liu%20-author%3Ayiz-liu) to help users resolve issues. ---- Peng Yu [@paulyu12](https://github.com/paulyu12) --- The main reasons for his nomination are his expertise and key contributions to the LORA and sustained and major contributions (initial support/doc/bugfix) around Lora. **Sustained and Major Contributions:** @paulyu12 starts his contribution with [Lora and Mulit-Lora support](vllm-project@697908f) since Apr 2025, he contributed about [10+ commits and bugfixes](vllm-project@697908f) on vllm-ascend. **Review Quality‌ and Community Involvement‌:** He also helped more than 10+ users address [Lora related issues](https://github.com/vllm-project/vllm-ascend/pulls?q=is%3Apr+commenter%3Apaulyu12+-author%3Apaulyu12+is%3Aclosed). I believe his addition will further improve vLLM Ascend Lora support. ---- Jinqian Wei [@weijinqian0](https://github.com/weijinqian0) --- The main reasons for his nomination are his key contributions to the RL scene and the high quality of his code reviews. **Review Quality‌:** He has completed [60+ reviews](https://github.com/vllm-project/vllm-ascend/pulls?q=is%3Apr+commenter%3Aweijinqian0+is%3Aopen+-author%3Aweijinqian0) since June. 2025, include [#comment-3284055430](vllm-project#2791 (comment)), [discussion_r2332166704](vllm-project#2817 (comment)), [discussion_r2343289692](vllm-project#2846 (comment)) high quality review. **Sustained and Quality Contributions:** He has Deep understanding of ‌vLLM‌ and ‌vLLM Ascend‌ codebases and solid contributions in RL scene (about [10+ PR merged](https://github.com/vllm-project/vllm-ascend/pulls?q=is%3Apr+author%3Aweijinqian0+is%3Amerged+) and 10+ PRs merged as co-author. - Code Refactor: As a co-author, he participated in the refactoring of the MOE module vllm-project#2150 vllm-project#2706 vllm-project#2867 - Performance Enhancement for RL: Participated as a co-author in the design and development of the solution, contributing to the planning of core capabilities. vllm-project#1547 vllm-project#2120 and so on. So I think he's a great addition to the vLLM Ascend Maintainer team. ---- Chuanyu Qin [@nalinaly](https://github.com/nalinaly) --- The main reason I nominated Qinchuanyu is because he is the initial designer of aclgraph and torch-npu, two key components of vllm-ascend. Considering aclgraph will eventually become the main path for vllm-ascend's graph model, I propose to nominate him. **Sustained and Major Contributions:** In fact, chuanyu actively helped the users/developers of vllm-ascend since Mar 2025 ([vllm-discuss#162](https://discuss.vllm.ai/t/can-ascend-officially-draft-a-documentation-on-the-vllm-ascend-adaptation-for-graph-mode/162/5)), and also helped early users of vllm-ascend understand aclgraph. He provided lots of help in the process of integrating aclgraph with vllm-ascend. **Community Involvement‌:** As speaker, he also presents help users understand aclgraph and torch_npu [《The design philosophy of torch_npu and the high performance principle of aclGraph》](https://github.com/PyTorch-China/pytorch-meetup/blob/main/beijing-2025/%E3%80%905%E3%80%91torch_npu%20%E7%9A%84%E8%AE%BE%E8%AE%A1%E5%93%B2%E5%AD%A6%E4%B8%8E%20aclGraph%20%E9%AB%98%E6%80%A7%E8%83%BD%E5%8E%9F%E7%90%86-%E7%A7%A6%E4%BC%A0%E7%91%9C-0920.pdf) ---- They have activate contribution to vllm-ascend or have rich experience for ascend AI. Welcome! - vLLM version: v0.11.0rc3 - vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0 Signed-off-by: wangxiyuan <[email protected]>

…nalinaly

…nalinaly (vllm-project#3406) I'd like to nominate 4 new maintainers for vllm-ascend: ---- Yizhou Liu [@yiz-liu](https://github.com/yiz-liu) ---- **Review Quality‌**: He has completed [40+ reviews](https://github.com/vllm-project/vllm-ascend/pulls?q=is%3Apr+commenter%3Ayiz-liu) and provided solutions or guides for [10+ issues](https://github.com/vllm-project/vllm-ascend/issues?q=is%3Aissue%20commenter%3Ayiz-liu), which includes many quality review like [#issue-3428408401](vllm-project#3002 (comment)), [#discussion_r2224572309](vllm-project#1803 (comment)), [#issuecomment-2982470226](vllm-project#1261 (comment)), [#issuecomment-2903621197](vllm-project#836 (comment)), [#issuecomment-2857678691](vllm-project#778 (comment)). **Sustained and High-Quality Contributions:** He has contributed more than [30+ commits](https://github.com/vllm-project/vllm-ascend/commits?author=yiz-liu) since Mar.2025, especially, aclgraph, DP, and EP related contributions are the main reason why I nominated him. As the owner of aclgraph support, he continuously improves aclgraph stability and performance as well as fixes key bugs. he laid the groundwork for EP-related functionality and delivered multiple foundational improvements **Community involvement:** He has a very good habit of logging issues：vllm-project#1649 and is also very active and involved in [many issues](https://github.com/vllm-project/vllm-ascend/issues?q=is%3Aissue%20state%3Aopen%20commenter%3Ayiz-liu%20-author%3Ayiz-liu) to help users resolve issues. ---- Peng Yu [@paulyu12](https://github.com/paulyu12) --- The main reasons for his nomination are his expertise and key contributions to the LORA and sustained and major contributions (initial support/doc/bugfix) around Lora. **Sustained and Major Contributions:** @paulyu12 starts his contribution with [Lora and Mulit-Lora support](vllm-project@697908f) since Apr 2025, he contributed about [10+ commits and bugfixes](vllm-project@697908f) on vllm-ascend. **Review Quality‌ and Community Involvement‌:** He also helped more than 10+ users address [Lora related issues](https://github.com/vllm-project/vllm-ascend/pulls?q=is%3Apr+commenter%3Apaulyu12+-author%3Apaulyu12+is%3Aclosed). I believe his addition will further improve vLLM Ascend Lora support. ---- Jinqian Wei [@weijinqian0](https://github.com/weijinqian0) --- The main reasons for his nomination are his key contributions to the RL scene and the high quality of his code reviews. **Review Quality‌:** He has completed [60+ reviews](https://github.com/vllm-project/vllm-ascend/pulls?q=is%3Apr+commenter%3Aweijinqian0+is%3Aopen+-author%3Aweijinqian0) since June. 2025, include [#comment-3284055430](vllm-project#2791 (comment)), [discussion_r2332166704](vllm-project#2817 (comment)), [discussion_r2343289692](vllm-project#2846 (comment)) high quality review. **Sustained and Quality Contributions:** He has Deep understanding of ‌vLLM‌ and ‌vLLM Ascend‌ codebases and solid contributions in RL scene (about [10+ PR merged](https://github.com/vllm-project/vllm-ascend/pulls?q=is%3Apr+author%3Aweijinqian0+is%3Amerged+) and 10+ PRs merged as co-author. - Code Refactor: As a co-author, he participated in the refactoring of the MOE module vllm-project#2150 vllm-project#2706 vllm-project#2867 - Performance Enhancement for RL: Participated as a co-author in the design and development of the solution, contributing to the planning of core capabilities. vllm-project#1547 vllm-project#2120 and so on. So I think he's a great addition to the vLLM Ascend Maintainer team. ---- Chuanyu Qin [@nalinaly](https://github.com/nalinaly) --- The main reason I nominated Qinchuanyu is because he is the initial designer of aclgraph and torch-npu, two key components of vllm-ascend. Considering aclgraph will eventually become the main path for vllm-ascend's graph model, I propose to nominate him. **Sustained and Major Contributions:** In fact, chuanyu actively helped the users/developers of vllm-ascend since Mar 2025 ([vllm-discuss#162](https://discuss.vllm.ai/t/can-ascend-officially-draft-a-documentation-on-the-vllm-ascend-adaptation-for-graph-mode/162/5)), and also helped early users of vllm-ascend understand aclgraph. He provided lots of help in the process of integrating aclgraph with vllm-ascend. **Community Involvement‌:** As speaker, he also presents help users understand aclgraph and torch_npu [《The design philosophy of torch_npu and the high performance principle of aclGraph》](https://github.com/PyTorch-China/pytorch-meetup/blob/main/beijing-2025/%E3%80%905%E3%80%91torch_npu%20%E7%9A%84%E8%AE%BE%E8%AE%A1%E5%93%B2%E5%AD%A6%E4%B8%8E%20aclGraph%20%E9%AB%98%E6%80%A7%E8%83%BD%E5%8E%9F%E7%90%86-%E7%A7%A6%E4%BC%A0%E7%91%9C-0920.pdf) ---- They have activate contribution to vllm-ascend or have rich experience for ascend AI. Welcome! - vLLM version: v0.11.0rc3 - vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0 Signed-off-by: wangxiyuan <[email protected]> Signed-off-by: luolun <[email protected]>

…nalinaly

…nalinaly (vllm-project#3406) I'd like to nominate 4 new maintainers for vllm-ascend: ---- Yizhou Liu [@yiz-liu](https://github.com/yiz-liu) ---- **Review Quality‌**: He has completed [40+ reviews](https://github.com/vllm-project/vllm-ascend/pulls?q=is%3Apr+commenter%3Ayiz-liu) and provided solutions or guides for [10+ issues](https://github.com/vllm-project/vllm-ascend/issues?q=is%3Aissue%20commenter%3Ayiz-liu), which includes many quality review like [#issue-3428408401](vllm-project#3002 (comment)), [#discussion_r2224572309](vllm-project#1803 (comment)), [#issuecomment-2982470226](vllm-project#1261 (comment)), [#issuecomment-2903621197](vllm-project#836 (comment)), [#issuecomment-2857678691](vllm-project#778 (comment)). **Sustained and High-Quality Contributions:** He has contributed more than [30+ commits](https://github.com/vllm-project/vllm-ascend/commits?author=yiz-liu) since Mar.2025, especially, aclgraph, DP, and EP related contributions are the main reason why I nominated him. As the owner of aclgraph support, he continuously improves aclgraph stability and performance as well as fixes key bugs. he laid the groundwork for EP-related functionality and delivered multiple foundational improvements **Community involvement:** He has a very good habit of logging issues：vllm-project#1649 and is also very active and involved in [many issues](https://github.com/vllm-project/vllm-ascend/issues?q=is%3Aissue%20state%3Aopen%20commenter%3Ayiz-liu%20-author%3Ayiz-liu) to help users resolve issues. ---- Peng Yu [@paulyu12](https://github.com/paulyu12) --- The main reasons for his nomination are his expertise and key contributions to the LORA and sustained and major contributions (initial support/doc/bugfix) around Lora. **Sustained and Major Contributions:** @paulyu12 starts his contribution with [Lora and Mulit-Lora support](vllm-project@697908f) since Apr 2025, he contributed about [10+ commits and bugfixes](vllm-project@697908f) on vllm-ascend. **Review Quality‌ and Community Involvement‌:** He also helped more than 10+ users address [Lora related issues](https://github.com/vllm-project/vllm-ascend/pulls?q=is%3Apr+commenter%3Apaulyu12+-author%3Apaulyu12+is%3Aclosed). I believe his addition will further improve vLLM Ascend Lora support. ---- Jinqian Wei [@weijinqian0](https://github.com/weijinqian0) --- The main reasons for his nomination are his key contributions to the RL scene and the high quality of his code reviews. **Review Quality‌:** He has completed [60+ reviews](https://github.com/vllm-project/vllm-ascend/pulls?q=is%3Apr+commenter%3Aweijinqian0+is%3Aopen+-author%3Aweijinqian0) since June. 2025, include [#comment-3284055430](vllm-project#2791 (comment)), [discussion_r2332166704](vllm-project#2817 (comment)), [discussion_r2343289692](vllm-project#2846 (comment)) high quality review. **Sustained and Quality Contributions:** He has Deep understanding of ‌vLLM‌ and ‌vLLM Ascend‌ codebases and solid contributions in RL scene (about [10+ PR merged](https://github.com/vllm-project/vllm-ascend/pulls?q=is%3Apr+author%3Aweijinqian0+is%3Amerged+) and 10+ PRs merged as co-author. - Code Refactor: As a co-author, he participated in the refactoring of the MOE module vllm-project#2150 vllm-project#2706 vllm-project#2867 - Performance Enhancement for RL: Participated as a co-author in the design and development of the solution, contributing to the planning of core capabilities. vllm-project#1547 vllm-project#2120 and so on. So I think he's a great addition to the vLLM Ascend Maintainer team. ---- Chuanyu Qin [@nalinaly](https://github.com/nalinaly) --- The main reason I nominated Qinchuanyu is because he is the initial designer of aclgraph and torch-npu, two key components of vllm-ascend. Considering aclgraph will eventually become the main path for vllm-ascend's graph model, I propose to nominate him. **Sustained and Major Contributions:** In fact, chuanyu actively helped the users/developers of vllm-ascend since Mar 2025 ([vllm-discuss#162](https://discuss.vllm.ai/t/can-ascend-officially-draft-a-documentation-on-the-vllm-ascend-adaptation-for-graph-mode/162/5)), and also helped early users of vllm-ascend understand aclgraph. He provided lots of help in the process of integrating aclgraph with vllm-ascend. **Community Involvement‌:** As speaker, he also presents help users understand aclgraph and torch_npu [《The design philosophy of torch_npu and the high performance principle of aclGraph》](https://github.com/PyTorch-China/pytorch-meetup/blob/main/beijing-2025/%E3%80%905%E3%80%91torch_npu%20%E7%9A%84%E8%AE%BE%E8%AE%A1%E5%93%B2%E5%AD%A6%E4%B8%8E%20aclGraph%20%E9%AB%98%E6%80%A7%E8%83%BD%E5%8E%9F%E7%90%86-%E7%A7%A6%E4%BC%A0%E7%91%9C-0920.pdf) ---- They have activate contribution to vllm-ascend or have rich experience for ascend AI. Welcome! - vLLM version: v0.11.0rc3 - vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0 Signed-off-by: wangxiyuan <[email protected]> Signed-off-by: hwhaokun <[email protected]>

…nalinaly

…nalinaly (vllm-project#3406) I'd like to nominate 4 new maintainers for vllm-ascend: ---- Yizhou Liu [@yiz-liu](https://github.com/yiz-liu) ---- **Review Quality‌**: He has completed [40+ reviews](https://github.com/vllm-project/vllm-ascend/pulls?q=is%3Apr+commenter%3Ayiz-liu) and provided solutions or guides for [10+ issues](https://github.com/vllm-project/vllm-ascend/issues?q=is%3Aissue%20commenter%3Ayiz-liu), which includes many quality review like [#issue-3428408401](vllm-project#3002 (comment)), [#discussion_r2224572309](vllm-project#1803 (comment)), [#issuecomment-2982470226](vllm-project#1261 (comment)), [#issuecomment-2903621197](vllm-project#836 (comment)), [#issuecomment-2857678691](vllm-project#778 (comment)). **Sustained and High-Quality Contributions:** He has contributed more than [30+ commits](https://github.com/vllm-project/vllm-ascend/commits?author=yiz-liu) since Mar.2025, especially, aclgraph, DP, and EP related contributions are the main reason why I nominated him. As the owner of aclgraph support, he continuously improves aclgraph stability and performance as well as fixes key bugs. he laid the groundwork for EP-related functionality and delivered multiple foundational improvements **Community involvement:** He has a very good habit of logging issues：vllm-project#1649 and is also very active and involved in [many issues](https://github.com/vllm-project/vllm-ascend/issues?q=is%3Aissue%20state%3Aopen%20commenter%3Ayiz-liu%20-author%3Ayiz-liu) to help users resolve issues. ---- Peng Yu [@paulyu12](https://github.com/paulyu12) --- The main reasons for his nomination are his expertise and key contributions to the LORA and sustained and major contributions (initial support/doc/bugfix) around Lora. **Sustained and Major Contributions:** @paulyu12 starts his contribution with [Lora and Mulit-Lora support](vllm-project@697908f) since Apr 2025, he contributed about [10+ commits and bugfixes](vllm-project@697908f) on vllm-ascend. **Review Quality‌ and Community Involvement‌:** He also helped more than 10+ users address [Lora related issues](https://github.com/vllm-project/vllm-ascend/pulls?q=is%3Apr+commenter%3Apaulyu12+-author%3Apaulyu12+is%3Aclosed). I believe his addition will further improve vLLM Ascend Lora support. ---- Jinqian Wei [@weijinqian0](https://github.com/weijinqian0) --- The main reasons for his nomination are his key contributions to the RL scene and the high quality of his code reviews. **Review Quality‌:** He has completed [60+ reviews](https://github.com/vllm-project/vllm-ascend/pulls?q=is%3Apr+commenter%3Aweijinqian0+is%3Aopen+-author%3Aweijinqian0) since June. 2025, include [#comment-3284055430](vllm-project#2791 (comment)), [discussion_r2332166704](vllm-project#2817 (comment)), [discussion_r2343289692](vllm-project#2846 (comment)) high quality review. **Sustained and Quality Contributions:** He has Deep understanding of ‌vLLM‌ and ‌vLLM Ascend‌ codebases and solid contributions in RL scene (about [10+ PR merged](https://github.com/vllm-project/vllm-ascend/pulls?q=is%3Apr+author%3Aweijinqian0+is%3Amerged+) and 10+ PRs merged as co-author. - Code Refactor: As a co-author, he participated in the refactoring of the MOE module vllm-project#2150 vllm-project#2706 vllm-project#2867 - Performance Enhancement for RL: Participated as a co-author in the design and development of the solution, contributing to the planning of core capabilities. vllm-project#1547 vllm-project#2120 and so on. So I think he's a great addition to the vLLM Ascend Maintainer team. ---- Chuanyu Qin [@nalinaly](https://github.com/nalinaly) --- The main reason I nominated Qinchuanyu is because he is the initial designer of aclgraph and torch-npu, two key components of vllm-ascend. Considering aclgraph will eventually become the main path for vllm-ascend's graph model, I propose to nominate him. **Sustained and Major Contributions:** In fact, chuanyu actively helped the users/developers of vllm-ascend since Mar 2025 ([vllm-discuss#162](https://discuss.vllm.ai/t/can-ascend-officially-draft-a-documentation-on-the-vllm-ascend-adaptation-for-graph-mode/162/5)), and also helped early users of vllm-ascend understand aclgraph. He provided lots of help in the process of integrating aclgraph with vllm-ascend. **Community Involvement‌:** As speaker, he also presents help users understand aclgraph and torch_npu [《The design philosophy of torch_npu and the high performance principle of aclGraph》](https://github.com/PyTorch-China/pytorch-meetup/blob/main/beijing-2025/%E3%80%905%E3%80%91torch_npu%20%E7%9A%84%E8%AE%BE%E8%AE%A1%E5%93%B2%E5%AD%A6%E4%B8%8E%20aclGraph%20%E9%AB%98%E6%80%A7%E8%83%BD%E5%8E%9F%E7%90%86-%E7%A7%A6%E4%BC%A0%E7%91%9C-0920.pdf) ---- They have activate contribution to vllm-ascend or have rich experience for ascend AI. Welcome! - vLLM version: v0.11.0rc3 - vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0 Signed-off-by: wangxiyuan <[email protected]> Signed-off-by: nsdie <[email protected]>

weijinqian_v1 added 12 commits July 1, 2025 09:51

[Feature]Moe alltoallv communication optimization for unquantized RL …

7ff288e

…training sence & alltoallv support dpo Signed-off-by: weijinqian_v1 <[email protected]>

[Feature]Moe alltoallv communication optimization for unquantized RL …

6d7b5b4

…training sence & alltoallv support dpo Signed-off-by: weijinqian_v1 <[email protected]>

[Feature]Moe alltoallv communication optimization for unquantized RL …

6a8e1a9

…training sence & alltoallv support dpo Signed-off-by: weijinqian_v1 <[email protected]>

[Feature]Moe alltoallv communication optimization for unquantized RL …

4805c5a

…training sence & alltoallv support dpo Signed-off-by: weijinqian_v1 <[email protected]>

[Feature]Moe alltoallv communication optimization for unquantized RL …

d68ce07

…training sence & alltoallv support dpo Signed-off-by: weijinqian_v1 <[email protected]>

[Feature]Moe alltoallv communication optimization for unquantized RL …

0aff693

…training sence & alltoallv support dpo Signed-off-by: weijinqian_v1 <[email protected]>

[Feature]Moe alltoallv communication optimization for unquantized RL …

f6ab19e

…training sence & alltoallv support dpo Signed-off-by: weijinqian_v1 <[email protected]>

[Feature]Moe alltoallv communication optimization for unquantized RL …

a94c094

…training sence & alltoallv support dpo Signed-off-by: weijinqian_v1 <[email protected]>

[Feature]Moe alltoallv communication optimization for unquantized RL …

91570d8

…training sence & alltoallv support dpo Signed-off-by: weijinqian_v1 <[email protected]>

[Feature]Moe alltoallv communication optimization for unquantized RL …

e7c0d2d

…training sence & alltoallv support dpo Signed-off-by: weijinqian_v1 <[email protected]>

[Feature]Moe alltoallv communication optimization for unquantized RL …

47439e8

…training sence & alltoallv support dpo Signed-off-by: weijinqian_v1 <[email protected]>

[Feature]Moe alltoallv communication optimization for unquantized RL …

cf3f1c8

…training sence & alltoallv support dpo Signed-off-by: weijinqian_v1 <[email protected]>

github-actions bot added module:tests module:ops module:core labels Jul 1, 2025

weijinqian_v1 added 3 commits July 1, 2025 14:03

[Feature]Moe alltoallv communication optimization for unquantized RL …

a4126f3

…training sence & alltoallv support dpo Signed-off-by: weijinqian_v1 <[email protected]>

[Feature]Moe alltoallv communication optimization for unquantized RL …

807aaf0

…training sence & alltoallv support dpo Signed-off-by: weijinqian_v1 <[email protected]>

[Feature]Moe alltoallv communication optimization for unquantized RL …

6f6efc1

…training sence & alltoallv support dpo Signed-off-by: weijinqian_v1 <[email protected]>

github-actions bot added the merge-conflicts label Jul 3, 2025

handle conflict

305a0eb

Signed-off-by: weijinqian_v1 <[email protected]>

github-actions bot added merge-conflicts and removed merge-conflicts labels Jul 8, 2025

github-actions bot removed the merge-conflicts label Jul 8, 2025

weijinqian_v1 and others added 5 commits July 9, 2025 16:25

add st:qwen3

5411ed6

Signed-off-by: weijinqian_v1 <[email protected]>

add st for moe token dispatcher

3f88769

Signed-off-by: weijinqian_v1 <[email protected]>

fix bug

854c149

Signed-off-by: weijinqian_v1 <[email protected]>

add st for moe token dispatcher

d0bd006

Signed-off-by: weijinqian_v1 <[email protected]>

add moe_block: AscendSparseMoeBlock

49e9771

Signed-off-by: weijinqian_v1 <[email protected]>

weijinqian added 8 commits July 12, 2025 23:43

handle clean code

267db60

Signed-off-by: weijinqian_v1 <[email protected]>

handle clean code

b0572c8

Signed-off-by: weijinqian_v1 <[email protected]>

handle clean code

e4f1050

Signed-off-by: weijinqian_v1 <[email protected]>

handle clean code

eaed83d

Signed-off-by: weijinqian_v1 <[email protected]>

handle clean code

b97baf4

Signed-off-by: weijinqian_v1 <[email protected]>

handle clean code

d232d49

Signed-off-by: weijinqian_v1 <[email protected]>

handle clean code

1e435e6

Signed-off-by: weijinqian_v1 <[email protected]>

handle clean code

8effdd0

Signed-off-by: weijinqian_v1 <[email protected]>

weijinqian0 force-pushed the v0.9.1-dev branch from 2ae837f to 8effdd0 Compare July 12, 2025 15:44

weijinqian_v1 added 2 commits July 12, 2025 23:44

handle code clean

a8136b7

Signed-off-by: weijinqian_v1 <[email protected]>

handle code conflict

d29deae

Signed-off-by: weijinqian_v1 <[email protected]>

github-actions bot removed the merge-conflicts label Jul 12, 2025

weijinqian_v1 and others added 2 commits July 12, 2025 23:52

handle code clean

94b7b5b

Signed-off-by: weijinqian_v1 <[email protected]>

fix header

c2f670d

Signed-off-by: duyangkai <[email protected]>

ganyi1996ppo approved these changes Jul 14, 2025

View reviewed changes

ganyi1996ppo merged commit 63944db into vllm-project:v0.9.1-dev Jul 14, 2025
16 checks passed

wangxiyuan added the no-main label Jul 14, 2025

weijinqian0 mentioned this pull request Jul 29, 2025

[main][Feature]Moe alltoallv communication optimization for unquantized RL training sence #2088

Merged

wangxiyuan mentioned this pull request Oct 13, 2025

[Community] Nominate new maintainers: @yiz-liu @paulyu12 @weijinqian0 @nalinaly #3406

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[0.9.1][Feature]Moe alltoallv communication optimization for unquantized RL training sence & alltoallv support dpo #1547

[0.9.1][Feature]Moe alltoallv communication optimization for unquantized RL training sence & alltoallv support dpo #1547

Uh oh!

weijinqian0 commented Jul 1, 2025 •

edited

Loading

Uh oh!

github-actions bot commented Jul 3, 2025

Uh oh!

github-actions bot commented Jul 8, 2025

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

9 participants

[0.9.1][Feature]Moe alltoallv communication optimization for unquantized RL training sence & alltoallv support dpo #1547

[0.9.1][Feature]Moe alltoallv communication optimization for unquantized RL training sence & alltoallv support dpo #1547

Uh oh!

Conversation

weijinqian0 commented Jul 1, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Introduction

Performance Improvements

Uh oh!

github-actions bot commented Jul 3, 2025

Uh oh!

github-actions bot commented Jul 8, 2025

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

9 participants

weijinqian0 commented Jul 1, 2025 •

edited

Loading