Commits · 3f87f83116ccda4738b98db7a6dee4cfec3c78e4 · zhaoyu6 / sglang · GitLab

23 Apr, 2025 1 commit
- Fuse q_a_proj and kv_a_proj (#5619) · 3f87f831
  Baizhou Zhang authored Apr 22, 2025
  
  3f87f831
22 Apr, 2025 6 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
- Fix FA3 DeepSeek prefill performance regression (#5624) · 4418f599
  JieXin Liang authored Apr 22, 2025
```
Co-authored-by: ispobock <ispobaoke@gmail.com>
```
  4418f599
- fix: gemma 3 not use softcap (#5622) · 04f2abcb
  Yineng Zhang authored Apr 22, 2025
  
  04f2abcb
- [fix] fix compile_deep_gemm missing kv_b_proj (#5620) · 506be6b8
  JieXin Liang authored Apr 22, 2025
  
  506be6b8
- Remove extra copy in deepseek forward absorb (#5578) · 11b23ae9
  Ke Bao authored Apr 22, 2025
```
Co-authored-by: saienduri <saimanas.enduri@amd.com>
```
  11b23ae9
21 Apr, 2025 4 commits
- [feature] enable pre compile jit deep_gemm (#5580) · c2942907
  JieXin Liang authored Apr 22, 2025
  
  c2942907
- [misc] remove is_cuda_available (#5319) · 97cb762b
  JieXin Liang authored Apr 21, 2025
  
  97cb762b
- Speedup shared expert weight construction by avoid cloning (#5188) · 5239d795
  fzyzcjy authored Apr 21, 2025
  
  5239d795
- Fix one more issue reported by torchfix (#4859) · b868526d
  Brayden Zhong authored Apr 20, 2025
  
  b868526d
20 Apr, 2025 3 commits
- enable DeepSeek V3 shared_experts_fusion in sm90 (#5571) · d9dd5298
  Xiaoyu Zhang authored Apr 21, 2025
  
  d9dd5298
- Fix BumpAllocator error when no input_ids (#5564) · 0a0dd34e
  fzyzcjy authored Apr 20, 2025
  
  0a0dd34e
- [perf] introduce deep gemm group_gemm_masked as bmm (#5432) · 99456bca
  JieXin Liang authored Apr 20, 2025
  
  99456bca
19 Apr, 2025 2 commits
- Remove one kernel in per_tensor_quant_mla_fp8 (#5549) · 613b197e
  fzyzcjy authored Apr 20, 2025
  
  613b197e
- simplify the control logic for using shared experts fusion (#5504) · d58e3544
  Xiaoyu Zhang authored Apr 20, 2025
  
  d58e3544
18 Apr, 2025 3 commits
- [Model] Adding Qwen3 and Qwen3MoE (#4693) · 4db463b1
  yhyang201 authored Apr 19, 2025
  
  4db463b1
- Introduce moe_dense_tp_size to fix dense layer errors in DeepSeek V3 + 4x8xH100 (#4836) · 53dcf388
  fzyzcjy authored Apr 18, 2025
  
  53dcf388
- Configuration qwen2_moe.py - qkv_bias now in transformers (#5512) · 1effba4c
  Michael Feil authored Apr 17, 2025
  
  1effba4c
17 Apr, 2025 5 commits
- [Bug fix] use correct func path in deepseek (#5496) · 06d0a3d9
  Xuchun Shang authored Apr 17, 2025
```
Signed-off-by: Xuchun Shang <xuchun.shang@linux.alibaba.com>
```
  06d0a3d9
- Refactor DeepSeek decoder layer branches (#5205) · 8beb356f
  fzyzcjy authored Apr 17, 2025
  
  8beb356f
- Feat/support encoder model (like bert) (#4887) · 3bface15
  woodx authored Apr 17, 2025
  
  3bface15
- Deprecate disable-mla (#5481) · 4fb05583
  Baizhou Zhang authored Apr 17, 2025
  
  4fb05583
- [Model Support] unsloth/Phi-4-mini bnb model (#4982) · 8f783c19
  eigen authored Apr 16, 2025
```
Co-authored-by: yhyang201 <yhyang201@gmail.com>
Co-authored-by: Liangsheng Yin <hnyls2002@gmail.com>
Co-authored-by: Chayenne <zhaochen20@outlook.com>
Co-authored-by: Yineng Zhang <me@zhyncs.com>
```
  8f783c19
16 Apr, 2025 3 commits
- Clean up imports (#5467) · 177320a5
  Lianmin Zheng authored Apr 16, 2025
  
  177320a5
- Support MHA with chunked prefix cache for DeepSeek chunked prefill (#5113) · a42736bb
  Baizhou Zhang authored Apr 15, 2025
  
  a42736bb
- Support BNB quantization for llama/mllama (#5038) · bc24205b
  ryang authored Apr 16, 2025
```
Co-authored-by: Yuhao Yang <yyh073@foxmail.com>
```
  bc24205b
15 Apr, 2025 1 commit
- [fix/misc] remove duplicate row in deepseek v2 model (#5279) · ee9d6ca6
  Yangcheng Li authored Apr 15, 2025
  
  ee9d6ca6
14 Apr, 2025 2 commits
- [perf] experimental enhance fp8 per-tensor quant (#5370) · bdde2375
  JieXin Liang authored Apr 15, 2025
  
  bdde2375
- Support for Qwen2.5-VL Model in bitsandbytes Format (#5003) · 072df753
  yhyang201 authored Apr 14, 2025
  
  072df753
13 Apr, 2025 1 commit
- [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
12 Apr, 2025 2 commits
- 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
11 Apr, 2025 2 commits
- [fix] fix mrope positions not picked up (#5265) · e53a0b3d
  Mick authored Apr 11, 2025
  
  e53a0b3d
- Optimize attention in llama4 (#5127) · cd7e32e2
  fzyzcjy authored Apr 11, 2025
  
  cd7e32e2
10 Apr, 2025 1 commit
- Fix DeepSeek error when using DeepEP mode (#5190) · e3c4bd31
  fzyzcjy authored Apr 10, 2025
  
  e3c4bd31
09 Apr, 2025 4 commits
- model: support mllama4 (#5144) · fbebcb7a
  Mick authored Apr 10, 2025
  
  fbebcb7a
- Support Llama4 fp8 inference (#5194) · 40652482
  HandH1998 authored Apr 09, 2025
```
Co-authored-by: laixinn <xielx@shanghaitech.edu.cn>
Co-authored-by: sleepcoo <sleepcoo@gmail.com>
Co-authored-by: zhyncs <me@zhyncs.com>
```
  40652482
- Optimize topk operation in llama4 (#5128) · 86a876d8
  fzyzcjy authored Apr 09, 2025
  
  86a876d8
- fix: use DeepEPDispatcher on CUDA (#5180) · 90caf06c
  Yineng Zhang authored Apr 08, 2025
  
  90caf06c