Commits · 8aab7fdb21e86b87a3141951ce498246fd4aae73 · change / sglang · GitLab

15 Apr, 2025 5 commits
- chore: upgrade sgl-kernel 0.0.9 (#5401) · 8aab7fdb
  Yineng Zhang authored Apr 14, 2025
  
  8aab7fdb
- [fix/misc] remove duplicate row in deepseek v2 model (#5279) · ee9d6ca6
  Yangcheng Li authored Apr 15, 2025
  
  ee9d6ca6
- Add H20 dtype fp8_w8a8 shared experts fused MoE kernel tuning configs for DeepSeek V3/R1 (#5291) · 2dd64894
  Ximingwang-09 authored Apr 15, 2025
```
Co-authored-by: ximing.wxm <ximing.wxm@antgroup.com>
```
  2dd64894
- Add A800 shared experts fused MoE kernel tuning configs for DeepSeek V3/R1 (#5368) · 61e7c4dd
  lambert0312 authored Apr 15, 2025
  
  61e7c4dd
- [Fix] Turn off DeepGEMM by default (#5263) · f6772f14
  Baizhou Zhang authored Apr 14, 2025
  
  f6772f14
14 Apr, 2025 8 commits
- apply fused moe gate in ds v3/r1 (#5371) · 38076dea
  Xiaoyu Zhang authored Apr 15, 2025
```
Co-authored-by: Yineng Zhang <me@zhyncs.com>
```
  38076dea
- Apply deepseek cuda rope (#5385) · 5e0a9b09
  Ke Bao authored Apr 15, 2025
```
Co-authored-by: Yineng Zhang <me@zhyncs.com>
```
  5e0a9b09
- [perf] experimental enhance fp8 per-tensor quant (#5370) · bdde2375
  JieXin Liang authored Apr 15, 2025
  
  bdde2375
- [PD Bug] fix MLA get_contiguous_buf_infos error (#5384) · e9fc2ac7
  ybyang authored Apr 14, 2025
  
  e9fc2ac7
- Fix PD disaggregation bugs (#5326) · 44afde82
  Liangsheng Yin authored Apr 14, 2025
  
  44afde82
- Support for Qwen2.5-VL Model in bitsandbytes Format (#5003) · 072df753
  yhyang201 authored Apr 14, 2025
  
  072df753
- Fix DeepSeek DP Attention + torch compile (#5367) · defede50
  fzyzcjy authored Apr 14, 2025
```
Co-authored-by: ispobock <ispobaoke@163.com>
```
  defede50
- Free metadata_buffer_index after transfer finished (#5364) · 14e8bd88
  Yongtong Wu authored Apr 14, 2025
  
  14e8bd88
13 Apr, 2025 8 commits
- [DeepEP] Reduce routed scaling overhead (#5277) · adca585b
  yulei authored Apr 14, 2025
```
Co-authored-by: Cheng Wan <54331508+ch-wan@users.noreply.github.com>
```
  adca585b
- fix #5322 (#5359) · 39e41138
  Yineng Zhang authored Apr 13, 2025
  
  39e41138
- fix MLATokenToKVPoolHost get_size_per_token bug (#5161) · 5fbafbb8
  huangtingwei authored Apr 14, 2025
```
Co-authored-by: AniZpZ <zhuangsen.zp@antgroup.com>
```
  5fbafbb8
- [PD] Add transfer backend abstraction (#5328) · a9499885
  Byron Hsu authored Apr 13, 2025
  
  a9499885
- Fix typo: infight -> inflight (#5357) · f7655790
  Liangsheng Yin authored Apr 14, 2025
  
  f7655790
- chore: upgrade sgl-kernel 0.0.8.post3 (#5342) · f58b929a
  Yineng Zhang authored Apr 13, 2025
  
  f58b929a
- Fix: Ensure tensors for dist.broadcast match NCCL backend device (#5322) · 8311b07f
  mlmz authored Apr 13, 2025
  
  8311b07f
- fix: determine if flashinfer is installed (#5336) · 7d3b7c87
  Yineng Zhang authored Apr 12, 2025
  
  7d3b7c87
12 Apr, 2025 8 commits
- Support server based rollout in Verlengine (#4848) · bc92107b
  tianlian yi authored Apr 13, 2025
```
Co-authored-by: Jin Pan <jpan236@wisc.edu>
Co-authored-by: Chayenne <zhaochen20@outlook.com>
Co-authored-by: Jinn <47354855+jhinpan@users.noreply.github.com>
```
  bc92107b
- refine fused_moe tuning docs (#5294) · 3e4794aa
  Xiaoyu Zhang authored Apr 13, 2025
  
  3e4794aa
- Delete python/sglang/srt/layers/moe/fused_moe_triton/configs/E=257,N=… (#5321) · 690ec205
  Xiaoyu Zhang authored Apr 13, 2025
  
  690ec205
- feat: use fa3 mla by default on hopper (#5210) · 57de7c6b
  Yineng Zhang authored Apr 12, 2025
```
Co-authored-by: yundai424 <yundai424@gmail.com>
Co-authored-by: hebiao064 <hebiaobuaa@gmail.com>
```
  57de7c6b
- Fix fa3 window size setup (#5316) · aea98512
  Qingquan Song authored Apr 11, 2025
  
  aea98512
- Fix w8a8_int8 model shared experts fusion load weights error (#5120) · 1b1b47a9
  lambert0312 authored Apr 12, 2025
  
  1b1b47a9
- [VLM] Adopt fast image processor by default (#5065) · 34ef6c81
  Mick authored Apr 12, 2025
  
  34ef6c81
- fix: use deepgemm only on hopper (#5310) · 61172091
  Yineng Zhang authored Apr 11, 2025
  
  61172091
11 Apr, 2025 9 commits
- feat: add blackwell Dockerfile (#5302) · f774a0d2
  Yineng Zhang authored Apr 11, 2025
  
  f774a0d2
- remove moe_align_block_size torch.zeros in small batch/expert mode (#5298) · 60bcbf2a
  Xiaoyu Zhang authored Apr 12, 2025
  
  60bcbf2a
- [PD] fix: skip warmup request in disaggregation mode to prevent crash on timeout (#5292) · c35dcfdb
  Yusong Gao authored Apr 11, 2025
  
  c35dcfdb
- [fix] fix mrope positions not picked up (#5265) · e53a0b3d
  Mick authored Apr 11, 2025
  
  e53a0b3d
- Support `--enable-llama4-multimodal` (#5254) · 038bc5d5
  Cheng Wan authored Apr 11, 2025
  
  038bc5d5
- Optimize GPU memory usage in FlashAttentionBackend's strided indexing (#5262) · aee62d74
  Chang Su authored Apr 11, 2025
```
Co-authored-by: ch-wan <cwan39@gatech.edu>
```
  aee62d74
- Optimize attention in llama4 (#5127) · cd7e32e2
  fzyzcjy authored Apr 11, 2025
  
  cd7e32e2
- ROCm/AITER CK_MoE: update 2-stage kernels & support both Activations (#5228) · 88799448
  HAI authored Apr 10, 2025
  
  88799448
- Fix torch.compile cacheing (#5259) · a879811c
  Richard Zou authored Apr 10, 2025
```
Co-authored-by: zhyncs <me@zhyncs.com>
```
  a879811c
10 Apr, 2025 2 commits
- Update deps for mllama4 (#5215) · 1078396f
  Ke Bao authored Apr 11, 2025
  
  1078396f
- [PD] Add get_contiguous_buf_infos interface for MLATokenToKVPool (#5204) · 7e4f72dd
  Teng Ma authored Apr 10, 2025
  
  7e4f72dd