Commits · 86d10d220f665092f93a3f6e8a31a65a36a4f376 · change / sglang · GitLab

23 Aug, 2025 11 commits
- Update grok.py and tiktoken tokenizer (#9532) · 86d10d22
  Lianmin Zheng authored Aug 23, 2025
  
  86d10d22
- feat(hicache): Supports 3fs-hicache compatibility with dp-attention (#9372) · 83871aa1
  hzh0425 authored Aug 23, 2025
  
  83871aa1
- Partially unify triton per token group quant kernels (#9485) · b1b3f0b3
  fzyzcjy authored Aug 23, 2025
  
  b1b3f0b3
- Tiny make device_loading_context more static (#9478) · 34e5e11f
  fzyzcjy authored Aug 23, 2025
  
  34e5e11f
- Overlapped weight offload (#8034) · 2600fc0d
  fzyzcjy authored Aug 23, 2025
  
  2600fc0d
- [fix] Fix mxfp4 triton MoE tp bug (#9473) · ccd3fb94
  hlu1 authored Aug 23, 2025
```
Signed-off-by: Hao Lu <14827759+hlu1@users.noreply.github.com>
```
  ccd3fb94
- tool-call(dsv3): Improve deepseek-v3 chat template and tool_choice = `required` (#9525) · c9dd70fb
  Chang Su authored Aug 23, 2025
  
  c9dd70fb
- fix: blackwell dsv3 fp8 issue temporary solution (#9530) · 6b2b8bf0
  Yineng Zhang authored Aug 23, 2025
  
  6b2b8bf0
- Add enable_flashinfer_mxfp4_bf16_moe for higher precision and slower moe backend (#9004) · 0374304a
  fzyzcjy authored Aug 23, 2025
  
  0374304a
- Support GC Freezing to improve latency & throughput (#9241) · 127d4b0d
  Chanh Nguyen authored Aug 22, 2025
```
Co-authored-by: Chanh Nguyen <cnguyen@linkedin.com>
Co-authored-by: Liangsheng Yin <hnyls2002@gmail.com>
```
  127d4b0d
- Add support for extensions of interface and pre-registrations to NIXL HiCache (#9211) · 7e880286
  Moein Khazraee authored Aug 22, 2025
```
Co-authored-by: Zhiqiang Xie <xiezhq@stanford.edu>
```
  7e880286
22 Aug, 2025 17 commits
- Add Qwen3-30B-A3B-Thinking-2507 support on AMD GPUs. (#9456) · c4500233
  sogalin authored Aug 22, 2025
  
  c4500233
- [AMD] Fix Llama 4 FP8 accuracy issues on MI300X (#7699) · f445a1d9
  Hubert Lu authored Aug 22, 2025
  
  f445a1d9
- [MTP] Force greedy sampling on AMD (#9127) · 110a6598
  datdo-msft authored Aug 22, 2025
  
  110a6598
- Use Tensor Core Decode when gqa group size >= 4 (#8624) · 0f587e80
  Wenxuan Tan authored Aug 22, 2025
```
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
```
  0f587e80
- [HiCacheStorage] backup optimization for MLA model (#8865) · 6078d5fc
  huangtingwei authored Aug 22, 2025
```
Co-authored-by: Zhiqiang Xie <xiezhq@stanford.edu>
```
  6078d5fc
- 3fs zerocopy (#9109) · 70cf4abc
  pansicheng authored Aug 22, 2025
```
Co-authored-by: Zhiqiang Xie <xiezhq@stanford.edu>
```
  70cf4abc
- [bugfix] Make --enable-hierarchical-cache and --disable-radix-cache mutually exclusive (#9452) · cebf4599
  Xuchun Shang authored Aug 22, 2025
```
Signed-off-by: Xuchun Shang <xuchun.shang@linux.alibaba.com>
```
  cebf4599
- Disable torch.compile for get_last_loc_large_page_size_large_top_k (#9507) · 9c0c1e30
  Qiaolin Yu authored Aug 22, 2025
```
Co-authored-by: ispobock <ispobaoke@gmail.com>
```
  9c0c1e30
- minor: determine mm attn backend based on platforms (#9303) · a1f011d0
  Mick authored Aug 22, 2025
  
  a1f011d0
- Support speculative decoding in the trtllm_mha attention backend (#9331) · 9ec314c6
  Qiaolin Yu authored Aug 21, 2025
```
Co-authored-by: ispobock <ispobaoke@gmail.com>
```
  9ec314c6
- Add trtllm_mla and cutlass_mla for ragged fmha for chunked prefill (#9480) · 243e745d
  Elfie Guo authored Aug 21, 2025
  
  243e745d
- torch.compile() mrope (#9487) · 61a0e600
  timmy-feng authored Aug 22, 2025
  
  61a0e600
- Support MHA with chunked prefix cache for flashinfer/flashmla backend, support... · 9708d353
  Yongfei Xu authored Aug 22, 2025
```
Support MHA with chunked prefix cache for flashinfer/flashmla backend, support page size > 1 for MHA chunked prefix (#8616)
Co-authored-by: xuyongfei.xyf <xuyongfei.xyf@antgroup.com>
```
  9708d353
- Add Support for Page Size greater than 1 for Flashinfer MLA Backend (#8593) · 3cc3d9b9
  Pavani Majety authored Aug 21, 2025
```
Signed-off-by: Pavani Majety <pmajety@nvidia.com>
```
  3cc3d9b9
- Revert "bugfix: Fix output_ids extraction in detokenizer_manager" (#9467) · 6c855db8
  Xinyuan Tong authored Aug 22, 2025
  
  6c855db8
- feat: update auto_choose_speculative_params (#9470) · 0f9318f7
  Yineng Zhang authored Aug 21, 2025
  
  0f9318f7
- fix: tmp revert gpt oss tp sharding on hopper (#9469) · 849957bc
  Yineng Zhang authored Aug 21, 2025
  
  849957bc
21 Aug, 2025 12 commits
- [FA3] Init Spec Page Table only when Spec is enabled to save ~40MB (#9455) · cded039b
  Stefan He authored Aug 21, 2025
  
  cded039b
- feat: add fused moe config for GLM-4.5-Air-FP8 on B200 (#9463) · 275f9df3
  zixuanzhang226 authored Aug 21, 2025
  
  275f9df3
- Add deepseek v3.1 thinking parser support and update docs (#9464) · e8449ab5
  Xinyuan Tong authored Aug 22, 2025
```
Signed-off-by: Xinyuan Tong <xinyuantong.cs@gmail.com>
```
  e8449ab5
- fix: support fb fp8 (#9462) · 4746aaea
  Yineng Zhang authored Aug 21, 2025
  
  4746aaea
- fix: should return a invalid request response when schema missing (#9461) · 10d34f74
  gongwei-130 authored Aug 21, 2025
  
  10d34f74
- accomendate reasoning_effort set in chat_template_kwargs (#9458) · 9ba72530
  gongwei-130 authored Aug 21, 2025
  
  9ba72530
- [5/n]decouple quantization implementation from vLLM dependency (#9454) · 9c8e4f69
  Hongbo Xu authored Aug 22, 2025
  
  9c8e4f69
- [fix] Fix mxfp4 weight loading bug with TP sharding in GPT-OSS (#9433) · dae9a80f
  hlu1 authored Aug 21, 2025
```
Signed-off-by: Hao Lu <14827759+hlu1@users.noreply.github.com>
Signed-off-by: Xinyuan Tong <xinyuantong.cs@gmail.com>
Co-authored-by: Xinyuan Tong <xinyuantong.cs@gmail.com>
```
  dae9a80f
- Fix quant kernel test errors and benchmark wrong output speeds (#7604) · e85cb1ce
  fzyzcjy authored Aug 21, 2025
  
  e85cb1ce
- Refactor weight offloading logic (#8521) · 55d336cb
  fzyzcjy authored Aug 21, 2025
  
  55d336cb
- ci: enhance xeon ci (#9395) · 029e0af3
  DiweiSun authored Aug 21, 2025
  
  029e0af3
- Enables speculative decoding for the trtllm_mla attention backend (#9238) · 64574ef8
  pranavm-nvidia authored Aug 21, 2025
  
  64574ef8