Commit Graph

6437 Commits

Author SHA1 Message Date
Lianmin Zheng
20ed3822bb [Auto Sync] Update piecewise_cuda_graph_runner.py (20260119) (#17313)
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
Co-authored-by: Binyao Jiang <byjiang1996@gmail.com>
2026-01-20 14:05:05 -08:00
Baizhou Zhang
6ea491e439 Overlap shared experts with deepep dispatch for single batch overlap on Blackwell (#17289) 2026-01-21 02:56:55 +08:00
GuoweiWangU
16802fb6b2 [FIX] fix mambaish model pp kv cache compute (#17334) 2026-01-20 23:59:05 +08:00
Ke Bao
d97066d209 Use attn_tp_group for all reduce in token embedding (#17403) 2026-01-20 23:56:21 +08:00
Ke Bao
ce2d686e94 Disable swa memory for trtllm-mha backend (#17429) 2026-01-20 23:55:00 +08:00
Qiaolin Yu
76b06bee03 [New Model] GLM4.7-Flash (#17247)
Co-authored-by: zRzRzRzRzRzRzR <2448370773@qq.com>
Co-authored-by: JustinTong0323 <justinning0323@gmail.com>
Co-authored-by: Xinyuan Tong <xinyuantong.cs@gmail.com>
Co-authored-by: Baizhou Zhang <sobereddiezhang@gmail.com>
2026-01-20 23:44:16 +08:00
BingjiaWang
612026ad2c [OPT] DeepSeekV3.2: optimize indexer weight_proj-mma performance (#17205) 2026-01-20 23:09:44 +08:00
Baizhou Zhang
55c616427d Add flag that enables NCCL mlp sync batch for overlap scheduler (#17288) 2026-01-20 23:06:55 +08:00
Ke Bao
91a4cd8648 Update default attn backend to fa3 for mimo model (#17419) 2026-01-20 21:22:27 +08:00
ympcMark
f7a5e425c3 [3/N] Achieve fault tolerance at the DP level (#11657)
Co-authored-by: UNIDY <unidy2002@outlook.com>
Co-authored-by: Hank Han <hanhan7630@outlook.com>
2026-01-20 18:47:08 +08:00
Yuan Luo
e6b7c04947 [Kimi-Linear] Refactor kimi-linear gate calculation to avoid duplicated code (#17160)
Co-authored-by: luoyuan.luo <luoyuan.luo@antgroup.com>
2026-01-20 14:29:24 +08:00
Thomas Wang
6988a0f570 Disable mla persistent kernel when not using fp8 kv_cache (#17327) 2026-01-19 20:29:54 -08:00
Ratish P
c560e14216 [diffusion] fix: enforce 16-pixel alignment for flux2 to prevent shape mismatch crash (#17302) 2026-01-20 12:03:33 +08:00
Shangming Cai
71cb9d0302 [PD] Fix PP dynamic chunking with DP attention (#17339)
Signed-off-by: Shangming Cai <csmthu@gmail.com>
2026-01-20 11:27:13 +08:00
shuwenn
8fb45523f3 feat: support bitsandbytes quantization algorithm (#15325) 2026-01-19 18:36:56 -08:00
Raghav Ravishankar
84aef37808 fix function calling for Trinity (#17364) 2026-01-19 18:33:17 -08:00
Aurick Qiao
09a9d214f7 Pipe customized_info through CudaGraphRunner output (#17088) 2026-01-19 15:48:49 -08:00
Hudson Xing
c1282da236 fix(ci): apply MMMU retry logic to all affected test files (#17329) 2026-01-19 10:00:29 -08:00
zhangheng
20b0523eca [RadixTree][1/N Refactor]: Support unified match_prefix params (#17142)
Co-authored-by: yizhang2077 <1109276519@qq.com>
Co-authored-by: pansicheng <sicheng.pan.chn@gmail.com>
2026-01-19 22:39:40 +08:00
Ke Bao
ce8a6ac690 Evict swa kv cache during decoding (#17220) 2026-01-19 22:36:52 +08:00
ybyang
ebca5879a1 Fix v32 continue_final_message not work (#16567) 2026-01-19 21:45:39 +08:00
Xiaoyu Zhang
cc410a1088 [Diffusion] Apply qknorm to flux2 and apply lightx2v rms_norm_one_pass kernel(without residual) (#17305)
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
2026-01-19 21:25:33 +08:00
b8zhong
f374623fa9 [Refactor] Set fp4-gemm-backend=auto on SM100 and rename fp4-gemm-backend with flashinfer_ prefix (#17309) 2026-01-19 20:09:07 +08:00
Shu Wang
5c02217746 Inclusion of nvfp4 blockscale in EPLB Rebalance (#17158) 2026-01-19 17:45:27 +08:00
Xiaoyu Zhang
a3d9a21882 Revert "[Perf] fuse q, k norm for Flux2Attention (#17241)" (#17332) 2026-01-19 15:24:11 +08:00
Minglei Zhu
64946679b5 [Perf] fuse q, k norm for Flux2Attention (#17241)
Co-authored-by: Minglei Zhu <zminglei@linkedin.com>
2026-01-19 14:33:34 +08:00
Kartik Ramesh
5836324c55 KV Cache Events with Attention DP bug fix (#16030) (#16412) 2026-01-19 14:13:48 +08:00
yudian0504
9fe56cd0fb Fix kernel selection in biased_grouped_topk_gpu (#17325) 2026-01-19 14:06:11 +08:00
Gaoji Liu
858a4d659b support new qwen3_coder_detector (#16744)
Co-authored-by: liugaoji.lgj <liugaoji.lgj@alibaba-inc.com>
2026-01-18 21:16:27 -08:00
Lianmin Zheng
e619f53113 [Auto Sync] Update tokenizer_manager.py (20260119) (#17317)
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
2026-01-18 20:32:19 -08:00
Yongfei Xu
d2105d4abd [DeepSeek v3.2] Opt MTP decode cuda batch sizes and nsa implementation (#16961) 2026-01-19 11:54:11 +08:00
Lee Nau
84c8390514 Use dsv3 optimized routing fused_topk_deepseek instead of moe_fused_gate (#15347) 2026-01-19 11:50:16 +08:00
Baizhou Zhang
ea879c7739 [Minor] Correct sglang version when installing from source (#17315) 2026-01-18 19:36:16 -08:00
Shangming Cai
0227db8926 [PD] Optimize MHA models pp util calculation logic (#17306) 2026-01-19 11:23:15 +08:00
Glen Liu
ad1b4e4728 [Feature] overlap LoRA weight loading with compute (#15512) 2026-01-19 10:43:17 +08:00
Koushik Dutta
d3eafc7357 [GLM 4.7] Add RTX 6000 Pro aka sm120 (#17235)
Co-authored-by: root <root@ubuntu-nvidia.localdomain>
2026-01-18 13:33:19 -08:00
Jinyan Chen
e00b43442d [jit-kernel] Add CuTe DSL GDN Decode Kernel (#15631)
Co-authored-by: Jinyan Chen <jinyanc@nvidia.com>
2026-01-18 12:54:36 -08:00
Todobe
733de6be31 [NPU]Support GPT-OSS for NPU (#14197) 2026-01-19 04:13:41 +08:00
Jerry Ji
93433726eb [Refactor] Split out deepseek v2 weight loader function into mixin (#16649) 2026-01-19 00:41:59 +08:00
Xiaoyu Zhang
330605cc88 [Diffusion] Apply jit qk_norm to flux1 (#17296) 2026-01-19 00:28:36 +08:00
Ch3ngY1
bb6055b43c [Bugfix] Fix PD accuracy when MTP is not configured on the prefill node (#17212)
Co-authored-by: Shangming Cai <csmthu@gmail.com>
2026-01-18 23:37:58 +08:00
b8zhong
4df74eb576 [Refactor] Add -fp4-gemm-backend to replace SGLANG_FLASHINFER_FP4_GEMM_BACKEND (#16534)
Co-authored-by: Vincent Zhong <207368749+vincentzed@users.noreply.github.com>
2026-01-18 23:25:46 +08:00
Xinyuan Tong
2069050d3f fix: Handle multiple named chat templates in HuggingFace tokenizers (#17236)
Signed-off-by: Xinyuan Tong <xinyuantong.cs@gmail.com>
2026-01-18 17:20:04 +08:00
YAMY
a45e0e5df4 [SPEC_V2] Enable cudagraph draft_extend for trtllm_mla_backend and Acclen Fix for DP under cudagraph mode (#16974) 2026-01-18 15:56:21 +08:00
Changyi Yang
8fd3399880 [diffusion] fix: set guidance_scale default to None (#17182) 2026-01-18 15:04:11 +08:00
Yuan Luo
6d29d8ab16 [VLM][Reland] Refactor load_mm_data to improve performance (#16152)
Co-authored-by: luoyuan.luo <luoyuan.luo@antgroup.com>
2026-01-18 14:11:17 +08:00
Ke Bao
e499258e97 Use swa radix cache and memory pool for gpt-oss model (#17261) 2026-01-18 13:47:07 +08:00
Mick
09491a9bcd cli: support sglang version (#17250) 2026-01-18 13:20:24 +08:00
Lancer
e486a4dac1 [diffusion] feat: support default 4-step inference for Flux2-Klein distilled models (#17225)
Signed-off-by: Lancer <maruixiang6688@gmail.com>
2026-01-18 10:15:50 +08:00
Hudson Xing
90399cbc07 fix(ci): recover from corrupted MMMU parquet cache (#17256) 2026-01-17 17:32:01 -08:00