Commits · a071dc4084ea08aea1ab2fe96baace67b5f29b5f · change / sglang · GitLab

21 May, 2025 8 commits
- 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
- 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 12 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
- [doc] add note for get_num_kv_splits in triton_backend (#6444) · 69af3ec3
  JieXin Liang authored May 20, 2025
  
  69af3ec3
- [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
- 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
- Revert "Implement `return_hidden_states` for the OpenAI API (#6137)" (#6440) · b1465557
  Yineng Zhang authored May 19, 2025
  
  b1465557
- [BUG] fix stop_profile crash (#6431) · b06215da
  Yi Zhang authored May 20, 2025
  
  b06215da
19 May, 2025 4 commits
- [Metrics] Add KV events publishing (#6098) · 7adf245b
  Trevor Morris authored May 19, 2025
  
  7adf245b
- Fix nodeepgemm init (#6417) · 844e2f22
  lukec authored May 19, 2025
  
  844e2f22
- Implement `return_hidden_states` for the OpenAI API (#6137) · 4f39bcf7
  kyle-pena-kuzco authored May 19, 2025
  
  4f39bcf7
- Fix request id error (#6401) · 31c9569b
  fzyzcjy authored May 19, 2025
  
  31c9569b
18 May, 2025 16 commits
- [Bugfix] Fix field error in v1_embedding_request (#6400) · 1be6956d
  Chang Su authored May 18, 2025
  
  1be6956d
- vlm: tensor hash kernel (#5974) · 626ccb7d
  Mick authored May 19, 2025
  
  626ccb7d
- Refactor DeepSeek MoE layer to unify the two forward branches (#6325) · 72bfb0ba
  fzyzcjy authored May 19, 2025
  
  72bfb0ba
- refactor: Extract repeated member variables in KVCache subclasses to base class. (#6323) · 15521495
  wangxiyu191 authored May 19, 2025
  
  15521495
- [Misc] Implement RankZeroFilter for rank-specific logging in model_runner.py (#6333) · ebe58d54
  Chang Su authored May 18, 2025
  
  ebe58d54
- [OAI] Add rid tracing for v1/embeddings and fix rid type in Chat (#6397) · 066cf445
  Chang Su authored May 18, 2025
  
  066cf445
- [fix] fix fa3 forward_decode with spec_decode (#6395) · 1f30c05d
  JieXin Liang authored May 19, 2025
```
Co-authored-by: Stefan He <hebiaobuaa@gmail.com>
```
  1f30c05d
- chore: tiny remove duplicated code (#6392) · 9d24c3ff
  doujiang24 authored May 18, 2025
```
Signed-off-by: doujiang24 <doujiang24@gmail.com>
```
  9d24c3ff
- The Gemma template is missing a newline after the user role. (#6331) · 24161c59
  Yury Sulsky authored May 18, 2025
```
Co-authored-by: Yury Sulsky <ysulsky@tesla.com>
```
  24161c59
- Add pipeline parallelism for Qwen2 and Qwen3 Model (#6250) · 11553c1a
  libra authored May 18, 2025
  
  11553c1a
- refactor: minor refactors regarding multimodal processing (#6187) · 01dd39ba
  Mick authored May 18, 2025
  
  01dd39ba
- Do not use FA3 for mistral (#6379) · b3f3d610
  Lianmin Zheng authored May 17, 2025
  
  b3f3d610
- chore: upgrade sgl-kernel v0.1.3 (#6377) · f07c6a00
  Yineng Zhang authored May 17, 2025
  
  f07c6a00
- Fix CI tests (#6362) · 4bb816d4
  Lianmin Zheng authored May 17, 2025
  
  4bb816d4
- [Fix Chat API] add request id for chat/completion for tracing (#6364) · c250939e
  ybyang authored May 18, 2025
  
  c250939e
- fix: allow `launch_dummy_health_check_server` to start inside of running asyncio loop (#6330) · b6909aa2
  ishandhanani authored May 17, 2025
  
  b6909aa2