Commits · 327fdf18ac3d37331763bfc8ec97a4df5448d821 · OpenDAS / vllm_cscc

24 Nov, 2025 1 commit
- feat: pipeline_parallel新增pp域请求数均衡,VLLM_USE_PP_BALANCE控制，默认开启 · f4b01cd4
  jujl1 authored Oct 31, 2025
  
  f4b01cd4
15 Nov, 2025 1 commit
- [feat]1.w8a8 marlin适配deepep低延迟;2.非naive ep模式，去掉多余的dp padding,避免allreduce耗时 · 4a943d35
  王敏 authored Nov 15, 2025
  
  4a943d35
13 Nov, 2025 3 commits

add VLLM_USE_PD_SPLIT to split prefill and decode · 97fed613
zhuwenwen authored Nov 13, 2025

97fed613

zhuwenwen authored Nov 13, 2025

- 在环境变量中引入 `VLLM_SCHED_ENABLE_MINIMAL_INJECTION` 以控制流水线并行调度的最小注入。
- 调整 Scheduler 逻辑以使用新的最小注入功能。
- 更新调度逻辑以利用输出占位符，确保在解码过程中避免 0-token 停滞。
- 增强 Scheduler，根据批次队列状态管理最小进度注入。

613edd7d

[feat]w4a8和w8a8适配deepep低延迟 · 92761bde
王敏 authored Nov 13, 2025

92761bde

10 Nov, 2025 1 commit
- [fix]修复mtp中的笔误 · 91c0a497
  王敏 authored Nov 10, 2025
  
  91c0a497
08 Nov, 2025 1 commit
- [fix]解决开启mtp后，在极端情况碰到显存不足时，导致mla中申请的tensor数据错乱问题 · b6074f33
  王敏 authored Nov 08, 2025
  
  b6074f33
07 Nov, 2025 1 commit
- 请求进入到waitting队列后，增加request.is_finished()判断 · 46d16787
  zhuwenwen authored Nov 07, 2025
  
  46d16787
06 Nov, 2025 2 commits
- [fix]解决同时开启chunked-prefill和并行解码出现的指标统计错误问题 · 9dab474d
  王敏 authored Nov 06, 2025
  
  9dab474d
- [fix]修复开启mtp并且显存不足时发生的超出维度限制问题 · a3e6d682
  王敏 authored Nov 06, 2025
  
  a3e6d682
27 Oct, 2025 1 commit
- [feat]w4a8适配deepep ht模式，解决开启dp时mtp>1时卡住问题 · 98b7432a
  王敏 authored Oct 27, 2025
  
  98b7432a
17 Oct, 2025 1 commit
- 支持pd分离p2p_async & 解决oom问题 · 6ca3d790
  zhuwenwen authored Oct 17, 2025
  
  6ca3d790
13 Oct, 2025 2 commits
- 去掉all2all ep相关代码 · 0b467604
  王敏 authored Oct 13, 2025
  
  0b467604
- 新增fp8—e5m2 · ed3cdc81
  zhuwenwen authored Oct 13, 2025
  
  ed3cdc81
11 Oct, 2025 2 commits
- close profile.StartTracer · a92daffa
  maxiao1 authored Oct 11, 2025
  
  a92daffa
- fix pd send async perfomance · 1707bebe
  maxiao1 authored Oct 11, 2025
  
  1707bebe
10 Oct, 2025 3 commits
- token split by token adapt to pd separation & p2p can be used async · 1a135a9d
  maxiao1 authored Oct 10, 2025
  
  1a135a9d
- keep same version · fc5bfc66
  maxiao1 authored Oct 10, 2025
  
  fc5bfc66
- pd分离_tbo · 12291212
  maxiao1 authored Oct 06, 2025
  
  12291212
09 Oct, 2025 1 commit
- Revert "Merge branch 'v0.9.2-dev-lmcache-pd' into 'v0.9.2-dev'" · 7b30ecd8
  zhuwenwen authored Oct 09, 2025
```
This reverts merge request !219
```
  7b30ecd8
06 Oct, 2025 1 commit
- modify engied_id str · 4bb8c6af
  yangshj1 authored Oct 06, 2025
  
  4bb8c6af
30 Sep, 2025 1 commit
- [feat]优化mori计算逻辑，支持cudagraph，按照bs*ep_size截断fused_moe的输入，共享专家不tp切分，去掉最后的allreduce · d2e57a90
  王敏 authored Sep 30, 2025
  
  d2e57a90
28 Sep, 2025 1 commit
- 修复kvcache-fp8—e5m2的不能开cp的bug · b2d14ba3
  yangql authored Sep 28, 2025
  
  b2d14ba3
26 Sep, 2025 2 commits
- update cat kernel · e0fdf4e8
  zhuwenwen authored Sep 26, 2025
  
  e0fdf4e8
- [fix] pp+mtp bs 1 correctness · 22d6f9af
  lizhigong authored Sep 26, 2025
  
  22d6f9af
25 Sep, 2025 1 commit
- fix connector metadata · 1586d41f
  yangshj1 authored Sep 25, 2025
  
  1586d41f
24 Sep, 2025 6 commits
- update VLLM_USE_OPT_CAT · 4d97c5fc
  zhuwenwen authored Sep 24, 2025
  
  4d97c5fc
- [kernel] add lightop's moe_sum(mul+add) fusion operator for deepseek · 8d2cac26
  zhuwenwen authored Sep 24, 2025
```
[FIX] 修复mtp和VLLM_USE_TRITON_CAT不能一起开的bug
```
  8d2cac26
- 修复mtp和VLLM_USE_TRITON_CAT不能一起开的bug · cc2dca96
  SAC_fanth authored Sep 24, 2025
  
  cc2dca96
- fix scheduler issu in pp + mtp · 2cb921da
  lizhigong authored Sep 24, 2025
  
  2cb921da
- add env · c6b7a44b
  yangshj1 authored Sep 24, 2025
  
  c6b7a44b
- add engine id · f6fcc8ff
  xuxz authored Sep 24, 2025
  
  f6fcc8ff
23 Sep, 2025 1 commit
- 修复kvcachefp8与cp/pc的冲突 · 8b7daa0d
  yangql authored Sep 23, 2025
  
  8b7daa0d
16 Sep, 2025 3 commits
- add pd lmcache · f4fd3a96
  yangshj1 authored Sep 16, 2025
  
  f4fd3a96
- 接入新的concat算子，包含decode和prefill，并根据size的不同进行选择 · 2233e3d0
  王敏 authored Sep 16, 2025
  
  2233e3d0
- 添加mori ep · 1a56d6cb
  王敏 authored Sep 16, 2025
  
  1a56d6cb
14 Sep, 2025 1 commit
- [fix]修复eagle 创建cu_num_tokens类型错误问题 · b4610c06
  王敏 authored Sep 14, 2025
  
  b4610c06
13 Sep, 2025 1 commit

update triton cat scheduling · b1329ff2

zhuwenwen authored Sep 13, 2025

update the default values of VLLM_USE_TRITON_CAT and VLLM_USE_LIGHT_OP to True

b1329ff2

12 Sep, 2025 1 commit
- update the cat implementation of triton's non contiguous memory for the decode phase · 2c169409
  zhuwenwen authored Sep 12, 2025
  
  2c169409
10 Sep, 2025 1 commit
- update triton kernel to optimize torch cat for ds prefill · 90e40f49
  zhuwenwen authored Sep 10, 2025
  
  90e40f49