fzyzcjy
ce647cbcfd
Add a SimplePhaseChecker for execution-phase assertions ( #26803 )
2026-05-31 09:51:39 +08:00
fzyzcjy
7f93952f79
Add a debug toggle for selectively reverting PR fixes ( #26802 )
2026-05-31 09:50:12 +08:00
fzyzcjy
656e75b798
Add a nullcontext placeholder in the forward path for KV-canary ( #26801 )
2026-05-31 09:49:32 +08:00
fzyzcjy
bad83ab427
Fix the EAGLE chunked-prefill next-token chain ( #26329 ) ( #26800 )
2026-05-31 09:48:30 +08:00
fzyzcjy
45194794d0
Apply gemma's position offset out-of-place instead of in-place ( #26799 )
2026-05-31 09:47:50 +08:00
fzyzcjy
13ca55afa3
Make qwen3's set_embed_and_head idempotent ( #26798 )
2026-05-31 09:47:07 +08:00
Liangsheng Yin
9b4be9c574
[core] Compute dimensions/return_pooled_hidden_states in ForwardBatch.init_new ( #26779 )
2026-05-30 17:38:01 -07:00
Bruce Changlong Xu
7ed53d15f3
[config] Recognize custom hybrid SWA models via hf_text_config.is_hybrid_swa ( #23988 )
2026-05-30 23:21:10 +08:00
e279b0bf72
Optimize large add_constant tensors ( #24755 )
...
Co-authored-by: Codex <codex@example.com >
Co-authored-by: BBuf <xiaoyu.zhang@radixark.net >
2026-05-30 22:25:19 +08:00
b421e60eed
【NPU】【bugfix】fix server error when mtp unquant ( #26389 )
...
Co-authored-by: cen121212 <luochen23@huawei.com >
Co-authored-by: Even Zhou <even.y.zhou@outlook.com >
2026-05-30 15:01:19 +03:00
Liangsheng Yin
282c46133f
[Scheduler] Defer prefill input_ids H2D to forward stream, unify resolve via future_map ( #25945 )
2026-05-30 02:58:13 -07:00
Haichuan Hu
acd689b407
feat: add SGLANG_RAY_BUNDLE_INDICES for fine-grained Ray bundle index control ( #24667 )
...
Signed-off-by: Haichuan Hu <kaisennhu@gmail.com >
2026-05-30 02:19:50 -07:00
ZeyuanChen2000
90eb894564
[NPU] fix model llava-onevision-qwen2-7b-ov torch compiles error in npu case ( #26573 )
2026-05-30 17:10:58 +08:00
Bruce Changlong Xu
714bcd84e2
[parallel] Support moe_dense_tp_size == attn_tp_size to share the attention TP group ( #23996 )
2026-05-30 02:10:07 -07:00
a952e9174f
[MLX] Support Qwen3.5 (dense) Model ( #25754 )
...
Signed-off-by: Xiaodong Ye <yeahdongcn@gmail.com >
Co-authored-by: Alex Nails <alex.nails@radixark.ai >
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com >
2026-05-30 17:05:02 +08:00
Qichao Li and Qichao Li
7c5708cba7
[DeepSeek-V4] Add mhc_fused_post_pre kernel ( #25976 )
...
Co-authored-by: Qichao Li <liqichao@baidu.com >
2026-05-30 02:04:51 -07:00
Yuan Luo and luoyuan.luo
0d9a2a9de3
[MoE Refactor] Migrate SM90 Cutlass W4A16 to MoeRunner ( #26489 )
...
Co-authored-by: luoyuan.luo <luoyuan.luo@antgroup.com >
2026-05-30 02:02:56 -07:00
huangtingwei
c93a559e5a
Fix MoE LoRA wrapper exposing moe_runner_config ( #26710 )
2026-05-30 01:59:15 -07:00
3b61a1f935
[Bugfix] Optimize metadata allocation and transfer for mooncake intraNode NVLink ( #26707 )
...
Co-authored-by: 百麒 <yaozhong.lyz@alibaba-inc.com >
Co-authored-by: Shangming Cai <csmthu@gmail.com >
2026-05-30 16:52:01 +08:00
Alison Shao
edfe8d34e8
[CI] ci-coverage-overview: schedule + manual only, include XPU/MUSA/multimodal_gen ( #26619 )
2026-05-30 16:39:18 +08:00
Alison Shao
b4bf489dea
ci: allow /rerun-test to dispatch nightly/weekly CUDA tests ( #26624 )
2026-05-30 16:35:01 +08:00
Jimmy Shong
716e670d3d
[bugfix]: size CuteDSL MoE allgather buffers for the worst-case forward ( #26696 )
2026-05-30 00:27:20 -07:00
Zhangheng
7662210406
[UnifiedTree]: Support eviction priority ( #26549 )
2026-05-30 15:19:43 +08:00
Zhangheng
c048ebd10d
[Hicache]: skip flaky test ( #26764 )
2026-05-30 15:16:12 +08:00
gjsheu
02aeed5387
[NPU] DFlash Speculative Decoding Support NPU ( #23122 )
2026-05-30 15:13:59 +08:00
AndyLi429
fe4b29d391
[Bugfix] Fix Ascend NPU CP attention for batch size > 1 ( #26705 )
2026-05-30 15:07:39 +08:00
silencejade
23a825c694
[DOC] [NPU] add qwen3.5-397b best practice to doc_new ( #26709 )
2026-05-30 14:20:24 +08:00
Liangsheng Yin
1c79015434
Drop dead ScheduleBatch return_routed_experts/return_indexer_topk fields ( #26760 )
2026-05-29 23:16:09 -07:00
6f1c9fc77b
[RL] Fix crash when the reqs in a batch have a mix of return_routed_experts = True and False. ( #26423 )
...
Co-authored-by: root <root@slurm-h200-209-231.slurm-compute.tenant-slurm.svc.cluster.local >
Co-authored-by: Cursor <cursoragent@cursor.com >
Co-authored-by: Cheng Wan <54331508+ch-wan@users.noreply.github.com >
2026-05-29 20:46:10 -07:00
fzyzcjy
804f01a823
Allow PR test and lint workflows to trigger on non-main bases ( #26721 )
2026-05-29 19:22:59 -07:00
Liangsheng Yin
1f850e67f2
[core] Fix crashes on the gpu_only spec_v2 path ( #26738 )
2026-05-29 19:12:03 -07:00
Liangsheng Yin
c2ac37dbcc
[Bug] ngram verify: keep batch.seq_lens_sum in sync after accept ( #26753 )
2026-05-29 18:00:11 -07:00
Erik Wijmans
95cd2fd29f
[lora] More efficient pinned memory ( #20876 )
2026-05-30 09:04:59 +09:00
Bruce Changlong Xu
a5e6a8887a
[attention] Fallback to Triton merge_state when FlashInfer hits CUDA thread limit ( #23993 )
2026-05-29 16:30:49 -07:00
Byron Hsu and Byron Hsu
6ea69efb7f
[RL] Forward Kimi K2.5 weight hooks to language model ( #26744 )
...
Co-authored-by: Byron Hsu <24364830+ByronHsu@users.noreply.github.com >
2026-05-29 15:08:58 -07:00
Chao Shi
6ce49e5f4c
[Utils] Support configure log level at runtime ( #26583 )
2026-05-29 14:49:06 -07:00
cf66693b35
[Model] Add Qwen3-MoE MTP ( #26468 )
...
Co-authored-by: Byron Hsu <byronhsu@Byrons-MacBook-Pro.local >
Co-authored-by: Cursor <cursoragent@cursor.com >
Co-authored-by: root <root@slurm-h200-209-231.slurm-compute.tenant-slurm.svc.cluster.local >
2026-05-29 14:31:05 -07:00
Lianmin Zheng and cctry
4ff1296f5e
Optimize get load calls (/v1/loads) using shared-memory load snapshots ( #26348 )
...
Co-authored-by: cctry <cctry@meta.com >
2026-05-29 13:40:26 -07:00
Qiaolin Yu
3cecc77ccb
[perf] Fuse NVFP4 gate_up_gemm + swiglu + output FP4 quant ( #26626 )
2026-05-29 13:16:24 -07:00
Liangsheng Yin and Qiaolin-Yu
6b5f0d0ccb
[core] Make spec_v2 seq_lens_cpu optional via backend needs_cpu_seq_lens; Triton opts out ( #26128 )
...
Co-authored-by: Qiaolin-Yu <liin1211@outlook.com >
2026-05-29 13:00:32 -07:00
Cheng Wan and Claude Sonnet 4.6
ff8ed7a302
[refactor] unify cuda-graph capture/replay across attention backends ( #26665 )
...
Co-authored-by: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com >
2026-05-29 12:46:42 -07:00
7fb7b41a3e
[docs] Qwen3.5 cookbook: multi-node, MTP TP overrides, dense mamba flag ( #26695 )
...
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com >
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
2026-05-29 12:10:14 -07:00
Bingxu Chen and Cursor Agent
f113ece5cc
Revert "improve: combine vit calls for images from different reqs from one batch ( #25910 )" ( #26442 )
...
Co-authored-by: Cursor Agent <cursoragent@cursor.com >
2026-05-29 11:13:34 -07:00
Cheng Wan and Claude Sonnet 4.6
ec075d8bc5
Fix DRAFT_EXTEND_V2 CG metadata: align test fixture and Triton with production seq_lens convention ( #26651 )
...
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com >
2026-05-29 02:46:45 -07:00
Zhangheng
eb5d4827e8
[UnifiedTree]: Split unified tree kl ci into multiple files to reduce GPU usage. ( #26666 )
2026-05-29 02:46:34 -07:00
akhoroshev
4585f8eb95
[refactor] remove unused op_mlp ( #26673 )
2026-05-29 02:38:56 -07:00
chenxb002 and kjuuii
8652001b6a
fix: use req.req_pool_idx instead of loop variable for req_to_token i… ( #26534 )
...
Co-authored-by: kjuuii <1375341936@qq.com >
2026-05-29 02:34:42 -07:00
Liangsheng Yin
ed85bcf8c3
pin kernels<0.15 ( #26704 )
2026-05-29 01:46:57 -07:00
Teng Ma
544f3039d5
[PD] Fix IB device validation for JSON mappings ( #26114 )
2026-05-29 16:44:49 +08:00
Rita Brugarolas
9062f583db
[ROCm] Eliminate redundant contiguous copy in MLA attention on ROCm MXFP4 ( #25463 )
...
Signed-off-by: Rita Brugarolas Brufau <rita.brugarolasbrufau@amd.com >
2026-05-29 01:28:32 -07:00
fzyzcjy