kvcache-ai/sglang
1
0
Fork 0
mirror of https://github.com/kvcache-ai/sglang.git synced 2026-07-01 12:17:09 +00:00
Code Issues Packages Projects Releases Wiki Activity
12,015 Commits 157 Branches 0 Tags
d84470079dc6b46b9eebebe39bc7226c2ef3ebda
Commit Graph

7965 Commits

Author SHA1 Message Date
Bingxu Chen
d84470079d [AMD] Fix Grok-2 nightly: avoid multimodal misdetection from auto-populated vision_config (#23383) 2026-04-26 21:54:36 -07:00
Jia Guo
bead2e3470 perf: optimize PCG inductor path for FP8 models (redo of #21734) (#23227) 
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
 2026-04-26 20:34:27 -07:00
Byron Hsu
85376a6119 refactor(moe): centralize post-experts all-reduce skip predicate (#23748) 
Co-authored-by: Byron Hsu <byron@periodiclabs.ai>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
 2026-04-26 20:29:59 -07:00
iridiumine
32c3513816 [NPU] Support MTP for Qwen3.5 (#20918) 2026-04-27 10:44:17 +08:00
Kangyan-Zhou
35591c7d51 fix(lora): don't assert on non-LoRA lm_head adapter weights (#23433) 
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
 2026-04-26 12:10:07 -07:00
Mick
a392ae8879 [diffusion] feat: accelerate multiple-outputs generation (#23759) 2026-04-27 01:47:33 +08:00
jianan-gu
10fd0faccd [CPU] Add Qwen3.5 model optimization for CPU (#19484) 
Co-authored-by: Zheng, Beilei <beilei.zheng@intel.com>
Co-authored-by: Ma Mingfei <mingfei.ma@intel.com>
Co-authored-by: Xinyuan Tong <115166877+JustinTong0323@users.noreply.github.com>
 2026-04-26 10:12:36 -07:00
Liwansi
7d49564431 [NPU]Fix support_triton bug (#23604) 2026-04-26 21:34:56 +08:00
Cheng Wan
c7878dbb6d [MoE] Deprecate act_and_mul_triton; fold filter_expert into JIT silu/gelu_and_mul (#23707) 
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
 2026-04-26 01:41:35 -07:00
Mick
d49a0377de [diffusion] refactor: make timestep scheduler request-local (#23716) 2026-04-26 15:59:53 +08:00
sglang-bot
9003f24e2b chore: bump sglang-kernel version to 0.4.1.post1 (#23733) 
Co-authored-by: sglang-bot <sglang-bot@users.noreply.github.com>
Co-authored-by: Kangyan Zhou <zky314343421@gmail.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
 2026-04-25 23:23:49 -07:00
Byron Hsu
ba4e9d2ac2 Apply should_use_dp_reduce_scatterv guard to remaining MoE models (follow-up to #23731) (#23732) 
Co-authored-by: Byron Hsu <byronhsu@noreply.github.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-authored-by: Kangyan-Zhou <zky314343421@gmail.com>
 2026-04-25 20:36:16 -07:00
Byron Hsu
71029abd64 Fix Qwen3 MoE: also guard EP all-reduce with not use_reduce_scatter (follow-up to #23731) (#23734) 
Co-authored-by: Byron Hsu <byron@periodiclabs.ai>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
 2026-04-25 20:35:52 -07:00
Byron Hsu
99b59b279c Fix Qwen3 MoE double-reduce when DP attention + EP + reduce_scatterv (#23729) (#23731) 
Co-authored-by: Byron Hsu <byronhsu@noreply.github.com>
 2026-04-25 15:28:28 -07:00
AlbeeSo
e0a4522370 [typo] fix typo in parallel_state (#23710) 2026-04-25 09:33:33 -07:00
Mick
03849496ad jit_kernel: tolerate FA3 kernels without out arg (#23717) 2026-04-25 23:42:33 +08:00
1874.
046c14a3ed [NPU] Support GGUF quantization for Ascend NPU (dense + MoE) (#17883) 
Co-authored-by: ronnie_zheng <zl19940307@163.com>
 2026-04-25 17:16:47 +03:00
gjsheu
e708ea6d94 [diffusion] fix: restore cache-dit support for LTX2 (#23235) 
Co-authored-by: gengjinsong <gengjinsong@huawei.com>
 2026-04-25 18:10:43 +08:00
Aleksi Vesanto
50ce2708ca [diffusion] fix: Fix FLUX.1/2 graph breaks (#23648) 2026-04-25 17:54:52 +08:00
kk
393252f514 [AMD] fused qk gemma norm kernels to reduce four kernels (#23575) 
Co-authored-by: root <root@smci355-ccs-aus-g12-26.cs-aus.dcgpu>
 2026-04-25 00:30:01 -07:00
Артем Савкин
bd523dd60d [NPU] [Bugfix] [Diffusion] Fixed gray images at the generation output (#23266) 
Co-authored-by: ronnie_zheng <zl19940307@163.com>
 2026-04-25 10:20:38 +03:00
Yujing
6175946db7 [Feature]Add MSProbe dump support in SGLang (#18349) 2026-04-25 10:12:50 +03:00
Yujun Dong
21835fb0af [HiCache] Prevent move_hybrid_indices from polluting radix-tree node host state (#23427) 
Co-authored-by: hzh0425 <hzh0425@apache.org>
 2026-04-25 14:27:42 +08:00
DarkSharpness
82254bd9c5 [JIT Kernel] Reland JIT activation (#22094) 
Co-authored-by: Cheng Wan <54331508+ch-wan@users.noreply.github.com>
Co-authored-by: Cheng Wan <chwan@rice.edu>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
 2026-04-24 23:00:28 -07:00
YC Yen-Ching Tseng
adc59325bc [AMD] Optimize MiniMax-M2.5 - enable fused Triton kernel for FP8 KV cache write in aiter decode path (#23620) 2026-04-24 22:23:49 -07:00
YC Yen-Ching Tseng
fb272d27db [AMD] Optimize MiniMax-M2.5 - use aiter biased_grouped_topk for sigmoid scoring in MoE routing (#23611) 2026-04-24 22:18:08 -07:00
Shenxiu Liu
8471c9ebe6 Skip torch.cuda.empty_cache() in weight update flush path (#22998) 2026-04-25 12:41:39 +08:00
Yuhao Yang
4a3fe2a091 model: support parakeet nemotron encoder (#23568) 
Co-authored-by: trangdough <trangtdo22@gmail.com>
 2026-04-25 11:00:23 +08:00
Jackey Hua
465abadd3c Add fused moe triton config for Qwen3.5-397B-A17B-FP8 (#23682) 2026-04-24 18:35:32 -07:00
Xinyi Song
76da28f6d6 [AMD][bugfix] add gate rocm >= 7.2 for bpreshuffle (#23671) 2026-04-24 13:26:16 -07:00
Jia Guo
587fd15bd2 perf: eliminate attention DtoD copy by passing pre-allocated output to FA (#21985) 
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
 2026-04-24 12:05:16 -07:00
Xinyuan Tong
6d03861476 support Hy3 preview (#23533) 
Co-authored-by: pengmeng <pengmeng@tencent.com>
Co-authored-by: Qiaolin-Yu <liin1211@outlook.com>
Co-authored-by: chengvjiang <chengvjiang@tencent.com>
Co-authored-by: russellfeng <russellfeng@tencent.com>
 2026-04-24 12:03:24 -07:00
Lianmin Zheng
6344b546c8 Deprecate --collect-tokens-histogram, auto-collect with --enable-metrics (#23595) 2026-04-24 12:00:16 -07:00
Mick
05696527ea [diffusion] feat: support LoRA for LTX2.3 (#23649) 2026-04-25 01:52:41 +08:00
Kang Yifei
baa0aa670f [HiCache & HybridModel] 3FS backend support DSA & mamba model (#23241) 
Co-authored-by: 墨已 <kangyifei.kyf@alibaba-inc.com>
Co-authored-by: hzh0425 <hzh0425@apache.org>
 2026-04-25 00:48:01 +08:00
Kangrui Du
92d262f710 [diffusion] RL: add per-step rollout options for SDE and trajectory capture (#23151) 2026-04-24 23:26:16 +08:00
Siju Samuel
bca3dd958a [Intel GPU] Enable pipeline parallelism on XPU (#23645) 2026-04-24 19:52:44 +08:00
Yuwei An
60bbb800db [Experimental] Breakable Piecewise Cuda Graph (#22218) 
Signed-off-by: Oasis-Git <ayw.sirius19@gmail.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-04-24 04:33:05 -07:00
Mick
b3b03369a5 [diffusion] fix: unify LTX-2.3 HQ codepath gates for all LTX-2.3 variants (#23624) 2026-04-24 17:44:08 +08:00
Shangming Cai
b8d883398d Revert "[Intel GPU] Enable pipeline parallelism on XPU" (#23641) 2026-04-24 17:36:35 +08:00
Hubert Lu
4cb0c4e1f3 [AMD] Fix memory access fault when --page-size > 1 with speculative decoding on AMD GPUs (#23596) 2026-04-23 23:56:36 -07:00
Mick
cd1fa7506a [diffusion] model: support LTX2.3 high quality pipeline (#23366) 2026-04-24 14:18:20 +08:00
Shaojun Zhou
59724e90a9 model: support Moss-VL (#23454) 2026-04-24 11:14:29 +08:00
Siju Samuel
bf98eb3ab7 [Intel GPU] Enable pipeline parallelism on XPU (#23472) 
Co-authored-by: Shangming Cai <csmthu@gmail.com>
 2026-04-24 10:41:51 +08:00
popsiclexu
b35213be11 [MUSA][16/N] Add MUSA backend support for layers and DeepSeek models (V2/V3/R1) (#22774) 
Co-authored-by: popsiclexu <zhenxue.xu@mthreads.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
 2026-04-23 18:59:51 -07:00
R0CKSTAR
87e50f20f6 [Apple Silicon][MLX] Cache seq_lens-derived tensors in BatchedDecodeContext (#23470) 
Signed-off-by: Xiaodong Ye <yeahdongcn@gmail.com>
 2026-04-23 18:12:26 -07:00
Mick
c0166355ae [diffusion] CI: minor refactor CI (#23576) 2026-04-24 08:48:31 +08:00
Cheng Wan
d9c72bdd2b Skip unselected experts in flashinfer_trtllm (#23493) 2026-04-23 17:30:19 -07:00
Cheng Wan
000a2525e1 Move expert_mask_gpu from FusedMoE layer to StandardDispatcher (#23585) 
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
 2026-04-23 17:17:27 -07:00
Lianmin Zheng
95d021b523 Pre-set SWA cache location in CudaGraphRunner (#23552) 2026-04-23 16:51:29 -07:00