586355a83b
[Refactor] attention: Route every FlashAttention call through the selected kernel ( #46695 )
...
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com >
Co-authored-by: Alex Nails <alex.nails@radixark.ai >
2026-10-27 17:11:58 -07:00
Michael
20b439ca77
[CI] [AMD] miles-nightly: Split the 8-GPU suite into two jobs ( #46750 )
2026-10-28 08:10:08 +08:00
Liangsheng Yin
655059819c
[Fix] dsv41: Use --attn-dp-size in the Engram docs and restore env overrides on exceptions ( #46896 )
2026-10-27 16:49:10 -07:00
zijiexia
e0641da8f8
Revert "[Doc] beam: Use lmsysorg/sglang:dev image in the cookbook" ( #46895 )
2026-10-27 16:34:42 -07:00
2c83dc6656
[Perf] [AMD] vattn-asm: Split small-batch MTP verify attention over all CUs on gfx950 ( #45555 )
...
Co-authored-by: jacky.cheng <yichiche@amd.com >
Co-authored-by: Cursor <cursoragent@cursor.com >
2026-10-27 16:26:16 -07:00
zijiexia and Claude Opus 5.5
ab5602731f
[Doc] beam: Use lmsysorg/sglang:dev image in the cookbook ( #46886 )
...
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com >
2026-10-27 16:25:34 -07:00
Khoa Pham and Claude Opus 5.5
a0d4034d5b
[Fix] dsv4: Build the DSpark draft block metadata without host sequence lengths ( #46891 )
...
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com >
2026-10-27 16:19:02 -07:00
Liangsheng Yin
90a7339419
[Feat] dsv41: Choose where Engram tables live with --engram-config ( #46893 )
2026-10-27 16:15:19 -07:00
f9f2665b6e
[Fix] [AMD] moe/topk: Preserve fp32 bias across aiter routing paths ( #44797 )
...
Preserve fp32 correction bias through select_experts and prepare one logits/bias pair before the AITER versus radix4 dispatch. This keeps precision consistent across the radix4 batch cutoff and avoids an unused bias downcast.
Consolidate CPU dispatch coverage into the existing correction-bias cache test file and cover the real AITER/radix4 path in the existing AMD test file.
Fixes #44794 .
Co-authored-by: Junyi Yao <64238008+Kylinny@users.noreply.github.com >
Co-authored-by: Cheng Wan <54331508+ch-wan@users.noreply.github.com >
Co-authored-by: Emil Bogomolov <25350960+zetyquickly@users.noreply.github.com >
2026-10-27 16:04:43 -07:00
Chunan Zeng and Claude Opus 5.5
7e142dd9fd
[Perf] moe: Always pass the global token count to the Mega MoE tile heuristic ( #46093 )
...
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com >
2026-10-27 15:35:29 -07:00
baa349af7b
[Perf] eplb: Fuse expert-count recording into one kernel on CUDA decode graphs ( #46497 )
...
Co-authored-by: weireweire <20922698+weireweire@users.noreply.github.com >
Co-authored-by: kpham-sgl <khoa.pham@radixark.ai >
2026-10-27 15:28:45 -07:00
Olga Miroshnichenko
3145c7ae66
[Perf] [AMD] qwen3_5: Quantize the BF16 GDN projections to MXFP4 online ( #43839 )
2026-10-27 15:27:16 -07:00
Jzz1943
7f07519943
[Fix] DSA/KDA: Handle MoE padding in hybrid MTP verification ( #43271 )
2026-10-27 15:21:36 -07:00
Liangsheng Yin
010eb729be
[Refactor] bcg: Delete capture leftovers and rename capture_stub to on_capture ( #46792 )
2026-10-27 15:05:58 -07:00
YAMY
b74fd392e1
[Perf] [NVIDIA] speculative: Move the host work between Qwen3.5 MTP decode graphs into the graphs (1/2) ( #45474 )
2026-10-27 14:42:45 -07:00
148bea7302
[Perf] QSA: Remove the D2D copy in prefill generated kv gather ( #45753 )
...
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com >
Co-authored-by: Shunkang <182541032+Shunkangz@users.noreply.github.co >
2026-10-27 14:41:14 -07:00
YAMY
8f1b7691ca
[Feat] [NVIDIA] megakernels: Qwen3.5 binding and EAGLE entry wiring for NVIDIA megakernel plugins ( #45266 )
2026-10-27 14:02:53 -07:00
6be218a0f2
[Fix] rust-server: Report the launched --tokenizer-path from the info endpoints ( #46879 )
...
Co-authored-by: metamergebot <metamergebot@users.noreply.github.com >
Co-authored-by: shodoco <shodoco@users.noreply.github.com >
2026-10-27 13:56:46 -07:00
Liangsheng Yin
bda3718df0
[Refactor] bcg: Record graph breaks at capture without running them ( #46719 )
2026-10-27 13:53:11 -07:00
37e9bdb7a7
[Feat] jit: Split a header-only module's kernel instantiations across parallel translation units ( #46863 )
...
Co-authored-by: metamergebot <metamergebot@users.noreply.github.com >
Co-authored-by: mxz297 <mxz297@users.noreply.github.com >
2026-10-27 13:49:32 -07:00
zanes-ops and Shangming Cai
e80120b4fb
[Fix] eplb: Load all local replicas of each expert ( #45363 )
...
Signed-off-by: Zane <zanes@nvidia.com >
Co-authored-by: Shangming Cai <csmthu@gmail.com >
2026-10-27 13:48:19 -07:00
maocheng23
4a58dc9838
[Feat] scheduler: Add forward observer and deferred responses for plugins ( #43450 )
2026-10-27 13:37:12 -07:00
Jan Schalkamp and Xiaoyu Zhang
a4a19332ab
[Perf] paged-experts: fewer host syncs in host-planned steps ( #46567 )
...
Co-authored-by: Xiaoyu Zhang <1182563586@qq.com >
2026-10-28 03:36:11 +08:00
Avaya Aggarwal and Xiaoyu Zhang
69767e5aa6
[Fix] diffusion: float32 Plucker rays for bf16 camera poses ( #45638 )
...
Co-authored-by: Xiaoyu Zhang <1182563586@qq.com >
2026-10-28 03:35:48 +08:00
Avaya Aggarwal
c45ebb507d
[Fix] diffusion: HunyuanImage-3 MoE shared MLP and timestep schedule on CUDA ( #46600 )
2026-10-28 03:34:46 +08:00
Avaya Aggarwal
dc41126cde
[Fix] diffusion: LoRA format alpha and rank scaling ( #45833 )
2026-10-28 03:34:18 +08:00
Thomas Wang
5044779f0f
[Feat] [AMD] mono-kernels: Register specs by model type ( #46036 )
2026-10-27 12:33:04 -07:00
Qiaolin Yu
41fb719cf2
[Feat] model: Add Flint (RecursiveSI/flint) support ( #45624 )
2026-10-27 12:21:10 -07:00
maocheng23
b068417bfc
[Feat] dflash: Serve SpecForge MLA drafts through the target MLA attention ( #44676 )
2026-10-27 12:14:03 -07:00
e7c55f6959
[Fix] mem_cache: Clamp SWA tombstone recovery to the request's protected prefix ( #46700 )
...
Co-authored-by: metamergebot <metamergebot@users.noreply.github.com >
Co-authored-by: xiezhq-hermann <xiezhq-hermann@users.noreply.github.com >
2026-10-27 11:46:40 -07:00
b28f4fc065
[Perf] disaggregation: Restore host-received decode KV without blocking the scheduler ( #46689 )
...
Co-authored-by: metamergebot <metamergebot@users.noreply.github.com >
Co-authored-by: ZYHowell <ZYHowell@users.noreply.github.com >
2026-10-27 11:46:34 -07:00
b7ac4c0e72
[Perf] radix-tree: Keep the GIL for tree-core flag, size and digest reads unless the core is busy ( #46674 )
...
Co-authored-by: metamergebot <metamergebot@users.noreply.github.com >
Co-authored-by: Jialin <Jialin@users.noreply.github.com >
2026-10-27 11:46:28 -07:00
b48aba306b
[Test] rust-server: Drop two unit tests that mirror one-line delegations ( #46629 )
...
Co-authored-by: metamergebot <metamergebot@users.noreply.github.com >
Co-authored-by: shodoco <shodoco@users.noreply.github.com >
2026-10-27 11:46:22 -07:00
Michael
cb55d77d37
[CI] [AMD] deps: Install openai 2.8.0 in AMD CI for the Responses API ( #46774 )
2026-10-27 09:53:57 -07:00
Michael
ec54c220f8
[CI] [AMD] diffusion: Skip the CUDA-only Nomas fused-kernel and FlexAttention unit tests on ROCm ( #46686 )
2026-10-28 00:37:54 +08:00
Xinyuan Tong
33e8e0fabd
[Doc] cookbook: Add LFM3 cookbook ( #46859 )
2026-10-28 00:18:53 +08:00
Yuhao Yang
de26bb16a9
[Fix] custom-all-reduce: Remove multi-node VMM allocator gate ( #46853 )
2026-10-27 22:53:54 +08:00
shun8686
635430d073
[CI] [NPU] precise-test: skip suites that register no PR test ( #46844 )
2026-10-27 21:16:48 +08:00
7fe5b127e4
[Fix] disaggregation: Avoid extra GPU contexts with PD ( #45678 )
...
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com >
Co-authored-by: Po-Han Huang <pohanh@nvidia.com >
2026-10-27 20:19:23 +08:00
Po-Han Huang (NVIDIA) and Claude Opus 5.5
359bfc1950
[Doc] cookbook: Refresh SemiAnalysis AgentX launches with DeepSeek-V4 B200 DSpark6 disagg and Muse Spark GB300 points ( #46807 )
...
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com >
2026-10-27 19:53:31 +08:00
amote-i
82b90efc84
[Doc] [NPU] docs: update npu docs and fix some issues ( #46736 )
2026-10-27 19:52:56 +08:00
Yihao Wang and Alison Shao
7e5cd57028
[CI] slash-command: Drive opt-in CI tiers from a registry ( #45735 )
...
Co-authored-by: Alison Shao <54658187+alisonshao@users.noreply.github.com >
2026-10-27 04:23:39 -07:00
YC Yen-Ching Tseng
f9257d6c7c
[Perf] [AMD] rccl: upgrade rccl to 2.32.3 for rocm10.1 ( #45576 )
2026-10-27 19:18:52 +08:00
a6f5bd397b
[Perf] [AMD] fp8: Use the AITER CK blockscale GEMM by default on gfx942 ( #43815 )
...
Signed-off-by: siliangchen-amd <SiLiang.Chen@amd.com >
Co-authored-by: Cursor <cursoragent@cursor.com >
Co-authored-by: Xiaoyu Zhang <1182563586@qq.com >
2026-10-27 18:58:11 +08:00
Avaya Aggarwal
8a455f1b90
[Fix] diffusion: pipeline and model-helper bug fixes ( #44888 )
2026-10-27 18:39:21 +08:00
Jan Schalkamp
8063296528
[Perf] paged-experts: fuse slot lookup into decide ( #46457 )
2026-10-27 18:38:59 +08:00
3927a73a34
[Fix] kda: Use raw beta in prefill CUDA graphs ( #44778 )
...
Co-authored-by: Hrithvik Alex <hrithvik@baseten.co >
Co-authored-by: kpham-sgl <khoa.pham@radixark.ai >
2026-10-27 02:29:33 -07:00
9d45fdda46
[Fix] qwen3_coder: Preserve structured values in qwen3_coder mixed unions ( #44124 )
...
Signed-off-by: Ryan Stewart <rystewart@nvidia.com >
Co-authored-by: Po-Han Huang (NVIDIA) <53919306+nvpohanh@users.noreply.github.com >
Co-authored-by: Po-Han Huang <pohanh@nvidia.com >
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com >
2026-10-27 17:01:36 +08:00
922e595a94
[Chore] [NVIDIA] docker: Update to torch 2.15rc1 in CUDA 13.4 container ( #45858 )
...
Co-authored-by: Po-Han Huang <pohanh@nvidia.com >
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com >
2026-10-27 16:52:08 +08:00
weireweire and weireweire
24d55621a3
[Fix] dspark: Draw the simulated-acceptance bonus at the simulated row ( #46404 )
...
Co-authored-by: weireweire <20922698+weireweire@users.noreply.github.com >
2026-10-27 01:48:20 -07:00
