Commit Graph

4977 Commits

Author SHA1 Message Date
Binyao Jiang
cf0478d602 [Glm46v] Bug fix for accuracy drop and unable to launch server (#14585)
Co-authored-by: yhyang201 <yhyang201@gmail.com>
Co-authored-by: zRzRzRzRzRzRzR <2448370773@qq.com>
Co-authored-by: Minglei Zhu <mingleizhu1122@gmail.com>
2025-12-07 23:45:02 -08:00
Tiwei Bie
36361adcbf [DLLM] Add initial cuda graph support (#14203) 2025-12-08 14:12:35 +08:00
Qiaolin Yu
661e9775d0 [2/2] Add rope kernel in sgl-kernel (#14452) 2025-12-07 21:37:29 -08:00
Nicholas
f57d4fe78e [feat] use cachebuffer to store mm feature to speedup hash (#14386) 2025-12-08 10:35:20 +08:00
wentx
b7b7524e95 [Tool Call] Fix DeepSeekV32Detector skipping functions with no params in streaming mode (#14573) 2025-12-07 18:32:10 -08:00
Xiaoyu Zhang
03b835e7d1 Refactor tuning block wise kernel and opt Qwen/Qwen3-VL-32B-Instruct-FP8 (#14141) 2025-12-08 09:24:58 +08:00
b8zhong
3b47973af8 [CI] Tiny speed up VLM CI (#14517)
Co-authored-by: Brayden Zhong <b8zhong@users.noreply.github.com>
2025-12-07 13:30:41 -08:00
Hudson Xing
84efe54bc4 Fix FP8 KV Triton type issue and add regression test (#14553) 2025-12-07 10:51:46 -08:00
khalilzhk
948b6acee8 [BugFix] fix prefixcache performance and accuracy on ascend (#13573) 2025-12-08 02:16:20 +08:00
Vladimir Serov
f124539a01 [NPU]LoRA: Adding Torch Native backend (#14132) 2025-12-08 02:16:07 +08:00
AichenF
c8683ae305 [diffusion] cli: profiling utilities support (#14185)
Co-authored-by: jianyingzhu <53300651@qq.com>
Co-authored-by: Jianying <53503712+jianyingzhu@users.noreply.github.com>
Co-authored-by: Mick <mickjagger19@icloud.com>
2025-12-08 00:59:45 +08:00
Liangsheng Yin
125e17efd5 Add small model test for spec v2 + dp + trtllm_mla (#14576) 2025-12-07 23:55:00 +08:00
Liangsheng Yin
88c459c6c8 Tiny remove wrong import from python.sglang (#14577) 2025-12-07 22:07:49 +08:00
Xiaoyu Zhang
ae6a6630e4 Add Expert Parallelism (EP) support for kimi-k2-thinking (#13725) 2025-12-07 20:28:57 +08:00
Yuan Luo
26d95008b6 [apply][2/2] Fused qk_norm_rope for Qwen3-MoE (#13998)
Co-authored-by: luoyuan.luo <luoyuan.luo@antgroup.com>
2025-12-07 20:25:18 +08:00
Tiwei Bie
9abcab3ffa [DLLM] feat: Add threshold based parallel decoding support (#14412)
Co-authored-by: Jinwei Yao <jinweiy@illinois.edu>
Co-authored-by: 赵晨阳 <zhaochen20@outlook.com>
2025-12-07 18:25:33 +08:00
Alison Shao
41d61faa99 [FLA] Add explicit kernel arguments to kda.py for Kimi Linear support (#14561) 2025-12-06 22:45:41 -08:00
Chen1022
3c7886ec4c Fix attention backend logic for Qwen3-Next on SM100 (#14560) 2025-12-06 22:03:34 -08:00
b8zhong
6d5d76ad97 remove unecessary dual stream token threshold from the rest of models (qwen moe, kimi linear, etc.) (#14337) 2025-12-06 19:57:26 -08:00
Rain H
32a32cf7d2 Enhance prefill PP node robustness (#14494) 2025-12-06 18:00:54 -08:00
Minglei Zhu
be4a3ec376 support piecewise cuda graph for Olmo models (#14476) 2025-12-06 17:57:44 -08:00
almaslof
ff6e3ea934 [docs] Add missing word in argument description (#14205) 2025-12-06 17:56:54 -08:00
sglang-bot
d2b42477c7 chore: bump sgl-kernel version to 0.3.18.post3 (#14518) 2025-12-06 13:15:16 -08:00
Baizhou Zhang
9dfa01a435 [Misc]Register and refactor some environs for dpsk-fp4 and DeepEp (#14538) 2025-12-06 12:29:16 -08:00
Hanming Lu
e592ee6545 [Qwen3-next] remove heuristics and add radix cache kl test (#14520) 2025-12-06 12:11:40 -08:00
Baizhou Zhang
bc388471d2 [1/n] Fix hanging during DeepGemm Warmup (#14493) 2025-12-06 10:44:02 -08:00
gongwei-130
3e40c63674 fix "GrammarMatcher has terminated after accepting the stop token, but is trying to find the next token mask" when both reasoning and spec are enabled (#14464) 2025-12-06 06:15:22 -08:00
WenhaoZhang
80122e4f4c [diffusion] lora: fix LoRA dtype handling and weight attribute access for z-image model (#14543)
Co-authored-by: niehen6174 <nihen6174@gmail.com>
2025-12-06 22:14:44 +08:00
Xiaoyu Zhang
6d41791823 [diffusion] perf: add QKV fusion optimization for Flux models (#14505)
Co-authored-by: Mick <mickjagger19@icloud.com>
2025-12-06 20:44:16 +08:00
Mick
35a9a07370 [diffusion] refactor: simplify sampling params' override logic (#14539) 2025-12-06 20:23:49 +08:00
Rain Jiang
ea177372bd support mtp with deepseek r1 nvfp4 model (#13115)
Co-authored-by: Trevor Morris <tmorris@nvidia.com>
2025-12-06 00:45:54 -08:00
Baizhou Zhang
42fcf5438f Revert "tiny remove deprecated endpoint call" (#14533) 2025-12-05 23:48:54 -08:00
Mick
d881f31488 [diffusion] chore: temporarily upgrade diffusers to make Z-image compatible with Cache-DiT (#14530) 2025-12-06 12:39:37 +08:00
Vincent Zhong
2ac5b98395 fix: fix rmsnorm -> layernorm in qwen3 omni (#11791)
Co-authored-by: Brayden Zhong <b8zhong@users.noreply.github.com>
2025-12-06 12:12:57 +08:00
Alison Shao
b988c18eae Fix safetensors validation to catch corruption after download (#14465) 2025-12-05 16:04:00 -08:00
fzyzcjy
3d1b591aa1 Tiny use trtllm_mha as default when possible (#14291) 2025-12-05 14:26:03 -08:00
b8zhong
ec7b2c16d9 tiny remove deprecated endpoint call (#13607) 2025-12-05 09:54:49 -08:00
Hudson Xing
38daa29466 Add fused FP8 KV cache write kernel for TRTLLM MHA backend (#14093)
Co-authored-by: Qiaolin Yu <liin1211@outlook.com>
2025-12-06 00:53:55 +08:00
blahblah
66984a8b3d [diffusion] feat: support cache-dit integration (#14234)
Co-authored-by: shuxiguo <shuxiguo@meituan.com>
Co-authored-by: DefTruth <qiustudent_r@163.com>
Co-authored-by: Mick <mickjagger19@icloud.com>
2025-12-06 00:52:22 +08:00
roikoren755
889b46ea50 [Spec] Mamba2 support in target models (#13434) 2025-12-06 00:50:46 +08:00
Mick
a89045603b [diffusion] chore: set allowing overriding protected fields of sampling params as default behavior (#14471) 2025-12-06 00:22:42 +08:00
Alison Shao
662809874c Add Mistral Large 3 to nightly CI tests (#14459) 2025-12-05 23:16:27 +08:00
elvischenv
205f041e96 Add Mistral Large 3 Eagle Support (#14466)
Co-authored-by: Linda-Stadter <57756729+Linda-Stadter@users.noreply.github.com>
2025-12-05 23:11:41 +08:00
Simo Lin
7235a7fbe9 [misc] add model arch and type to server info and use it for harmony (#14456) 2025-12-05 06:51:00 -08:00
Yuxuan Zhang
8fce9e7b2a support GLM-V vision model dp (#14097) 2025-12-05 21:03:54 +08:00
Xiaoyu Zhang
5347732219 [diffusion] fix: Fix profiler trace missing Python stack in diffusion pipeline (#14499) 2025-12-05 12:12:35 +00:00
roikoren755
2ce121a1c3 Enable RadixCache for Mamba2 models (#13584) 2025-12-05 18:23:58 +08:00
WenhaoZhang
35ba6fe19e [diffusion] fix: fix CLIP text encoder attention mask not used (#14364)
Co-authored-by: niehen6174 <niehen.6174@gmail.com>
Co-authored-by: Mick <mickjagger19@icloud.com>
2025-12-05 16:30:10 +08:00
GMI Xiao Jin
7c744d137d [diffusion] cli: add argument --adjust-frames and --override-protected-fields (#13996)
Co-authored-by: dev <devnull@example.com>
Co-authored-by: Mick <mickjagger19@icloud.com>
2025-12-05 15:32:06 +08:00
zyksir
46b05ef58f [diffusion] fix: fix bug about pin memory when offloading (#14472) 2025-12-05 15:26:30 +08:00