Chair: @paulyu12
Topic:

1. vLLM Weekly update improvement @wangxiyuan
2. model improvement sync @ Jun Han
3. Flash Comm v1 for VL models @realliujiaxu
4. KVComp长序列稀疏化算法——GQA全显存方案

Chair: @jianzs
Topic:

1. EPLB：Ascend vs. Community - @shenchuxiaofugui
2. [interface] User interface for flashcomm2 and layer_shard features - @zzhx1
3. vLLM-Ascend support HMA and cross layer KVCache - @zzzzwwjj

Chair: @wangxiyuan
Topic:

1. vLLM Community weekly Update - @david6666666
   https://docs.google.com/presentation/d/1_qBc9BhK4baQ9jeGraI1JpQiHGcjMaVf2KW7gjjkHFs/edit?slide=id.g399750550f0_1_0#slide=id.g399750550f0_1_0
2. vLLM Ascend progress @ Jun Han
3. Async scheduler @Ronald1995
4. FA refactor @weijinqian0
   [RFC]: Refactor Attention module #4629
5. Custom ops @zzzzwwjj

Chair: Jun Han
Topic:

1. vLLM Community weekly Update - @MengqingCao
2. vllm-ascend support sequence parallel for DS v3.2 model - @AlvisGong
3. vllm-ascend VL modeling files removing - @shen-shanshan
4. MindSpore model support - @wangtiance
5. Add any topick below

Chair: JinQi Yu
Topic:

1. vLLM Community weekly Update - @david6666666
2. vllm-ascend support dump data in eager mode - @Tjh-UKN
3. In vllm-ascend, KV Pool Elimination of Redundancy - @baxingpiaochong
4. vllm-ascend support xlite graph wrapper - @lulina

Chair: @realliujiaxu
Topic:

1. vLLM Community weekly Update - @wangxiyuan
2. vllm-ascend mla operator support all gather the result of W8A8_matmul(hiddenstate and wdq+wqdkv) - @chenlongxiao
3. torch binding issues when adding custom ops into vllm-ascend - @ChenxiQ
4. magicmtp - QingSen Han

Chair: @paulyu12
Topic:

1. vLLM Community weekly Update - @MengqingCao

• vLLM important PRs in last week
• Q4: triton for Ascend

2. token level re-inference - Li ShiLin

• Exception handling and recovering: kvcache, weights, activate value HBM UCE failed, network package losing...
• intrusive to ModelRunner.

3. vLLM Ascend Model Support List guide revisit - Shen Xinjie

• Apply to Document. For each model,
  • introduction
  • main applied scenario for the model
  • feature support
  • link to model weight
  • link to deploy tutorial

4. vLLM Ascend Q4 RoadMap - Han Jun

• Q3
  • mooncake pooling
  • chunk prefill
  • aclgraph full graph
  • long seq
  • multi-stream parallel
  • dynamic eplb
  • w4a8 w4a4
  • kvcache resharding
  • kvcache int8 quantization
• Q4
  • system-level schedule optimization
  • hostbound between decoding process: async scheduling...
  • RAS, DFX
  • multimodal
  • triton for Ascend
  • inductor
  • auto adaptor for short-long seq
  • model support: tier 2
  • P/D quantization

Chair: @ApsarasX
Topic:

1. vLLM Community weekly Update - @david6666666
   https://docs.google.com/presentation/d/1_qBc9BhK4baQ9jeGraI1JpQiHGcjMaVf2KW7gjjkHFs/edit?slide=id.p#slide=id.p

2. vLLM Ascend Model Suport List guide - @evakang777 , Shen Xinjie
   The customer hopes that the model support list in the vllm ascend community can provide some specifications and features. Together with PAE, we have sorted out the table headers to be added to the model support list based on customer requirements. It is expected that the vllm-ascend model supports the information displayed on the customer plane in the list.

3. Qwen3-Next support the ACL Graph feature. - @xueliangyang-oeuler

4. vllm-ascend PD Separation and TP Asymmetry Solution - @liziyu179, Zhou Xuerong

   For the PD separation layerwise push + TP asymmetry solution, the overlay of DP characteristics in stress testing scenarios has shown instability. We have proposed a new solution centered on D-node scheduling.