Commits · 7a80f56513752e2ed63beed542daaf28119cd89a · change / sglang · GitLab

22 May, 2025 10 commits
- Support dynamically rebalancing experts using EPLB (#6469) · 7a80f565
  fzyzcjy authored May 22, 2025
  
  7a80f565
- Support logging expert balancedness metrics (#6482) · 9484eba4
  fzyzcjy authored May 22, 2025
  
  9484eba4
- [RL] Remove the w13 weight_scale and input_scale for UnquantizedEPMoE… (#6308) · e9feb488
  Zilin Zhu authored May 22, 2025
  
  e9feb488
- Support updating expert locations dynamically (#6388) · fc992a09
  fzyzcjy authored May 22, 2025
  
  fc992a09
- Add main for merge state tests (#6492) · 121f92c5
  Yuan Luo authored May 22, 2025
```
Co-authored-by: luoyuan.luo <luoyuan.luo@antgroup.com>
```
  121f92c5
- [PD] Abort request if transfer fails (#6504) · 3bde1010
  Byron Hsu authored May 21, 2025
  
  3bde1010
- [PD] Add doc and simplify sender.send (#6019) · 75135580
  Byron Hsu authored May 21, 2025
  
  75135580
- [1/2] Support Qserve (#6457) · 4d643f6c
  HandH1998 authored May 22, 2025
```
Co-authored-by: yych0745 <1398089567@qq.com>
Co-authored-by: sleepcoo <sleepcoo@gmail.com>
```
  4d643f6c
- Apply constraint grammar to EAGLE (#6499) · 6ce0ed07
  Ke Bao authored May 22, 2025
```
Co-authored-by: merrymercy <lianminzheng@gmail.com>
```
  6ce0ed07
- Recover from corrupted cache file in bench serving (#6510) · 969660c7
  fzyzcjy authored May 22, 2025
  
  969660c7
21 May, 2025 17 commits
- doc: Update README.md with adding deepwiki badge to enable weekly auto-refresh (#6508) · 16d4f680
  Xinyuan Tong authored May 21, 2025
  
  16d4f680
- fix: EXAONE when using tie_word_embeddings (#5759) · ada268fd
  Kyungmin Lee authored May 22, 2025
  
  ada268fd
- [CPU] Fix build issue (#6419) · cfe48c59
  blzheng authored May 22, 2025
  
  cfe48c59
- [Fix]Fix capture fail bug for DeepSeek (#6275) · d4c038da
  Baizhou Zhang authored May 21, 2025
  
  d4c038da
- Fix bench_one_batch_server (#6503) · 55f6005f
  fzyzcjy authored May 22, 2025
  
  55f6005f
- Let bench_one_batch_server use sharegpt data to make expert distribution more natural (#5573) · 7222e1da
  fzyzcjy authored May 21, 2025
  
  7222e1da
- Tiny make Lint CI show diff (#6445) · 505eec4d
  fzyzcjy authored May 21, 2025
  
  505eec4d
- Support redundant experts in expert parallel (#6461) · ccfe5c00
  fzyzcjy authored May 21, 2025
  
  ccfe5c00
- Tiny add stage assertions to DeepEPDispatcher to avoid misuse (#6467) · a071dc40
  fzyzcjy authored May 21, 2025
  
  a071dc40
- Fix num_qps_per_rank computation when providing custom DeepEP configuration (#6468) · a40aecc5
  fzyzcjy authored May 21, 2025
  
  a40aecc5
- Refactor DeepSeek attention dispatching (#6476) · d6e1d28c
  fzyzcjy authored May 21, 2025
  
  d6e1d28c
- [RL] allow weight updation with dp attention enabled (#6311) · 7c347259
  Zilin Zhu authored May 21, 2025
  
  7c347259
- [router] support http2 in router (#6487) · 669caa0a
  Zilin Zhu authored May 21, 2025
  
  669caa0a
- Implement Siglip Vision model, and support BNB quantization for gemma3-mm (#5339) · 4024e1d2
  Jiajun Li authored May 20, 2025
  
  4024e1d2
- aiter attention-backend (default enabled on AMD/ROCm) (#6381) · 5c0b38f3
  HAI authored May 20, 2025
  
  5c0b38f3
- Refactor group_concurrent_contiguous in NIXL (#6214) · 30ca18f4
  Yuan Luo authored May 21, 2025
```
Co-authored-by: luoyuan.luo <luoyuan.luo@antgroup.com>
```
  30ca18f4
- Disable all two stream overlap on amd (#6475) · 03886917
  Lianmin Zheng authored May 20, 2025
  
  03886917
20 May, 2025 13 commits
- [docs] Fix torch version (#6472) · 66324895
  Wenxuan Tan authored May 20, 2025
  
  66324895
- 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
- [doc] add note for get_num_kv_splits in triton_backend (#6444) · 69af3ec3
  JieXin Liang authored May 20, 2025
  
  69af3ec3
- Update extend/decode attention kernel for CPU in sgl-kernel and add UTs (#6405) · 32cc66ef
  YanbingJiang authored May 20, 2025
```
Co-authored-by: mingfeima <mingfei.ma@intel.com>
```
  32cc66ef
- [QuickFix] fix gptq model initialize (#6429) · 83f2d9d4
  PGFLMG authored May 20, 2025
  
  83f2d9d4
- Address performance regression: disable multiple streams on ROCm (#6412) · 6317c5c6
  HAI authored May 19, 2025
  
  6317c5c6
- Support DeepSeek EPLB algorithm with static distributions (#6387) · cba1cdbc
  fzyzcjy authored May 20, 2025
  
  cba1cdbc
- Support loading weights when physical experts are different from logical experts (#6386) · c471d39e
  fzyzcjy authored May 20, 2025
  
  c471d39e
- Refactor DeepSeek logic into atomic operations (#6326) · d0443275
  fzyzcjy authored May 20, 2025
  
  d0443275
- Remove `Cargo.lock`, add it into .gitignore (#6438) · 17d080b7
  Liangsheng Yin authored May 20, 2025
  
  17d080b7
- 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