-
a3bf25dc62
[AMD] Clamp MORI intranode grid GPUs (#39763)

inkcherryandDuyi-Wang
2026-09-16 17:13:59 +08:00
-
7d5696b3a1
dsv4.1: communication kernels and wrappers (#39653)
+4









2026-09-16 02:09:40 -07:00
-
5a0c1e21e9
[AMD] Preserve deterministic inference when Lean Attention is enabled (#37740)
Bingxu Chen
2026-09-16 17:09:08 +08:00
-
2cbfaefbf9
[AMD] Avoid the FP8 wo_a path when the weight is BF16 (#38453)
Bingxu Chen
2026-09-16 16:57:15 +08:00
-
5212797880
[NPU] fix npu docker workspace directory (#39732)

Jensenandronnie_zheng
2026-09-16 16:51:38 +08:00
-
fc5a979f21
[misc] Merge FlashInfer autotune caches across spec workers, pad MXFP4 TP shards, drop dead ngram attrs (#39678)

Liangsheng YinandBBuf
2026-09-16 01:50:29 -07:00
-
0e528dc9ff
bugfix:fix unifiedcache c128 radix cache management (#39426)
vstone-w
2026-09-16 16:47:42 +08:00
-
8baeded6f3
[Router] Pin to the prefix owner when the whole fleet is queueing (--saturation-queue-floor) (#39169)


2026-09-16 01:16:24 -07:00
-
ad28b91fae
ci: fix always-failing coverage job, add by-GPU-count view (#39697)
Alison Shao
2026-09-16 01:13:08 -07:00
-
9c8d4641ff
[Fix] Fix GLM5 mHC PP forward (#39720)
Void
2026-09-16 15:32:53 +08:00
-
54f0d72adf
[Router] Fleet-wide sampling contract 2/3: enforce and inject per request (#39001)


2026-09-16 00:29:33 -07:00
-
7835f1de9a
[Router] Add the shutdown-drain configuration surface (#39015)


2026-09-16 00:26:50 -07:00
-
5beb2fd552
[NPU] Avoid device synchronization in Ascend sampling (#39404)
Jensen
2026-09-16 15:15:16 +08:00
-
e2d56bbbfc
[Router] Bound the e2e worker memory budget so prefill graph capture stops OOMing (#39713)


2026-09-16 00:13:09 -07:00
-
d634320e48
[Router] Count open HTTP exchanges until their response body finishes (#39014)


2026-09-16 00:04:38 -07:00
-
faaff1eca8
dsv4.1: Top-k kernels and candidate selection (#39648)
+4









2026-09-15 23:54:35 -07:00
-
a64be2e430
[Kernel] Add H20 block-FP8 MoE configs for GLM-5.3-Flash EP4/EP8 (#38913)
Hank Han
2026-09-16 14:42:18 +08:00
-
6b33338242
[HiCache] Forward prefix metadata to v2 storage calls (#39567)

Shuwen WangandZhangheng
2026-09-16 14:05:42 +08:00
-
91f691c490
dsv4.1: standalone kernels and Python wrappers (#39646)
Liangsheng Yin
2026-09-15 22:31:18 -07:00
-
afde31a2f5
[router] Speak cleartext h2c on both edges: serve it inbound, forward it outbound (#39006)


2026-09-15 22:30:31 -07:00
-
c1f5b4736a
[Router] Add --worker-queue-limit: stop sending cache-affinity traffic to a queueing worker (#39168)


2026-09-15 22:24:48 -07:00
-
3f8eb35ead
[PD] Do not admit intake-rejected requests to a PD handoff (#38935)

Shangming CaiandClaude Opus 5
2026-09-16 12:44:49 +08:00
-
100e1cd0d9
[Fix] Spawn, don't fork, the benchmark server process (#34712)

Dayananda VandClaude Opus 5
2026-09-16 10:14:09 +05:30
-
5f17e3a75f
[Router] Fleet-wide sampling contract 1/3: the config surface (#39000)


2026-09-15 21:39:06 -07:00
-
4678536df6
[CPU] Add fp8_per_tensor_scaled_mm_cpu kernel (#32618)

Xinguo ZhuandAKatydid
2026-09-16 12:31:05 +08:00
-
f60652a43e
[AMD] Update deepseek-v4 PDI and cache policy setting for agentic workload (#39702)
Thomas Wang
2026-09-16 12:16:30 +08:00
-
444b29c932
[Diffusion] Clean up obsolete worker plumbing, dead helpers, and tests (#39293)
Xiaoyu Zhang
2026-09-16 12:11:45 +08:00
-
954bb6804a
[AMD][bugfix] Fix DSV4 MTP crash (#39547)
Xinyi Song
2026-09-15 21:10:23 -07:00
-
537ac52c5b
[SRT] Clean up no-op compiler pass, dead helpers, and migration tests (#39295)
Xiaoyu Zhang
2026-09-16 12:04:21 +08:00
-
6331e43081
[AMD] Prefer HIP Top-K for GLM-5.x on ROCm (#39631)
Zhang, Jiejing
2026-09-15 20:55:57 -07:00
-
a9bb4d7d45
[AMD] Align Qwen3.5 MI355X HiCache cookbook with kernel / page_first (#39572)

jacky.chengandChangLiu0709
2026-09-16 11:22:28 +08:00
-
5f6dd44edc
[NPU] [bugfix] Fix undefined require_attn_tp_allgather (#39561)
chx96642264
2026-09-16 11:02:40 +08:00
-
935cbf24ec
Accept MXFP8 dispatch in FlashInfer A2A TRT-LLM MoE (#39613)

Mohammad Miadh AngkadandMohammad Angkad
2026-09-16 10:23:15 +08:00
-
d4ad368ed9
[Intel XPU] Enable fused_moe_triton tuning on XPU and add tuned DeepSeek-OCR-2 configs (#28723)
Rahul Vijayaraghavan
2026-09-16 07:50:22 +05:30
-
f11cd8ab0e
[XPU] Use torch scaled_mm for XPU block FP8 linear (#35605)
Cao E
2026-09-16 10:10:31 +08:00
-
2cb51f5d22
[CPU] [Diffusion] Add fused scale-shift and norm kernels for CPU (#33452)

Yuxingwang-intelandMa Mingfei
2026-09-16 10:01:29 +08:00
-
f920be4b09
[AMD] GLM-5.2 NextN: cast draft fused MoE to per-channel FP8 (#39155)

Zhang, Jiejingandfanxingran
2026-09-15 18:56:48 -07:00
-
7eedd57ab0
[ROCm] Fix EAGLE spec-decode verify silently sampling greedy on HIP (#37134)


2026-09-16 09:49:49 +08:00
-
2f7d04da5d
dsv4.1: Rust extension modules for image preprocessing, KV pool names, and PD bootstrap (#39677)
Liangsheng Yin
2026-09-15 18:11:48 -07:00
-
dbd7281b06
fix(npu): fix hybrid KV transfer with PP prefill in PD disaggregation (#38402)
Jinyan Yi
2026-09-16 09:10:17 +08:00
-
4e9e407d37
[HiCache] Label radix-cache metrics per rank and split the "shrunk" prefetch reason (#39280)
Zhiqiang Xie
2026-09-15 17:54:04 -07:00
-
c9a8fba991
[DSV4][BCG] Optimize the heavy memory use of C4 Indexer when BCG is enabled (#36534)

Siyuan ChenandYuwei An
2026-09-16 08:23:11 +08:00
-
bbef93e22a
[Router] Add a KV storage-tier coverage row to the Grafana dashboard (4/4) (#39111)


2026-09-15 17:09:11 -07:00
-
63845a1bb2
[router] Resolve a wire protocol per worker at registration (#39004)


2026-09-15 16:54:37 -07:00
-
0dabef3d30
[misc] Fix tool-call index, graph padded-row count, and prefill-graph input_embeds refresh (#39574)
Liangsheng Yin
2026-09-15 16:37:22 -07:00
-
fb91baedab
[Router] Real-GPU e2e coverage for storage-tier-aware cache routing (3/4) (#39110)


2026-09-15 15:48:51 -07:00
-
b803cfa0c4
Add external multimodal processors to the Rust frontend (#39329)
Lianmin Zheng
2026-09-15 15:38:47 -07:00
-
2929a39927
Use a shared byte budget for unified hybrid-SWA memory (#36729)


2026-09-15 15:27:00 -07:00
-
4da5599e93
Fix first-token metadata and reused attention-layer indexing (#39328)


2026-09-15 15:22:26 -07:00
-
08b1922405
fix(gemma4): set lm_head_is_tied for Gemma4UnifiedForConditionalGeneration (#35809)


2026-09-16 03:46:39 +05:30
-
87c9f78e99
[Fix][PD] Give prefill and decode their own RDMA NICs in disaggregation tests (#39584)
Liangsheng Yin
2026-09-15 15:07:42 -07:00
-
406c9c71d8
[Feature] Support BF16 and batch-invariant inference with DeepEP v2 (#38160)
Cheng Wan
2026-09-15 14:30:50 -07:00
-
d58342deab
[Fix] Release NCCL on scheduler exit and let the ASGI server own shutdown (#39560)
Liangsheng Yin
2026-09-15 13:39:15 -07:00
-
3c48c1e967
[Router] Expose the KV storage-tier stream and tree occupancy on /metrics (2/4) (#39109)


2026-09-15 13:30:05 -07:00
-
24874f90a3
[Test] Fix DeepGEMM batch invariance test output dtype (#39636)
YAMY
2026-09-15 12:40:28 -07:00
-
073fdf97cc
Generalize auxiliary outputs (#39164)

jasonjk-parkandjasonjk
2026-09-15 12:05:13 -07:00
-
49f540f3c0
[HiCache] Optimize buffer-mode storage existence bookkeeping (#39480)
Zhiqiang Xie
2026-09-15 10:49:58 -07:00
-
7f5dd19256
[HiCache] Rework the buffer-mode storage prefetch pipeline and retry bookkeeping (#39283)
Zhiqiang Xie
2026-09-15 10:48:49 -07:00
-
03ea13a545
[AMD][CI] Consolidate AMD workflows and retire ROCm 7.0 CI (#38632)
Bingxu Chen
2026-09-15 23:45:22 +08:00
-
47a157f257
[CI][Disaggregation] Fix EADDRINUSE flake in test_disaggregation_dwdp_gpt_oss (#39611)

Shangming CaiandClaude Opus 5
2026-09-15 23:05:05 +08:00
-
1aeeb25e86
[NPU][CI] Scope NPU nightly artifact dirs by image; fix stale glm5_2 case (#39585)
pllimax
2026-09-15 22:55:53 +08:00
-
832ec39cc0
[Intra-node PD][DSV4] Pack all layers into one batch for INTRA_NODE_NVLINK path (#38984)

Siyuan ChenandBJWang-ant
2026-09-15 21:08:33 +08:00
-
a2a261963f
[CI] Fix SWA decode radix cache NIXL import (#39600)
Po-Han Huang (NVIDIA)
2026-09-15 21:01:05 +08:00
-
1895cabfa8
Fix mooncake scale joiner groups (#38503)

Yoray ZackandPo-Han Huang
2026-09-15 14:07:25 +03:00
-
fb01a079a6
[NPU] [CI] Fix multimodal_gen filter leakage and add daily scheduled run (#39411)

Elizaveta Martirosianandronnie_zheng
2026-09-15 15:03:11 +04:00
-
17ba2c2e7c
Support non-strict GLM47 tool calls with EBNF constraints (#38890)

Yuxuan ZhangandXinyuan Tong
2026-09-15 16:41:48 +08:00
-
2c0a70960c
[qwen 3.8 next] reuse old cuda stream instead of endlessly creating streams (#39474)
Qiaolin Yu
2026-09-15 01:36:14 -07:00
-
8565b11003
[misc] Trim redundant variants from the 8-gpu-h20 disaggregation test suite (#39544)
Liangsheng Yin
2026-09-15 01:24:46 -07:00
-
4f52c9948b
keeping router GEMM in fp32 for deterministic inference (DeepSeek V3/V4) (#38176)


2026-09-15 04:16:09 -04:00
-
37ebacb50f
[EPLB] Drop defensive getattr for ep_dispatch_algorithm (#31804)

Beihao ZhouandBrayden Zhong
2026-09-15 01:14:41 -07:00
-
df254b0a11
[profiler] Label draft-runner steps DRAFT and target verify VERIFY in step spans (#38630)

Bi XueandLianmin Zheng
2026-09-15 00:57:05 -07:00
-
0e9b6bf8d8
[Router] Extract worker selection into policies::selection (no behavior change) (#39322)


2026-09-15 00:49:31 -07:00
-
e4a6b090f4
[CI] Run test_unified_radix_cache_kl_dcp on cutedsl_mla with bf16 KV cache (#39553)
Khoa Pham
2026-09-15 00:35:01 -07:00
-
8a20062652
[NPU][Bugfix] Disable pinned memory to fix DeepSeek-V2 DP-attention hang (#39423)
iridiumine
2026-09-15 14:59:31 +08:00
-
575163ff32
[NPU] [DOC] delete unsupported models in npu docs (#39555)
amote-i
2026-09-15 14:55:17 +08:00
-
06992c92ad
[Fix] Aggregate all IPC weight update responses (#39534)


2026-09-14 23:25:46 -07:00
-
860fa83a9f
[CI] Wait for a killed test server's GPU memory before the next launch (#39545)
Liangsheng Yin
2026-09-14 23:24:37 -07:00
-
e687b8d6af
[CPU] Implement fused QK Norm and RoPE kernels (#37748)
Cui Lily
2026-09-15 14:17:57 +08:00
-
2c37b90ad6
[Fix][DCP] Localize widened KV ids in MLA retraction CPU backup/restore (#39487)

Khoa PhamandClaude Fable 5.1
2026-09-14 22:33:41 -07:00
-
ddd4600197
[Fix] HiCache startup ImportError on the pinned kernel wheel (#39516)



2026-09-15 13:31:19 +08:00
-
5298d85218
fix: stop shadowing the DSpark shared-experts fusion guard (#39366)
Shuwen Wang
2026-09-15 13:24:49 +08:00
-
b510881157
[Fix][Qwen-VL] Normalize <image> sentinel on artifact fast path (#39278)

ashwini rathiandClaude Opus 4.7
2026-09-15 10:49:59 +05:30
-
6c514ab025
Force reasoning mode for GLM-5.3 chat templates (#39227)




2026-09-15 13:16:26 +08:00
-
ebd37705e4
[PD][LoRA] Gate decode admission on adapter slots (#39332)
Yanbin Jiang
2026-09-14 22:11:04 -07:00
-
3f871a246c
feat(agent sessions): attribute stored KV cache blocks to sessions (#37482)
ishandhanani
2026-09-14 21:34:49 -07:00
-
c9fbe5f655
[MM] Add flag to force Kimi image preprocessing onto CPU (#39148)

Byron HsuandByron Hsu
2026-09-14 21:19:39 -07:00
-
45b511b2ef
[Router] Honor KV-event storage tiers in the cache-aware tree (1/4) (#39108)


2026-09-14 21:18:15 -07:00
-
4b186cfea5
[CI][PD] Skip the flaky decode HiCache file-backend disaggregation test (#39368)
Shangming Cai
2026-09-15 11:19:03 +08:00
-
bdf8886ad3
[NPU] [DOC] Rename NPU hardware to Ascend A2/A3 Series product (#39389)
amote-i
2026-09-15 10:33:14 +08:00
-
5dde6e8f02
Fix MegaMoE buffer allocation and caching for effective SM budgets (#39223)

Aurick QiaoandAurick Qiao
2026-09-14 19:24:02 -07:00
-
99060191e7
[KDA] Support ReplaySSM ring-write in the fused chain-verify kernel (#36821)

Yuan Luoandluoyuan.luo
2026-09-14 19:22:41 -07:00
-
a23fd557ed
[Kernel] Add OOT dispatch for clamp position (#38687)

JoeandXiaoyu Zhang
2026-09-14 22:22:21 -04:00
-
e89d8facab
[sgl-router] Prepare dynamo-render dependencies (#39457)

Kan WuandClaude Fable 5.1
2026-09-14 19:08:16 -07:00
-
a25f213bc4
[diffusion] docs: give the RTX 5090 its own H3 recipe, measured on a physical desktop (#39373)


2026-09-15 09:15:13 +08:00
-
fc4193a63c
[HiCache] Remove duplicate benchmark result fields (#38097)
Charles xu
2026-09-15 08:56:26 +08:00
-
07e1918924
[Rust] Use Dynamo native renderers when chat templates are missing (#38939)
Kan Wu
2026-09-14 17:29:10 -07:00
-
8874c51a96
[Benchmark] Add an opt-out for the token-capacity check (#39284)
Xingyu Liu
2026-09-14 17:17:12 -07:00
-
c0b8725f5a
Fix /model_info serialization when a config value is a class (#39237)

raghothamandLianmin Zheng
2026-09-14 17:10:44 -07:00
-
276663a79d
[model-loader] Split weight loading from postprocessing (#34981)

Yang Liuandyangliu991
2026-09-14 17:10:19 -07:00
-
a2b4e8888f
Allow CUDA VMM feature transport with the Rust frontend (#39347)
Lianmin Zheng
2026-09-14 17:09:49 -07:00