This website requires JavaScript.
11e661fd45
[Fix] Don't free the multi-CTAs KV counter the decode graphs captured (#39175 )
2026-09-21 19:58:48 +00:00
44bdf225d8
Fix lint failure from MXFP8 reserved-slot test location (#40618 )
Cheng Wan
2026-09-21 12:31:11 -07:00
bccf691b22
Bringing the parallel runtime up becomes a phase, not a side effect (#40345 )
Cheng Wan
2026-09-21 12:29:50 -07:00
1d3243d05f
Take the parallel getters off the package's public surface (#40344 )
Cheng Wan
2026-09-21 12:27:50 -07:00
970e946e4f
Retire the per-runner parallel record (#40343 )
Cheng Wan
2026-09-21 12:26:40 -07:00
73f071db52
Deprecate the parallel getters the context answers, and ratchet them shut (#40342 )
Cheng Wan
2026-09-21 12:25:32 -07:00
65be3fa71a
A runner and the objects it builds freeze the placement they describe (#40341 )
Cheng Wan
2026-09-21 12:24:17 -07:00
2d0e94e3a3
Check the topology identities where the layout is written, and build at the published widths (#40340 )
Cheng Wan
2026-09-21 12:22:59 -07:00
d5fdab7022
chore: add NIXL owners and CI access (#40602 )
ishandhanani and Kangyan-Zhou
2026-09-21 12:21:30 -07:00
0db1a93adb
State the draft's whole topology in its scope, and read the rest from the context (#40339 )
Cheng Wan
2026-09-21 12:19:38 -07:00
ae7a516ba7
feat: use XGrammar V4.1 DSML parameter constraints (#39026 )
2026-09-21 15:12:28 -04:00
f0940fe3a6
Update DeepSeek-V4 Pro for B200 FP4 agentic PD disaggregation (#40610 )
Faradawn Yang
2026-09-21 12:10:03 -07:00
90b3f8544c
[AMD] Use Triton softmax routing for Qwen3.5 on gfx950 (#39986 )
jacky.cheng
2026-09-22 03:02:29 +08:00
f702a0be29
Revert "[Diffusion] migrate the whole _register_configs from registry.py to the model own config file" (#40611 )
ronnie_zheng
2026-09-21 21:05:19 +03:00
632919e498
[AMD] Fix DeepSeek-R1-MXFP4 accuracy with AITER FP8 (#37762 )
Bingxu Chen
2026-09-22 01:51:37 +08:00
e6931ca889
[Diffusion] migrate the whole _register_configs from registry.py to the model own config file (#40475 )
ronnie_zheng
2026-09-21 20:45:53 +03:00
7a6191c4b9
Preallocate HiCache MHA staging before post-capture KV sizing (#40256 )
cctry
2026-09-21 10:44:29 -07:00
7ad55e4386
[HiCache] TMA-staged host<->device KV transfer kernel (sm_90+) (#40278 )
cctry
2026-09-21 10:38:23 -07:00
0cb37c018c
[KDA] Enable ReplaySSM for GLM-5.3 Flash (#40517 )
William Hu
2026-09-21 13:34:08 -04:00
3c71bb018a
[AMD] Enable GLM DSA prefill top-k to the v2 kernel (#37889 )
Eric.Chin.AMD and Thomas Wang
2026-09-22 01:31:07 +08:00
5a6a1bb883
[mxfp8-kv] Skip writes to the reserved CUDA-graph padding slot (#35351 )
2026-09-21 13:21:25 -04:00
008470abd8
[sgl-router] Bound streaming lifetimes and release guards on idle disconnect (#40391 )
Kan Wu and Shangming Cai
2026-09-21 10:07:07 -07:00
800613a74b
[Test] Split the serving perf tests by topic into basic_perf/ and route their thresholds through a kit (#40505 )
Liangsheng Yin
2026-09-21 10:05:57 -07:00
14e9c40a72
[Observability] Expose python/rust frontend identity in /server_info (#39993 )
Sage
2026-09-21 18:32:52 +03:00
50ec9702d0
[diffusion] docs: update ComfyUI sections, trimmed examples, and the RTX 5090 DiT-resident recipe (1.42x) for Qwen-Image-2.1 cookbook (#40573 )
2026-09-21 21:23:11 +08:00
69d1e5cfe0
[Docs][NPU] Add MiMo-V2.5-Pro FP4 DFlash best practice on Ascend NPU (#40577 )
iridiumine
2026-09-21 20:32:52 +08:00
b410010087
[NPU] [DOC] Add kimi k3 cookbook for 950PR/DT Series (#40575 )
amote-i
2026-09-21 20:18:03 +08:00
0f6761b54f
[sgl-router] Add SGLang-compatible DeepSeek V4.1 Flash rendering (#40532 )
Kan Wu
2026-09-21 03:35:09 -07:00
0abb251a20
[sgl-router] Match DeepSeek V4 rendering to SGLang (#40530 )
2026-09-21 02:56:41 -07:00
2016f5e7a1
[sgl-router] Add Kimi-K3 rendering with SGLang parity (#40390 )
Kan Wu and Claude Fable 5.1
2026-09-21 02:38:49 -07:00
b86a30afba
[AMD][DI][CI] Add a SPUR cluster profile to AMD DI CI (#40113 )
Zhaoyi Li
2026-09-21 03:56:46 -05:00
8faa2d6731
[NPU][Diffusion] Disable loading latency checks in Ascend fixtures (#40544 )
hanwlax
2026-09-21 16:56:35 +08:00
630b1ef322
[sgl-router] Fix reorg admission proxy test build after BucketResolver::new (#40537 )
2026-09-21 16:42:29 +08:00
8d08dfdab7
[Fix] Add gigachat35 to the tool-call and reasoning parser name lists (#40554 )
Mohammad Miadh Angkad and Mohammad Angkad
2026-09-21 08:34:43 +00:00
a9871012ac
[sgl-router] refactor - generalized admission policy definitions (#40271 )
Kan Wu and Claude Fable 5.1
2026-09-21 01:26:30 -07:00
70b5b03e78
[NPU][CI] Fix paths-filter negation that makes every PR run the NPU tier (#40549 )
Shangming Cai
2026-09-21 16:07:27 +08:00
c2f860af1c
[ci][xpu] Record device time in the multimodal_gen perf lane (#39956 )
ashwini rathi
2026-09-21 13:09:29 +05:30
b63f8416b3
[Feature] Gigachat 3.5 support (#29189 )
2026-09-21 09:37:55 +03:00
b54d5b7c7b
disaggregation: Fix FakeKVSender queue accumulation (#28652 )
kpjeeja
2026-09-21 11:57:05 +05:30
f5f3c38aad
[Fix] Preserve YaRN scaling when extending rotary caches (#38786 )
skyler-apdx
2026-09-21 14:14:05 +08:00
d20cd9d77f
[XPU]Enable HiSparse hierarchical sparse KV cache on Intel XPU (#32792 )
AMRUTHA M and Ma Mingfei
2026-09-21 11:37:49 +05:30
fcb080bd40
[sgl-router] refactor - cache-aware policy (#40366 )
Kan Wu and Claude Fable 5.1
2026-09-20 22:53:57 -07:00
11ecdbf39f
Clean up startup logging and streamline log audits (#40526 )
Lianmin Zheng
2026-09-20 22:28:35 -07:00
292e3ccc0c
[ci][xpu] Re-seed the wan2_1_t2v_1.3b perf baseline on Arc Pro B60 (#39955 )
ashwini rathi
2026-09-21 10:34:35 +05:30
c74a4037fb
mooncake: probe pointers via cuda.bindings instead of ctypes libcudart
minke.yu
2026-09-21 12:46:25 +08:00
b963295489
mooncake: support host-memory payloads with intra-node NVLink transport
minke.yu
2026-09-21 11:53:26 +08:00
1da8ac10e1
Update to the cookbook for XPU-supported models (#33649 )
Ziru and Niu Ziru
2026-09-21 11:26:09 +08:00
62ba964848
Fix: post-load staging regression breaks offload meta/sharded_gpu modes (#38779 )
jianzhao-xu
2026-09-21 11:19:07 +08:00
ab03a8e7eb
[Perf] Fork-safe import: no CUDA context at import time, lighter argument parsing (#40201 )
Xueshen Liu
2026-09-20 22:48:35 -04:00
176dbcb85d
[npu]add chunk gdn kernel and unify ssm state layout for ascend gdn backend (#36187 )
zhaozx-cn
2026-09-21 09:51:19 +08:00
6ad78f2281
[diffusion] docs: add verified DGX Spark recipe for Qwen-Image 2.1 (#40487 )
Mick and Mick Qian
2026-09-21 08:49:35 +08:00
b912db67ea
[diffusion] fix: keep Qwen-Image 2.1 prefix KV per layer under Cache-DiT (#40472 )
WenhaoZhang
2026-09-21 08:48:31 +08:00
501b7851e4
[diffusion] CI: guard E2E/loading latency with runner-aware baselines (#39206 )
Mick and Mick Qian
2026-09-21 08:43:22 +08:00
3a0324fb9b
[diffusion] optimization: reduce Qwen-Image 2.1 vae and graph warmup memory (#40481 )
Mick and Mick Qian
2026-09-21 08:42:02 +08:00
76a9065bef
[Fix] Raise on undelivered embeddings in send_with_url, fix broken tests (#40502 )
Liangsheng Yin
2026-09-20 17:35:27 -07:00
4027740569
[sgl-router] refactor - session-aware policy (#40379 )
Kan Wu and Claude Fable 5.1
2026-09-20 17:22:36 -07:00
aedda8377e
[sgl-router] refactor - layout BucketResolver, Bucket, EngineGroup and implement PowerOfTwo (#40241 )
Kan Wu and Claude Fable 5.1
2026-09-20 16:57:07 -07:00
4a9dc5c4af
[sgl-router] refactor - move policy-required states under src/state (#40272 )
Kan Wu and Claude Fable 5.1
2026-09-20 16:49:23 -07:00
acd20a516e
[CI] Give the kernel lane a 5090 suite and move kernel-only tests off the general lane (#40496 )
Liangsheng Yin
2026-09-20 16:46:53 -07:00
42875bcd2a
fix(modelopt): dispatch NVFP4 MoE on the cached backend, not the live global (#38932 )
2026-09-20 16:28:54 -07:00
2fa6b94e34
[Perf] Fuse the glm5_next mHC attn->MLP boundary (#39200 )
Mohammad Miadh Angkad and mmangkad
2026-09-20 23:20:37 +00:00
983e643854
[Feature] support bf16 MoE router and mxfp4 MoE for MiMo V2 (#40448 )
ollybbmonster
2026-09-21 07:13:59 +08:00
2e2d8a2fda
[CI] Drive per-commit stage jobs from a runner table instead of copied job blocks (#40495 )
Liangsheng Yin
2026-09-20 15:43:46 -07:00
d97aed2c90
Fix TopK v2 fallback when 16-block cluster capacity is zero (#40163 )
luoroger37 and Hank Han
2026-09-21 06:41:41 +08:00
f31a7bd45c
Use pinned memory for asynchronous sampling metadata transfers (#39777 )
2026-09-21 06:31:31 +08:00
95521da18d
[DeepSeek-V4.1] Bound dense prefill indexer memory (#40217 )
Harmya Bhatt
2026-09-20 16:43:19 -05:00
c2c3629f2d
[Kimi-K3] O(1) expert weight lookup in load_weights (#38805 )
2026-09-21 05:39:45 +08:00
f6483e479f
[Test] Drop cause-less disabled tests, fix XPU lane, demote quality gates off base-c (#40288 )
Liangsheng Yin
2026-09-20 14:36:02 -07:00
d229952e25
[Fix] Preserve model runner contracts in prefill CUDA graphs (#35452 )
2026-09-20 13:53:14 -07:00
745de73ba3
Add CODEOWNERS entry for sglang-renderer (#40483 )
Shangming Cai
2026-09-21 01:37:50 +08:00
b3e4d198af
[PD] Bound cached-prefix DCP transfers by pack capacity (#40376 )
Khoa Pham
2026-09-20 10:17:15 -07:00
80da4432d0
[Simulator] Fix meta host memory budgets on constrained runners (#40440 )
Shuwen Wang
2026-09-21 00:42:55 +08:00
3dbdd700e2
[CI] update CI permissions (#40474 )
WenhaoZhang
2026-09-21 00:07:37 +08:00
e97614d10c
[Qwen4-Exp] Build the offloaded PLE table on the meta device so --ple-offload-embedding never materialises it on the accelerator (#39928 )
Jimmy Shong and Yangmin Li
2026-09-20 08:28:25 -07:00
5f017ffabb
Update test cases and performance testing framework (#40392 )
ZY Y
2026-09-20 22:42:00 +08:00
404dee10c0
[NPU] add coverage-based precision test selection pipeline (#38339 )
chenyang08056032
2026-09-20 22:40:55 +08:00
8923f4d779
[Test] Fix optimistic prefill disaggregation test after mamba radix cache removal (#40469 )
Mohammad Miadh Angkad and Mohammad Angkad
2026-09-20 14:39:34 +00:00
fa826e08b1
Reuse shared compressed KV dequantization in DeepSeek V4.1 CP prefill
abing
2026-09-19 10:24:01 -07:00
3810f531a8
Support V4.1 decode vision MegaMoE
SYChen123
2026-09-18 19:13:10 +08:00
2580c24d1b
Support V4.1 DP attention in DP-only DSpark PD
SYChen123
2026-09-18 19:11:24 +08:00
791c7850d0
[Diffusion] Enable shared RMSNorm dispatch for SenseNova-U1 (#39705 )
faceless void and ronnie_zheng
2026-09-20 22:08:02 +08:00
4f22146e51
Avoid host synchronization in DeepSeek V4.1 CP prefill
abing
2026-09-18 02:28:12 -07:00
12e3b82e52
run pass cp+megamoe+bcg
abing
2026-09-17 04:04:38 -07:00
8305f66fc8
fix bug
abing
2026-09-19 02:32:27 -07:00
21a4a16b4b
update code
abing
2026-09-17 19:43:41 -07:00
fc954b7e08
add test
abing
2026-09-16 00:26:02 -07:00
c2059c4fb2
run pass llm cp
abing
2026-09-15 23:41:20 -07:00
7b1c2ed0a4
[rust-renderer] Standalone preprocessing (#36718 )
2026-09-20 17:03:12 +03:00
6880a47955
[diffusion] docs: simplify Qwen-Image 2.1 cookbook (#40455 )
Mick and Mick Qian
2026-09-20 20:39:26 +08:00
c610c40399
[sgl-router] refactor - config and organize CLI options (#39867 )
Kan Wu
2026-09-20 04:40:04 -07:00
efa7be2091
[Simulator][Compatibility] Adapt to latest KV cache pool interfaces (#40418 )
Ruiyan Ma
2026-09-20 18:00:06 +08:00
0024efa0de
[CI] Derive registered-test kind from the registry call instead of the path (#40294 )
Liangsheng Yin
2026-09-20 02:01:22 -07:00
671630abf1
[sgl-router] refactor - main startup logic (#39861 )
Kan Wu and Cursor
2026-09-20 01:55:34 -07:00
2a0cb2f04e
[Diffusion][MiniMax-H3] Add SM120 Sage compute for SubBlock sparse attention (#40116 )
HuangJi
2026-09-20 16:40:28 +08:00
414adef060
[CI] skip srt rust extension builds for diffusion-only PRs (#40293 )
Mick and Mick Qian
2026-09-20 16:33:23 +08:00
dc002c85fc
[Test] Fix OOT DFlash hook test resolving the draft config over the network (#40427 )
Liangsheng Yin
2026-09-20 01:25:34 -07:00
9f3d275940
[HiCache] Fix sparse hybrid transfer layer IDs (#37870 )
Shuwen Wang and Seokhoon Kang
2026-09-20 16:24:24 +08:00
e54009240a
[AMD][DSV4] feat: enable DSpark with fp8 unified_kv on gfx950 (#38901 )
amd-danli103 and HAI
2026-09-20 16:16:39 +08:00
a8a4d86be9
Remove swa and mamba radix cache (#40313 )
Ke Bao
2026-09-20 16:16:27 +08:00
5c69e32abe
[NPU] [DOC] fix typos, heading levels and terminology in NPU docs (#40402 )
amote-i
2026-09-20 15:15:58 +08:00