Commits · 6b3bb3aea2d6d091273d1070c26d551bb27e6363 · OpenDAS / vllm_cscc

18 Mar, 2026 1 commit
- sbo-deepep-gemm based on v0.9.2-dev-0316-dp · 6b3bb3ae
  chenhw5 authored Mar 18, 2026
  
  6b3bb3ae
17 Mar, 2026 1 commit
- [PD]支持dp的分支 · 236266a9
  xuxz authored Mar 17, 2026
  
  236266a9
13 Mar, 2026 1 commit
- remove cpp · 8c16f318
  yangshj1 authored Mar 13, 2026
  
  8c16f318
12 Mar, 2026 1 commit
- feat: 元宝 prefill融合算子优化 · d7bee8b6
  wanghl6 authored Mar 12, 2026
  
  d7bee8b6
11 Mar, 2026 1 commit
- [PD]回退p2pncclconnector · 73032f48
  xuxz authored Mar 11, 2026
  
  73032f48
09 Mar, 2026 4 commits
- [PD]移除dp · a997359c
  xuxz authored Mar 09, 2026
  
  a997359c
- 添加dspk prefill atten前FUSE_CAT_AND_CAST_FP8 · badaff2d
  wanghl6 authored Mar 09, 2026
  
  badaff2d
- cpp fix · 004a1ef4
  yangshj1 authored Mar 09, 2026
  
  004a1ef4
- add cpp, async tensor transfer between pp ranks · 5f308e68
  yangshj1 authored Mar 09, 2026
  
  5f308e68
26 Feb, 2026 1 commit
- fix: zero overhead KERNEL VMFault · 808a9ed9
  jujl1 authored Feb 26, 2026
  
  808a9ed9
25 Feb, 2026 2 commits
- 完善之前提交 · 6f866c45
  Your Name authored Feb 25, 2026
  
  6f866c45
- [PD][Feat]支持pd模式下dp并行 · d48d8d6d
  Your Name authored Feb 25, 2026
  
  d48d8d6d
12 Feb, 2026 1 commit
- [fix]修复tp1 dpsk启动失败 · b6a7135f
  王敏 authored Feb 12, 2026
  
  b6a7135f
11 Feb, 2026 3 commits
- support fuse cat + q to fp8 + mla · ad7c14d5
  zhuwenwen authored Feb 11, 2026
  
  ad7c14d5
- update Q/K/V_SCALE_CONSTANT · ab674544
  zhuwenwen authored Feb 11, 2026
  
  ab674544
- for qwen3 dense,set VLLM_USE_OPT_RESHAPE_AND_CACHE=1 · 39ff5a5a
  zhuwenwen authored Feb 11, 2026
  
  39ff5a5a
09 Feb, 2026 5 commits
- fix: 修复ep的变量未定义 · bee0b4e8
  wujl5 authored Feb 09, 2026
  
  bee0b4e8
- set VLLM_USE_FUSED_QA_KVA_GEMM=1 · 3c74c91a
  zhuwenwen authored Feb 09, 2026
  
  3c74c91a
- fix indentation · b6f5a16a
  zhuwenwen authored Feb 09, 2026
  
  b6f5a16a
- pd separation uses default scheduling (for all models) · 2aedb72b
  zhuwenwen authored Feb 09, 2026
  
  2aedb72b
- pd separation uses default scheduling · 7539745c
  zhuwenwen authored Feb 09, 2026
  
  7539745c
08 Feb, 2026 1 commit
- [fix]解决宽松mtp引入的同步问题 · b70256d7
  王敏 authored Feb 08, 2026
  
  b70256d7
06 Feb, 2026 3 commits
- 去掉DTBmm的一个冗余条件 · 8b6ef0da
  wujl5 authored Feb 06, 2026
  
  8b6ef0da
- perf: mla后面的DTbmm融合 · 8be144d5
  wujl5 authored Feb 06, 2026
  
  8be144d5
- feat:w4a8Linear调用apply_int8_linear，以支持blaslt · d435d1cd
  jujl1 authored Feb 06, 2026
  
  d435d1cd
05 Feb, 2026 2 commits
- fix: 修复重复判断逻辑 · 7acefb55
  jujl1 authored Feb 05, 2026
  
  7acefb55
- feat: Support shared experts fusion. · bbe4df8b
  wanglong3 authored Jan 29, 2026
```
feat: support moe sum when topk==9

bugfix: Fix mtp model load error when eable shared experts fusion.
```
  bbe4df8b
03 Feb, 2026 1 commit
- fix: 修复pp资源抢占bug · cb7d90a7
  jujl1 authored Feb 03, 2026
  
  cb7d90a7
02 Feb, 2026 1 commit
- fuse_moe_fp8接入marlin算子 · 3c283de3
  SAC_fanth authored Feb 02, 2026
  
  3c283de3
28 Jan, 2026 4 commits

fix(v1)：修复抢占恢复时 BlockTable 溢出 · 87ce2db9

laibao authored Jan 28, 2026

  - 缺失请求统一延后重新加入，避免同一轮重复写入
  - 抢占恢复用覆盖写入（add_row），正常运行用增量追加（append_row）
  - 保持正常请求的追加语义不变

87ce2db9

feat(moe)：新增 VLLM_USE_MOE_W16A16_TRTION 强制 Triton MoE · cedfe391

laibao authored Jan 28, 2026

增加环境变量开关，禁用 Marlin W16A16 MoE 路径
强制 Triton 且权重已是 Marlin packed 时给出明确报错
Marlin 支持探测改为 best-effort（不再依赖 VLLM_USE_LIGHTOP）

cedfe391

update self.kv_cache_dtype · 989a3fad
zhuwenwen authored Jan 28, 2026

989a3fad
update kv_cache_dtype · a9c755ac
zhuwenwen authored Jan 28, 2026

a9c755ac

27 Jan, 2026 4 commits
- fix: 解决原版0消耗chunk-prefill崩溃问题 · 2b1be0e8
  jujl1 authored Jan 27, 2026
  
  2b1be0e8
- remove redundant kv_cache_dtype_str · 9c95f8b0
  zhuwenwen authored Jan 27, 2026
  
  9c95f8b0
- V0.9.2 dev channel lxh · 747cd248
  wanglong3 authored Jan 27, 2026
  
  747cd248
- fp8增加fused_moe_gate参数 · 091cff41
  SAC_fanth authored Jan 27, 2026
  
  091cff41
26 Jan, 2026 1 commit
- fix local kv_cache_dtype_str · 718337a7
  zhuwenwen authored Jan 26, 2026
  
  718337a7
23 Jan, 2026 1 commit
- support fa kvcache fp8, add VLLM_USE_QUERY_QUANT to not use q quant(todo) · b3062dab
  zhuwenwen authored Jan 23, 2026
  
  b3062dab
22 Jan, 2026 1 commit
- 零消耗添加epsp · cc4d1002
  王敏 authored Jan 22, 2026
  
  cc4d1002