Repository navigation
[Community] Weekly Meeting Agenda #10928
Description
We hold technical meeting on Friday 10:00am - 12:00am(UTC+8) every week. The meeting is by invitation only currently and any vllm-ascend related topic is welcome. If you have anything want to discuess, feel free to submit your topic under this issue.
Previous Agenda Link: #3642
Meeting Note: https://docs.google.com/document/d/1u_MbjsBPpXsgXJA6ckX0TgyWB6wZUxxttsbV28-m3ZM/edit?usp=sharing
2026.11.13
Chair: @ningjingbengxiaohai
- [Feature] Add reuse-prioritized LRU for sparse KV offload - @wangbiao0814
- [RFC] Standardizing how dynamic KV-cache quantization is enabled - @kunpengW-code
- [Feature][MoE] Add opt-in DSpark expert-pool routing for expert offload - @jiemeng6
History
2026.11.06
Chair: @wangxiyuan
- Make NPUWorker inherit from the vLLM GPU Worker - @Ronald1995
- Add slipcover for UT coverage checking - @wenjinhust
- Remove model-level torch.compile dependency in vLLM Ascend - @yiz-liu
- Observability items — metrics levels, dynamic log level, garbled-text detection -@djh95
- HiSparse landing in vllm-ascend — PD mixed deployment, TPOT regression, A3 KV offload - @yanghaocsg
- Add int8 quantized all-reduce for large bf16 TP collectives on Atlas A2 - @chujiangang2006
2026.10.30
Chair: @wangxiyuan
- Multi-Model Daemon — Shared Memory Pool and Elastic KV Cache on Ascend NPU - @Id-Vulkan
- Shared-Index Prefetch for Sparse KV Cache Offloading in Decode - @banyannnn
- Plugin Marketplace on Model-Layer Interfaces: Encouraging Deep Customization for Specific Models and Scenarios - @LittleJiafang
2026.10.23
Chair: @ningjingbengxiaohai
Topic:
- On-demand main-KV loading for long context and high prefix hit rate DSA prefill - @ZD-euler
- Decoder append prefill for agentic scenario prefiller offloading - @Ydr-pku
- Implementation of PIVOT for vllm-ascend - @qqq1357962
- Second Sunset Review of Existing Models in the Community Document "Supported Models" - @liqian79
- MoE Expert Weight Offloading for Single-Device Inference - @florian0125
- KV Trace for Request-Level KV Block Access Tracing and Validation - @Tjh-UKN
2026.10.16
Chair: @realliujiaxu
Topic:
- A5 AI QoS — Request-Level KV Transfer and Per-Domain Collective Communication - @A-zhudong
- AIRDrafter - @ZhouZhichengSteven
- release note progress update - @wangxiyuan
2026.9.18
Chair: @ningjingbengxiaohai
Topic:
- Virtual Pipeline Parallelism (VPP) for Efficient Long-Context Prefill (PR [Feature] [Feature]Virtual Pipeline Parallel #16618 ) @shiyan1121
- Block State Validity Check for KV Cache Runtime Guard @djh95
- Full Attention Scheme Based on Turboquant @xinrunxue
2026.9.11
Chair: @ningjingbengxiaohai
Topic:
- Community Info Sync - new model and so on. @ningjingbengxiaohai
- Suggested PR Review Metrics for Size, Review, and Testing @hl843901190
2026.9.4
Chair: @wangxiyuan
Topic:
- Agent Hint For vllm-ascend @HiC4Sh1e
- LayerSplit for Prefill KV Cache @recky-c
- Patch removal sync - @wangxiyuan
- FFN operator (Ascend950) integration into vllm-ascend @chenchris2
2026.8.28 (cancel)
2026.8.21 (cancel)
2026.8.14:
Chair: @ningjingbengxiaohai
Topic:
- Support Extensible Metrics Configuration and Handlers for vLLM-Ascend -@chaseche77
- vllm Multi-Process Log Rotation Solution Design - @wenjinhust
- [Feature]：Research on Joint Optimization Methods for Long-Sequence KV Cache in Rollout - @hanyunpeng
- DFX Runtime Observability & Anomaly-Triggered Forensic Dump for vLLM-Ascend - @djh95
2026.8.7:
Chair: @zzzzwwjj
Topic:
- Model Runner V2 Pipeline Parallelism - @LostFox11
- Community Testing Coverage and Quality Improvement Discussion -@winson-00178005
- support 910PremiumA NPU for vllm - @perrypeng-peng
- Keep the community images consistent with the developing images - @wenjun91
2026.7.31:
Chair: @LCAIZJ
Topic:
- Low-latency NPU Operators for Decode KVcache Offloading - @xwLearnsLLM
- Introducing the vLLM AFD Plugin: Attention–FFN Disaggregation for Flexible MoE Serving - @jiangkuaixue123
- Introducing Test and Doc Review into RC Release Process - @hl843901190
2026.7.24
Chair: @ningjingbengxiaohai
Topic:
- More time-consuming logs and trace points by vllm/vllm-ascend - @huangdong2022
- KV cache offloading to support longer sequence length - @zhangsicheng5
- BLASST Sparse Attention NPU Implementation for Long-Context Inference - @iskezh
- The TurboQuant fusion operator NPU implementation for long-context inference. - @XiaolongShen1888
2026.7.17
Chair: @realliujiaxu
Topic:
- vLLM Q3 Roadmap [Roadmap] vLLM Roadmap Q3 2026 vllm#48168 - @wangxiyuan
- Distributed Edge-Cloud Collaborative Inference - @ninahqq
- Add EVS support for Qwen3.5 models - @garrygale
2026.7.10
Chair: @ningjingbengxiaohai
Topic:
- Review of the strategy for integrating PR with precision testing and nightly health checks. - @shenhui-cli
- [RFC] vLLM-Ascend logging quality standard - @xiaoshudian555
- Where to place troubleshooting guides in vllm-ascend repository - @xiaoshudian555
- Cross-DP Communication Bubble Optimization for Decoder Serving - @Ydr-pku
- Add score encoder cache manager - @sereinnwang
2026.7.3 (cancel)
2026.6.26
Chair: @wangxiyuan
Topic:
- vLLM weekly Sync - @wangxiyuan
- Rethinking Context Parallel - @pisceskkk
- Layerwise KV cache pool with prefill layer reuse support - @ader47
Activity
ader47 commentedon Jun 25, 2026
shenhui-cli commentedon Jun 29, 2026
xiaoshudian555 commentedon Jul 1, 2026
xiaoshudian555 commentedon Jul 1, 2026
Ydr-pku commentedon Jul 7, 2026
sereinnwang commentedon Jul 9, 2026
ninahqq commentedon Jul 14, 2026
garrygale commentedon Jul 15, 2026
huangdong2022 commentedon Jul 22, 2026
zhangsicheng5 commentedon Jul 23, 2026
iskezh commentedon Jul 23, 2026
XiaolongShen1888 commentedon Jul 23, 2026
xwLearnsLLM commentedon Jul 28, 2026
jiangkuaixue123 commentedon Jul 29, 2026
27 remaining items
Id-Vulkan commentedon Oct 22, 2026
florian0125 commentedon Oct 22, 2026
Tjh-UKN commentedon Oct 23, 2026
Id-Vulkan commentedon Oct 29, 2026
banyannnn commentedon Oct 29, 2026
LittleJiafang commentedon Oct 30, 2026
Ronald1995 commentedon Nov 2, 2026
wenjinhust commentedon Nov 3, 2026
yiz-liu commentedon Nov 5, 2026
djh95 commentedon Nov 5, 2026
yanghaocsg commentedon Nov 6, 2026
chujiangang2006 commentedon Nov 6, 2026
wangbiao0814 commentedon Nov 10, 2026
kunpengW-code commentedon Nov 12, 2026
jiemeng6 commentedon Nov 13, 2026
Topic for 11.13: [Feature][MoE] Add opt-in DSpark expert-pool routing for expert offload [#21680](#21680)
Background
The MoE expert-offload path keeps expert weights in host memory and loads missing experts into a fixed set of NPU slots. During DSpark verification, different token positions can select different experts, increasing the number of distinct expert weights that may need to be transferred in a single step. [Expert-offload PR #18604](#18604)
Proposal
We propose opt-in P/E expert-pool routing for DeepSeek on ModelRunner V2. At each MoE layer, the first P verification positions of each request, including the anchor, contribute their Top‑6 experts to a shared pool. The remaining positions each contribute their Top‑E candidates and select their final Top‑6 from the completed pool. Two Triton kernels build the pool and produce expert IDs and normalized routing weights for the existing offload path.
On a single Ascend A5 running DeepSeek‑V4‑Flash W4A8 with MoE offload, 84 resident expert slots per layer, and 7 draft tokens per DSpark step, the proposed routing reduced host-to-device expert transfers by 44%. TPOT fell from 26.1 to 21.1 ms/token (−19.2%), and output throughput rose from 35.3 to 42.7 tokens/s (about +20%), with less than 1% accuracy loss.