Commits · 989a0a2b31073333e1ca94edf8d74481e842903b · OpenDAS / vllm_cscc

08 Jan, 2026 1 commit
- feat: Support enable rms quant and shared expert overlap at same time. · 989a0a2b
  wanglong3 authored Jan 08, 2026
  
  989a0a2b
07 Jan, 2026 1 commit
- DS量化模型重构atten和moe调用rmsquant融合逻辑。 · 89b62a25
  wujl5 authored Jan 07, 2026
  
  89b62a25
05 Jan, 2026 1 commit
- feat: enable shared expert overlap. · ee19dca6
  wanglong3 authored Jan 03, 2026
  
  ee19dca6
04 Jan, 2026 1 commit
- perf: DS-量化模型融合qa和kva的gemm · 577eb49f
  wujl5 authored Jan 04, 2026
  
  577eb49f
22 Dec, 2025 3 commits
- [fix]1.解决ep sequence parallel优化引入的mtp报错;2.解决共享专家无法和combine overlap问题 · dc01fce4
  王敏 authored Dec 22, 2025
  
  dc01fce4
- [feat]1.优化ep sequence parallel,区分主模型和mtp逻辑;2.ep sequence parallel添加cudagraph... · 62f05dde
  王敏 authored Dec 22, 2025
```
[feat]1.优化ep sequence parallel,区分主模型和mtp逻辑;2.ep sequence parallel添加cudagraph padding到tp_size;3.修复共享专家和deepep combine overlap
```
  62f05dde
- 修复ep的auto模式的崩溃bug · 29523973
  yangql authored Dec 22, 2025
  
  29523973
18 Dec, 2025 2 commits
- [feat]合入基于deepep的大EP · 13130b89
  王敏 authored Dec 18, 2025
  
  13130b89
- [feat]优化dp attention，减少1次allgather耗时，高吞吐提升明显 · 3e386c3b
  王敏 authored Dec 18, 2025
  
  3e386c3b
17 Dec, 2025 1 commit
- remove fuse_rmsnorm_rope_quant_gfx938 · 99981972
  zhuwenwen authored Dec 17, 2025
  
  99981972
16 Dec, 2025 2 commits
- add fuse_rmsnorm_rope_quant_gfx938 to support use fp8_e4m3 mla · 0ce3b670
  zhuwenwen authored Dec 16, 2025
  
  0ce3b670
- up auto deepep · 0d3ae2fc
  yangql authored Dec 16, 2025
  
  0d3ae2fc
15 Dec, 2025 1 commit
- [feat]1.支持高吞吐模式ep_scatter+deepgemm contiguous+ep_gather方案；2.支持高吞吐模式下ETP,例如dp4 tp4 · 3833018c
  王敏 authored Dec 15, 2025
  
  3833018c
10 Dec, 2025 1 commit
- [fix]修复deepep 高吞吐模式vmfault问题 · 916b5876
  王敏 authored Dec 10, 2025
  
  916b5876
08 Dec, 2025 1 commit
- [feat]支持deepep ETP，dp4 tp4 ep16相比dp32 tp1 ep32提升明显 · 6cabbf16
  王敏 authored Dec 08, 2025
  
  6cabbf16
04 Dec, 2025 1 commit
- add VLLM_USE_LIGHTOP_RMS_ROPE_CONCAT when use USE_FUSED_RMS_QUANT and... · de87d606
  zhuwenwen authored Dec 04, 2025
```
add VLLM_USE_LIGHTOP_RMS_ROPE_CONCAT when use USE_FUSED_RMS_QUANT and USE_FUSED_CUSTOM_ALL_REDUCE_RMS_QUANT
```
  de87d606
02 Dec, 2025 2 commits
- update weight and cache · fd559b9f
  zhuwenwen authored Dec 02, 2025
  
  fd559b9f
- [feat]支持deepep低延迟与共享专家overlap · 1ae8f58c
  王敏 authored Dec 02, 2025
  
  1ae8f58c
01 Dec, 2025 1 commit
- add VLLM_USE_OPT_RESHAPE_AND_CACHE (test) · 64e307c7
  zhuwenwen authored Dec 01, 2025
  
  64e307c7
26 Nov, 2025 1 commit
- [pref]: DS_v2_w8a8模型融掉moe.quant · 68972532
  wujl5 authored Nov 26, 2025
  
  68972532
20 Nov, 2025 2 commits
- DS_v2_w4a8_CRQ增加slilu_mul_quant支持 · 075841f3
  wujl5 authored Nov 20, 2025
  
  075841f3
- MLA模块CRQ融合分支增加rms_cuda_opt · 8c646ebe
  wujl5 authored Nov 20, 2025
  
  8c646ebe
18 Nov, 2025 1 commit
- deepseekv2-w4a8支持custom-rms-quant融合 · 3e6729e0
  wujl5 authored Nov 18, 2025
  
  3e6729e0
07 Nov, 2025 1 commit
- add contiguous+rmsnorm to replace triton_ · dc54fefe
  zhuwenwen authored Nov 07, 2025
  
  dc54fefe
04 Nov, 2025 1 commit
- add contiguous+rmsnorm to replace triton_ · 2afbdacb
  zhuwenwen authored Nov 04, 2025
  
  2afbdacb
01 Nov, 2025 1 commit
- [feat]整合mori和deepep相关代码 · d698d6f2
  王敏 authored Nov 01, 2025
  
  d698d6f2
29 Oct, 2025 1 commit
- remove redundant envs · c26cfd1a
  zhuwenwen authored Oct 29, 2025
  
  c26cfd1a
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
24 Oct, 2025 2 commits

add VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD · c2e6f453

zhuwenwen authored Oct 24, 2025

support prefix cache on kme
fix the error in test_moe caused by moe align not supporting 511 and 211
multi-modal switching to torch implementation on z100l&k100

c2e6f453

set USE_FUSED_RMS_QUANT=0 and USE_FUSED_SILU_MUL_QUANT=0 · 46dd30e7
zhuwenwen authored Oct 24, 2025
```
add VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD
```
46dd30e7

15 Oct, 2025 4 commits
- update deepseek_v2.py · 4b3e2d5e
  zhuwenwen authored Oct 15, 2025
  
  4b3e2d5e
- update deepseek_v2.py · 4ae3fc04
  zhuwenwen authored Oct 15, 2025
  
  4ae3fc04
- 删除DPSK_FP16_QUICK，以及增加awq和blockwiseint8的shared_output接口 · 50cb9270
  yangql authored Oct 15, 2025
  
  50cb9270
- 删除DPSK_FP16_QUICK，以及增加awq和blockwiseint8的shared_output接口 · 7f459b46
  yangql authored Oct 15, 2025
  
  7f459b46
13 Oct, 2025 2 commits
- 去掉all2all ep相关代码 · 0b467604
  王敏 authored Oct 13, 2025
  
  0b467604
- support dsv32 · 633f8199
  zhuwenwen authored Oct 13, 2025
  
  633f8199
10 Oct, 2025 1 commit
- fix deepseek_v2.py · 2e1d8e9a
  zhuwenwen authored Oct 10, 2025
  
  2e1d8e9a
30 Sep, 2025 3 commits
- 修复部分代码 · e0ba23b5
  王敏 authored Sep 30, 2025
  
  e0ba23b5
- 修复部分代码 · 46e26bf1
  王敏 authored Sep 30, 2025
  
  46e26bf1
- [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