Commits · 989a0a2b31073333e1ca94edf8d74481e842903b · OpenDAS / vllm_cscc

"docs/pages/design-docs/request-plane.md" did not exist on "d59b9d72ed98eed50e8863de8259d0871efad336"

08 Jan, 2026 2 commits
- feat: Support enable rms quant and shared expert overlap at same time. · 989a0a2b
  wanglong3 authored Jan 08, 2026
  
  989a0a2b
- fix：更新 Marlin W16A16 MoE 中支持的断言消息 · d77a3d5b
  laibao authored Jan 08, 2026
  
  d77a3d5b
07 Jan, 2026 2 commits
- 解决custom allreduce在dp情况下的其服务错误问题 · 912dc4c9
  xiabo authored Jan 07, 2026
  
  912dc4c9
- DS量化模型重构atten和moe调用rmsquant融合逻辑。 · 89b62a25
  wujl5 authored Jan 07, 2026
  
  89b62a25
06 Jan, 2026 2 commits
- bugfix: Fix the startup crash issue of the service when USE_FUSED_RMS_QUANT=1 · e6d32c6d
  wanglong3 authored Jan 06, 2026
  
  e6d32c6d
- fix: 修复deepseek在PP>1开chunkprefill精度问题 · 4b752578
  jujl1 authored Jan 06, 2026
  
  4b752578
05 Jan, 2026 3 commits
- feat: enable shared expert overlap. · ee19dca6
  wanglong3 authored Jan 03, 2026
  
  ee19dca6
- feat：为 GLM4 和 Llama 模型新增 MultiModalConfigProxy，以支持扁平配置与嵌套的多模态配置（text_config） · 952f0347
  laibao authored Jan 05, 2026
  
  952f0347
- 删除调试信息 · c773cc66
  yangql authored Jan 05, 2026
  
  c773cc66
04 Jan, 2026 5 commits
- 解决pd分离非对称切分通信组过多问题 · 294cc23a
  xiabo authored Jan 04, 2026
  
  294cc23a
- 区分pcie和hglink custom allreduce的使用 · 1b5aa25e
  xiabo authored Jan 04, 2026
  
  1b5aa25e
- fix: 只有当kv block中不含有MTP的假数据时才会被cached，以修复cache_full_blocks同一个kv block保存两次的bug · f7cb8c7b
  jujl1 authored Jan 04, 2026
  
  f7cb8c7b
- perf: 加速 v1 InputBatch.add_request 的 token_ids 拷贝 · fbe8587a
  laibao authored Jan 04, 2026
```
新增环境变量开关 VLLM_V1_FAST_TOKEN_ID_COPY（默认关闭）。开启后在 CachedRequestState 中缓存 prompt_token_ids 的 np.int32，并在 add_request 里用 np.copyto 写入 token_ids_cpu，避免长 prompt 场景反复 list->NumPy 转换开销（尤其是抢占/反复进出 batch 时）
```
  fbe8587a
- perf: DS-量化模型融合qa和kva的gemm · 577eb49f
  wujl5 authored Jan 04, 2026
  
  577eb49f
31 Dec, 2025 5 commits
- fix: chunk-prefill生成的token被错误返回 · 2c1de3fa
  jujl1 authored Dec 31, 2025
  
  2c1de3fa
- Revert "增加mtp>1的pad" · d8c36dbf
  yangql authored Dec 31, 2025
```
This reverts commit 059b1961.
```
  d8c36dbf
- [feat]解决deepep auto模式 mtp >1 时卡住 · 37d2b35d
  王敏 authored Dec 31, 2025
  
  37d2b35d
- 增加mtp>1的pad · 059b1961
  yangql authored Dec 31, 2025
  
  059b1961
- feat: 兼容MTP零消耗和主模型+MTP零消耗(VLLM_ZERO_OVERHEAD_ENHANCE=1)开启 · 76e22965
  jujl1 authored Dec 31, 2025
  
  76e22965
30 Dec, 2025 3 commits
- feat: 兼容MTP零消耗和主模型+MTP零消耗(VLLM_ZERO_OVERHEAD_ENHANCE=1)开启 · 8a413453
  jujl1 authored Dec 30, 2025
  
  8a413453
- 解决pd分离非对称切分通信组过多问题 · 84e5aba2
  xiabo authored Dec 30, 2025
  
  84e5aba2
- fix: 解决列表访问越界 · 5208b291
  jujl1 authored Dec 30, 2025
  
  5208b291
29 Dec, 2025 6 commits
- [feat]解决deepep auto模式卡住问题 · cb64c6bc
  王敏 authored Dec 29, 2025
  
  cb64c6bc
- fix: 解决同时处理prefill和decode时的prefill请求token计数错误 · bc945a5a
  jujl1 authored Dec 24, 2025
  
  bc945a5a
- fix: support chunk-prefill and fix bug in check_stop · 96197e48
  jujl1 authored Dec 23, 2025
  
  96197e48
- feat: triton kernel 实现 update_input · 89639c96
  jujl1 authored Dec 23, 2025
  
  89639c96
- feat: 主模型+mtp提前返回 · 0936ee97
  jujl1 authored Nov 28, 2025
  
  0936ee97
- 增加mtp的pad · e60a51b4
  yangql authored Dec 29, 2025
  
  e60a51b4
26 Dec, 2025 1 commit
- add VLLM_USE_FLASH_ATTN_FP8 to support fa fp8 · 43546076
  zhuwenwen authored Dec 26, 2025
  
  43546076
25 Dec, 2025 3 commits
- [fix]解决EP不启用mtp时某些size报错 · a9b4bc74
  王敏 authored Dec 25, 2025
  
  a9b4bc74
- [fix]解决EP不开启mtp时某些size报错 · 9925dd0e
  王敏 authored Dec 25, 2025
  
  9925dd0e
- 修复w8a8 triton config 择优位运算可能引发torch compile 编译错误，修复smquant w8a8 权重后处理位置 · 16d49763
  gaoqiong authored Dec 25, 2025
  
  16d49763
24 Dec, 2025 3 commits

[feat]解决EP开启pd_padding大输入卡住问题 · 7439b205
王敏 authored Dec 24, 2025

7439b205
[fix]解决EP开启pd_padding大输入卡住问题 · 639a11d2
王敏 authored Dec 24, 2025

639a11d2

V1 采样器：新增 reduced top-k/top-p 采样路径 · 9b1e03d4

laibao authored Dec 24, 2025

新增环境变量 VLLM_V1_USE_REDUCED_TOPK_TOPP_SAMPLER 用于开关控制
扩展 SamplingMetadata，增加 max_top_k 与 has_any_no_top_k
在 InputBatch 侧计算 top-k 的主机端汇总信息，避免 device 同步
更新 Sampler/TopKTopPSampler 传递并使用新参数以启用优化采样

9b1e03d4

23 Dec, 2025 5 commits
- mla模型P、D单实例单机的任意切分方式（满足D的tp>=P的tp)使用 · 4f51931d
  xiabo authored Dec 23, 2025
  
  4f51931d
- [feat]低延迟模式采用int8 dispatch · ce41e45b
  王敏 authored Dec 23, 2025
  
  ce41e45b
- [feat]低延迟模式采用int8 dispatch · 35e43dfb
  王敏 authored Dec 23, 2025
  
  35e43dfb
- update fuse_fill_rms_x2_concat · bac269d7
  zhuwenwen authored Dec 23, 2025
  
  bac269d7
- set VLLM_CUSTOM_CACHE=1 · bdae1255
  zhuwenwen authored Dec 23, 2025
  
  bdae1255