Commits · 65f091310ca63b96384c9fcf24c0c5aa0c1e4f39 · zhaoyu6 / sglang · GitLab

26 May, 2025 1 commit
- refactor qwen moe code, use communicator to support tp+dp (#6581) · 65f09131
  Yi Zhang authored May 26, 2025
  
  65f09131
25 May, 2025 3 commits
- Support overlapping two batches (#4068) · 0d477880
  fzyzcjy authored May 25, 2025
  
  0d477880
- Add back DeepSeek non-TBO branches (#6578) · b2388433
  fzyzcjy authored May 25, 2025
  
  b2388433
- Refactor attention into multiple stages (#6477) · a38376fa
  fzyzcjy authored May 25, 2025
  
  a38376fa
22 May, 2025 1 commit
- Support updating expert locations dynamically (#6388) · fc992a09
  fzyzcjy authored May 22, 2025
  
  fc992a09
21 May, 2025 4 commits
- [Fix]Fix capture fail bug for DeepSeek (#6275) · d4c038da
  Baizhou Zhang authored May 21, 2025
  
  d4c038da
- Support redundant experts in expert parallel (#6461) · ccfe5c00
  fzyzcjy authored May 21, 2025
  
  ccfe5c00
- Refactor DeepSeek attention dispatching (#6476) · d6e1d28c
  fzyzcjy authored May 21, 2025
  
  d6e1d28c
- Disable all two stream overlap on amd (#6475) · 03886917
  Lianmin Zheng authored May 20, 2025
  
  03886917
20 May, 2025 6 commits
- Fix master CI for DeepSeek (#6447) · 13feffd0
  fzyzcjy authored May 20, 2025
  
  13feffd0
- Support dispatching logical to physical experts (#6385) · e98afbe0
  fzyzcjy authored May 20, 2025
  
  e98afbe0
- Address performance regression: disable multiple streams on ROCm (#6412) · 6317c5c6
  HAI authored May 19, 2025
  
  6317c5c6
- Refactor DeepSeek logic into atomic operations (#6326) · d0443275
  fzyzcjy authored May 20, 2025
  
  d0443275
- Refactor communication logic of DeepSeek for extensibility and understandability (#6321) · 1b19df4b
  fzyzcjy authored May 20, 2025
  
  1b19df4b
- Expert distribution recording without overhead for EPLB (#4957) · f0653886
  fzyzcjy authored May 20, 2025
  
  f0653886
18 May, 2025 1 commit
- Refactor DeepSeek MoE layer to unify the two forward branches (#6325) · 72bfb0ba
  fzyzcjy authored May 19, 2025
  
  72bfb0ba
17 May, 2025 3 commits
- Speed up when having padding tokens in DeepEP (#6175) · 27168308
  fzyzcjy authored May 18, 2025
  
  27168308
- Minor code cleanup refactor for DeepSeek models (#6324) · 2df9d40a
  fzyzcjy authored May 17, 2025
  
  2df9d40a
- Fix one wasted kernel in DeepSeek and minor refactor (#6316) · 8dc191f2
  fzyzcjy authored May 17, 2025
  
  8dc191f2
15 May, 2025 1 commit
- Reduce MoE memory usage (#6147) · f194e14f
  fzyzcjy authored May 16, 2025
  
  f194e14f
13 May, 2025 1 commit

Fix two issues related to `--moe-dense-tp-size=1` (#5657) · b2e95f62

Cheng Wan authored May 13, 2025


Co-authored-by: liusy58 <liusy58@linux.alibaba.com>
Co-authored-by: 颉沆 <xiehang.lsy@alibaba-inc.com>

b2e95f62

12 May, 2025 1 commit
- Performing Vocabulary Parallelism for LM Head across Attention TP Groups (#5558) · 25c83fff
  Cheng Wan authored May 12, 2025
```
Co-authored-by: liusy58 <liusy58@linux.alibaba.com>
```
  25c83fff
11 May, 2025 1 commit
- Add typo checker in pre-commit (#6179) · 2ce87935
  applesaucethebun authored May 11, 2025
```
Co-authored-by: Brayden Zhong <b8zhong@uwaterloo.ca>
```
  2ce87935
10 May, 2025 1 commit
- [fix] fix determine_n_share_experts_fusion (#6118) · c178abda
  JieXin Liang authored May 10, 2025
  
  c178abda
09 May, 2025 1 commit
- opt flashinfer mla cat (#5822) · e30c273b
  xu-yfei authored May 09, 2025
```
Co-authored-by: xuyongfei.xyf <xuyongfei.xyf@antgroup.com>
```
  e30c273b
08 May, 2025 4 commits
- [perf] dsv3 bmm fallback to bf16 (#5662) · 5e023301
  JieXin Liang authored May 09, 2025
  
  5e023301
- DeepEP normal support deepgemm-contiguous (#5626) · acc816d8
  lukec authored May 08, 2025
```
Co-authored-by: Yingyi Huang <yingyihuang2000@outlook.com>
Co-authored-by: Cheng Wan <54331508+ch-wan@users.noreply.github.com>
Co-authored-by: Xuting Zhou <xutingz@nvidia.com>
Co-authored-by: ZhengHSI <zhenghsi@qq.com>
```
  acc816d8
- Clean logs for DeepSeek-V3 launching (#6079) · 73600673
  Baizhou Zhang authored May 07, 2025
  
  73600673
- [refactor] slightly tidy fp8 module (#5993) · b70957fc
  JieXin Liang authored May 08, 2025
  
  b70957fc
02 May, 2025 1 commit
- Overlap qk norm with two streams (#5977) · d8ab6011
  Ke Bao authored May 03, 2025
  
  d8ab6011
01 May, 2025 1 commit
- Remove extra contiguous (#5953) · de2faef9
  Ke Bao authored May 02, 2025
  
  de2faef9
30 Apr, 2025 1 commit
- [Feature] add support kimi vl model (#5383) · 8fefdd32
  liwenju0 authored Apr 30, 2025
```
Co-authored-by: wenju.li <wenju.li@deepctr.cn>
```
  8fefdd32
29 Apr, 2025 1 commit
- Auto set draft model path for MTP (#5793) · dd408ee4
  Ke Bao authored Apr 30, 2025
  
  dd408ee4
27 Apr, 2025 1 commit
- Fuse MLA set kv cache kernel (#5748) · 799c4bb5
  Ke Bao authored Apr 27, 2025
  
  799c4bb5
26 Apr, 2025 1 commit
- Reorder loop in shared expert weight loading (#5719) · c3948ba6
  Ke Bao authored Apr 26, 2025
  
  c3948ba6
24 Apr, 2025 1 commit
- [BugFix] Fix combination of MTP and `--n-share-experts-fusion`with R1 (#5707) · 5d93a950
  Yuhong Guo authored Apr 24, 2025
  
  5d93a950
23 Apr, 2025 2 commits
- Remove unnecessary `torch.full` in DeepSeek (#5601) · 71d1785f
  fzyzcjy authored Apr 23, 2025
  
  71d1785f
- Fuse q_a_proj and kv_a_proj (#5619) · 3f87f831
  Baizhou Zhang authored Apr 22, 2025
  
  3f87f831
22 Apr, 2025 2 commits
- Remove q concat in FA3 backend for DeepSeek decode (#5638) · 6b6e7487
  Ke Bao authored Apr 23, 2025
  
  6b6e7487
- Fix shared experts fusion error without quantization (#5632) · 76d17c7e
  lambert0312 authored Apr 23, 2025
  
  76d17c7e