This website requires JavaScript.
b20c375c10
Fix flaky decode cache-hit check in Inkling test (#34405 )
Ke Bao
2026-08-11 21:47:01 +08:00
f148eb6e6e
Add Hunyuan3 On Ascend Doc (#30223 )
Jinyan Yi
2026-08-11 21:20:30 +08:00
a50ab9cec7
[npu] [bugfix] Fix HiCache MHA backup for NPU (#34341 )
gjsheu
2026-08-11 21:09:45 +08:00
3add7e19ff
Add NVIDIA Nemotron 3.5 Lightning cookbook (#33481 )
Faradawn Yang and Ryan Stewart
2026-08-11 06:00:57 -07:00
2d193077f7
[JIT Kernel] Migrate per-token FP8 quantization from AOT to JIT (#34257 )
Mohammad Miadh Angkad
2026-08-11 20:40:40 +08:00
a0a76e4485
[DSV4] perf: Enable alt stream during BCG prefill (#29070 )
2026-08-11 14:10:19 +02:00
546965fc72
[diffusion] LTX-2: mount the bit-exact fused modulate at the 8 bare adaLN sites (ltx23-one-stage denoise -2.8% H100 / -2.6% H200) (#34315 )
Xiaoyu Zhang and Claude Fable 5
2026-08-11 18:23:08 +08:00
dd8c5849af
[diffusion] refactor: move dit execution capabilities to runtime models (#34249 )
Mick
2026-08-11 18:20:43 +08:00
071f0f1e9d
[diffusion] ERNIE-Image: fuse rotate-half RoPE + GELU-mul and hoist rope cos/sin (denoise -16.2% H100 / -12.7% H200, bit-exact) (#34306 )
Xiaoyu Zhang and Claude Fable 5
2026-08-11 18:18:37 +08:00
ba3dc16401
[diffusion] weight-only FP8: dequantize linear weights once at first use (Ideogram-4 denoise -18.8% H200 / -7.8% H100, bit-exact) (#34305 )
Xiaoyu Zhang and Claude Fable 5
2026-08-11 18:15:26 +08:00
5469faec45
HiSparse: shared-index (IndexShare) plan-then-IO swap-in prefetch (#34329 )
Zhiqiang Xie and Tingwei Huang
2026-08-11 01:58:28 -07:00
396722e490
[NPU] Disable failed test cases (#34377 )
2026-08-11 16:38:15 +08:00
6f3fe13a9c
[AMD] Install AITER's pinned Triton wheel in the ROCm 7.2 image (#34364 )
Bingxu Chen and YC Yen-Ching Tseng
2026-08-11 16:36:45 +08:00
d8a61c26a6
[CI] Add a scheduled workflow to close stale PRs (#34380 )
Liangsheng Yin
2026-08-11 00:54:02 -07:00
e74ea5b1d7
[ROCm/gfx95] Fix fp8 per-channel attention for Kimi-K2.7-code-mxfp4 o… (#31105 )
2026-08-11 15:51:57 +08:00
1c06c160f9
[Docs] Add Ling-3.0-flash INT4 and MXFP4 recipes (#34363 )
Xinyuan Tong
2026-08-11 15:35:27 +08:00
9d4be40124
Fix DSpark + DeepSeek V4 prefill CP compatibility (#33865 )
ybyang
2026-08-11 14:26:29 +08:00
afa2d5570b
[Diffusion] Add online FP8 support for Krea-2 (#34136 )
Tanmay patil
2026-08-11 11:42:05 +05:30
667e18d99d
[PD] Support pipeline-parallel prefill with Mooncake staging buffer (#33807 )
YAMY and Shangming Cai
2026-08-10 22:53:17 -07:00
13aeb91b6e
[Fix] Update multimodal CUDA VMM helper import (#34358 )
Xinyuan Tong
2026-08-11 13:41:21 +08:00
b3c02cbce7
[perf] Collapse the DP attention scheduler sync to a single D2H copy (#34338 )
Liangsheng Yin and Alex Nails
2026-08-10 21:52:29 -07:00
585c3c6816
[Refactor] Split the FlashInfer autotune dummy-run flag from the LM-head policy (#34336 )
Liangsheng Yin
2026-08-10 21:20:05 -07:00
a58fa0388e
[Fix] Correct W4AFP8 DeepEP scaling and mode-specific dtypes (#33669 )
EchO
2026-08-11 11:21:42 +08:00
704808ed27
[MiniMax-H3] SubBlock: training-free block-sparse attention for the DiT (#34148 )
forrestl and Claude Opus 5
2026-08-11 11:11:11 +08:00
dd20826e0a
[AMD] Preserve the AITER expert mask across torch_memory_saver pause/resume (#34220 )
Zhiyao Jiang
2026-08-10 22:59:15 -04:00
b498f46271
[CI] Add MegaMoE runner compatibility alias (#34326 )
Baizhou Zhang
2026-08-10 19:47:38 -07:00
8d050dd880
[AMD][CI] Run MI300 8-GPU stage-C shards two at a time (#34324 )
Michael
2026-08-10 18:49:08 -07:00
5af5183351
test: isolate metal profiler tests from ambient SGLANG_USE_MLX (#34300 )
Jae B.
2026-08-10 21:44:16 -04:00
aeab1de1de
[diffusion] optimization: support cuda graph for Pi-0.5 prefix encoding (#34256 )
Mick
2026-08-11 09:14:53 +08:00
df986c4d5e
Consolidate CUDA VMM allocation helpers (#34199 )
cctry
2026-08-10 18:11:11 -07:00
ba5183fe10
[diffusion] UX: fix CI server warmup progress logging (#34301 )
Mick
2026-08-11 09:09:32 +08:00
418975ba64
[EPD] feat: pipeline owner-only multimodal preprocessing (#34206 )
Mick
2026-08-11 09:08:05 +08:00
03c942dfab
[Fix] Make wait_port_available actually wait timeout_s seconds (#33086 )
elvischenv and Claude Opus 5
2026-08-11 08:35:50 +08:00
a92bbf2f24
docs: remove DSV4 low-latency chunked prefill size (#34333 )
Baizhou Zhang
2026-08-10 17:23:43 -07:00
d59c1ddf70
fix(dflash): account for DCP in draft KV pool sizing (#33912 )
milesial
2026-08-10 17:15:01 -07:00
0967885121
[DCP] Drop two per-layer launches from the MLA target-verify path (#34240 )
Khoa Pham and Claude Opus 5
2026-08-10 17:11:03 -07:00
2d5009d130
[DSV4] Avoid host syncs in EAGLE prefill (#33662 )
weireweire and weireweire
2026-08-11 08:09:55 +08:00
7c7326ccb3
Fix DeepSeek-V4/DeepSeek-V4-Pro DP-attention gather semantics (#31700 )
Michael Gschwind
2026-08-10 17:03:19 -07:00
56e8bb49b8
fix(ci): constrain NeMo Skills evaluator dependencies (#34322 )
Baizhou Zhang
2026-08-10 16:56:46 -07:00
c30872fa00
[Fix] Read trace level via get_global_trace_level() in trace_async (#34325 )
Liangsheng Yin
2026-08-10 16:46:54 -07:00
8c5d5f75bf
Fix DSV4 DSpark shared expert loading (#33312 )
Mohammad Miadh Angkad
2026-08-11 07:36:50 +08:00
77b8315b84
[CI] Fix GSM8K floating-point tolerance boundary (#34272 )
Mohammad Miadh Angkad
2026-08-11 07:27:11 +08:00
ceeaec2078
[PD] Support --enable-unified-memory with PD disaggregation (kimi-linear MLA hybrid-Mamba) (#33362 )
Cheng Wan and Claude Fable 5
2026-08-10 16:07:58 -07:00
8a7c8a72d6
Fix NaN logits from deterministic Triton extend on the unified memory pool (#33517 )
Cheng Wan and Claude Fable 5
2026-08-10 16:05:04 -07:00
0661eb1c50
plugins: don't directly set quant class (#34252 )
2026-08-10 16:01:31 -07:00
b86c90215f
[Docs] Muse Glimmer cookbook: drop --speculative-dflash-block-size 5 (#34323 )
Brayden Zhong and Brayden Zhong
2026-08-10 15:46:34 -07:00
2c72323d90
[AMD] Fix AITER custom reduce-scatter CUDA-graph capture crash under torch_memory_saver (#34203 )
Zhiyao Jiang
2026-08-10 18:36:15 -04:00
c80a38edcd
[Fix] Pin cuda-tile to 1.6.0rc5 to unblock Python 3.10 x86_64 installs (#34321 )
Liangsheng Yin
2026-08-10 15:09:47 -07:00
ca0f8a0f4c
perf(hisparse): fuse the DSv4 value and scale swap-in copy on ROCm (#33484 )
AMD-yanfeiwang
2026-08-11 05:30:15 +08:00
166c6f7181
[Kernel] cutedsl_bf16_gemm: trailing cluster barrier for 2-CTA TGV kernel exit (#32907 ) (#32954 )
QIN2DIM
2026-08-11 04:31:18 +08:00
1a8e4876b6
perf(hisparse): 128-bit non-temporal swap-in copy on ROCm (#33085 )
AMD-yanfeiwang
2026-08-11 04:27:15 +08:00
733c05c887
[spec decoding] support inkling dspark (#31847 )
Qiaolin Yu
2026-08-10 13:21:35 -07:00
e54c153ba6
Add Intern-S2-Mobius cookbook (#33820 )
2026-08-11 04:13:35 +08:00
a2161ce682
Support thinking budget for Inkling (#33146 )
Lifan Shen and Xinyuan Tong
2026-08-10 11:01:08 -07:00
955aab8db1
[DCP] Reuse partial output in natural-log LSE merge (#34213 )
Tan Trinh
2026-08-11 00:45:40 +07:00
7738062294
[unified memory] Support DSPARK speculative decoding + fix two NaN root causes (page hand-out zeroing, CuTe int32 slot-stride wrap) (#33974 )
Cheng Wan
2026-08-10 10:35:06 -07:00
ec9babe36c
[diffusion] fix: fix h3 rank-local fsdp qkv loading (#34294 )
Mick
2026-08-10 23:59:00 +08:00
d07ac32d05
[diffusion] feat: support --served-model-name in sglang serve (#34228 )
2026-08-10 22:27:09 +08:00
f5f0c3ee7a
[diffusion] Z-Image single-GPU BCG: fix the replay crash and make output bit-exact vs eager (#34183 ) (#34210 )
Xiaoyu Zhang and Claude Fable 5
2026-08-10 22:18:32 +08:00
fd3036523a
[diffusion] Clean up shared bitexact gates, helpers, and stale naming (#34180 )
2026-08-10 22:14:09 +08:00
77c90e7e54
Cookbook: add Ling-3.0-tiny (#34283 )
Xinyuan Tong
2026-08-10 21:25:17 +08:00
3e2a26708b
[diffusion] chore: expose architecture config at the dit runtime boundary (#34248 )
Mick
2026-08-10 20:08:45 +08:00
8ba9385097
[diffusion] chore: optimize model weight loading (#34064 )
Mick
2026-08-10 20:07:48 +08:00
4eaaeda004
[CI] Build patched Docker images for both amd64 and arm64 (#34276 )
Liangsheng Yin
2026-08-10 04:49:22 -07:00
01a2ef95b8
[NPU] Modified kernel tag version to 8.10 (#34254 )
黄孝君
2026-08-10 19:40:52 +08:00
d95e824a49
Muse Glimmer Cookbook: install from the PR branch (#34281 )
Brayden Zhong and Brayden Zhong
2026-08-10 04:39:55 -07:00
14ffd447a4
[FEAT] Decouple multimodal global cache from Mooncake (#30392 )
2026-08-10 19:23:19 +08:00
d96b1533ea
Muse Glimmer Cookbook: install from the PR branch (#34278 )
Brayden Zhong and Brayden Zhong
2026-08-10 03:49:02 -07:00
443b62db57
fix(vlm): stream-order cuda-ipc feature pool lifecycle and streamline multimodal transport module (#33949 )
Mick
2026-08-10 18:47:19 +08:00
0977b22431
[AMD] Restore K3 MLA verify kernel path blocked by can_handle() guard (#34261 )
Thomas Wang
2026-08-10 18:31:53 +08:00
a6c34df044
Muse Glimmer Cookbook (#34271 )
2026-08-10 03:21:13 -07:00
955569a2dc
[diffusion] feat: expose cosmos3 policies through the Action API (#34243 )
Mick
2026-08-10 18:16:20 +08:00
c971d7ac9c
Refactor staging registration metadata fields (#33910 )
YAMY and Shangming Cai
2026-08-10 02:53:32 -07:00
0b6189d0e8
[CI] Add output_tag input to the Patch Docker Image workflow (#34253 )
Liangsheng Yin
2026-08-10 02:48:03 -07:00
3c533acec6
[Hicache][2/2]Support Mamba branching in Unified Radix Cache with HiCache (#33639 )
Jincong Chen
2026-08-10 17:16:42 +08:00
430f38ea25
Update dspark draft path in Inkling small cookbook (#34250 )
Ke Bao
2026-08-10 17:12:51 +08:00
a76a167812
Fix/hisparse host backed max request length (#28753 )
2026-08-10 16:58:13 +08:00
b51bf9ec9e
[Spec] Budget the DFLASH draft KV pool from its own attention geometry (#34234 )
Liangsheng Yin and Brayden Zhong
2026-08-10 01:28:49 -07:00
3bb72bc72a
[CI] Keep the torch compilation cache instead of wiping it on install (#34231 )
Liangsheng Yin
2026-08-10 01:21:20 -07:00
f2a4c4c847
[AMD] [CI] Enable 3 nested unit tests needing harness stub fixes (#31843 )
Michael and HAI
2026-08-10 01:09:00 -07:00
fb3d1419fd
[tracing] sglang tracing v2: support exporting tracing data asynchronously (#30023 )
Feng Su
2026-08-10 15:23:05 +08:00
06f32bab6b
[BCG][5/N] MLA Fully Support (#33661 )
Yuwei An and Claude Opus 5
2026-08-09 23:52:14 -07:00
aea78d1e73
[misc] Pass FP8 scales in FlashInfer SWA prefill, autotune fp8 on SM120, and tighten is_image_understandable_model (#34217 )
Liangsheng Yin and Brayden Zhong
2026-08-09 23:45:11 -07:00
d6a066131c
[Feature] Support NVFP4 token embedding in ModelOpt mixed-precision checkpoints (#34222 )
Liangsheng Yin and Brayden Zhong
2026-08-09 23:37:54 -07:00
e226bb711c
enable TRT-LLM for MiniMax M3 by preserving SwiGLU params (#33962 )
Chunan Zeng
2026-08-09 23:30:05 -07:00
2969ab3d41
[MLX] Window-bounded SWA KV storage and in-graph sampling (#34166 )
2026-08-09 21:28:51 -07:00
449f0da78f
[Intel GPU] DeepSeek V4 15/N: Add silu_and_mul_clamp support to triton fused_moe for XPU (#33808 )
Polisetty V R K Jyothendra Varma
2026-08-10 09:52:05 +05:30
5a8e360e70
[PD] Skip speculative verify scratch on prefill servers (saves num_draft_tokens x mamba pool per rank) (#34191 )
YAMY
2026-08-09 20:56:00 -07:00
5d85f25f75
fix(minimax): use routed TRT-LLM for NVFP4 MoE auto on SM100 (#32229 )
Jan Bernlöhr and Mohammad Miadh Angkad
2026-08-10 05:45:14 +02:00
410088c91e
[NPU] Increase the retry count to 3 for the GSM8K. (#34103 )
2026-08-10 10:55:41 +08:00
7331287c1c
[JIT Kernel][DSv4] Optimize epilogue of c128 (#26671 )
2026-08-10 10:32:43 +08:00
accc51c6db
[DSA] Fix top-k v2 dropping non-primary ranks' output on CUDA 13.1+ (root cause for #33835 ) (#34167 )
DarkSharpness and Claude Opus 5
2026-08-10 10:31:47 +08:00
169783d42f
[diffusion] chore: make torch.compile opt-in for speed mode (#34173 )
Mick
2026-08-10 10:22:16 +08:00
ee3ee8393c
fix: preserve GQA head mapping in Triton DCP prefill (#34161 )
Khoa Pham
2026-08-09 18:52:18 -07:00
441910f926
[diffusion] LTX-2 quality=high fused RMSNorm+modulate + FFN GELU epilogue (H200 ltx23-one-stage denoise 45.85->43.24 s, ~matches torch.compile) (#34172 )
Xiaoyu Zhang and Claude Opus 4.8
2026-08-10 09:46:17 +08:00
c20e99bd22
fix(vlm): preserve Kimi-K3 GPU JPEG accuracy (#34163 )
Mick
2026-08-10 09:42:52 +08:00
553dc0f936
[MLX] Support gpt-oss: sliding-window attention, attention sinks, sm_scale (#30050 )
2026-08-10 09:32:36 +08:00
68b961e9fb
Add kda replayssm tests (#33630 )
Ke Bao
2026-08-10 09:11:20 +08:00
25b7015064
Fix stale track rows corrupting conv checkpoints under the prefill graph (#34184 )
Ke Bao
2026-08-10 08:57:23 +08:00
56ef810cad
[diffusion] BCG: auto-capture the default warmup resolution instead of hard-requiring --warmup-resolutions (H200 SANA denoise 0.73->0.457 s with a single flag) (#34174 )
Xiaoyu Zhang and Claude Opus 4.8
2026-08-10 08:37:32 +08:00