Commit Graph

Select branches

Hide Pull Requests

dsv41-pd

dsv41-pd-visioncp

main

Mono Color

• 7562e741e2 Retain SWA down to the last state checkpoint (#34729) Ke Bao 2026-08-15 01:12:38 +08:00
• c20aceeb88 Fix mamba checkpoint depth under dcp (#34808) Ke Bao 2026-08-15 00:34:00 +08:00
• 70e291b70f [Docs] Add GB300 cells and benchmarks for Qwen3.8-27B (#34863) 4 people 2026-08-14 08:42:50 -07:00
• 29c6be15a4 [Docs] Add Qwen3.8-27B cookbook page (#34860) 4 people 2026-08-14 08:23:39 -07:00
• 9c9a3273be [diffusion] Fix Helios denoising profiler stepping (#34826) Xiaoyu Zhang 2026-08-14 23:21:54 +08:00
• 5f2a6d6422 [diffusion] Fix symbolic replicated-mode counting under torch.compile (#34824) Xiaoyu Zhang 2026-08-14 23:20:29 +08:00
• b95a746948 [MoE] Fuse swiglu moe up gemm epilogue (#32944) Yuan Luoandluoyuan.luo 2026-08-14 22:11:34 +08:00
• 9d2f1584fa Fix MiniMax-H3 Cache-DiT BCG warning (#34848) Mohammad Miadh Angkad 2026-08-14 22:07:50 +08:00
• c939307e8a [MiniMax-M3] Overlap shared and routed experts (#34542) 3 people 2026-08-14 22:01:48 +08:00
• 1a178f7c7c [EPD] Batch embedding cache host-device range copies (#31574) 4 people 2026-08-14 22:01:12 +08:00
• 5e65dd01a7 Remove the torchao integration (--torchao-config) (#34304) Brayden ZhongandBrayden Zhong 2026-08-14 06:49:11 -07:00
• fe0c18effd [NPU] [DOC] Add Qwen3.8-Max deployment tutorial on Ascend NPUs (#34836) amote-i 2026-08-14 20:02:04 +08:00
• 18107e38d2 Skip oow slot freeing under eagle (#34823) Ke Bao 2026-08-14 17:13:21 +08:00
• 4d94f1d310 [diffusion] fix: warn when bcg disables cache-dit (#34242) chilltongxandchilltongx 2026-08-14 16:45:36 +08:00
• f2c84de022 [Perf] Publish the WAR read-done event at DSPARK verify (#34816) Liangsheng Yin 2026-08-14 00:46:52 -07:00
• a86edcdc0a [diffusion] feat: rebuild minimax-h3 adaln outputs on demand (#34650) triple-muandMick 2026-08-14 15:33:51 +08:00
• 2622e013eb [Fix] has_hf_quant_config crashes on local dirs without the config (#34774) 3 people 2026-08-14 03:17:42 -04:00
• 65d62109dd [AMD] Fix Triton 3.7 gfx950 extend-attention spills (#34741) kk 2026-08-14 15:05:42 +08:00
• 7c15b9b7d0 [diffusion] fix: fix cache-first fast path accepting a metadata-only snapshot (#34121) Dayananda V 2026-08-14 12:22:52 +05:30
• 827552bc1d Fix eager AMX backend probe imports (#34496) datdo-msftandXiaoyu Zhang 2026-08-13 23:44:14 -07:00
• ba1d980b35 [AMD] Accelerate AITER unified-attention decode with scaled FP8 Q (#31856) 3 people 2026-08-14 02:42:25 -04:00
• 46d84f4b48 feat(cli): add extensible serve backend plugins (#34753) Mick 2026-08-14 13:57:59 +08:00
• 463981922c [Cookbook] Add DeepSeek-V4-Pro-0813 (Pro Official) serving recipes (#34809) zijiexiaandClaude Opus 5 2026-08-13 22:53:41 -07:00
• b784726863 [kernel] Content-addressed JIT build cache, generated from our own ninja (#34274) 3 people 2026-08-14 13:41:04 +08:00
• 240a12b302 [AMD] Fuse shared_expert_gate GEMV into the MoE append kernel (HIP/aiter) (#28666) jacky.chengandsogalin_codegen 2026-08-14 12:32:27 +08:00
• 0a6bbbe128 [Fix] Restore layer-level DSV4 RoPE policy (#34788) Baizhou Zhang 2026-08-13 21:26:45 -07:00
• e1c4db9621 [quantization] Add tuned Triton tile configs for channelwise FP8 GEMM… (#34331) 3 people 2026-08-13 20:11:02 -07:00
• 85cdf1178d [CI] Prune redundant CPU test overhead (#34309) Xinyuan Tongandhnyls2002 2026-08-14 10:51:17 +08:00
• 6ad3f2d8fd docs: link dots3.note checkpoints, add H100 cells (#34797) Yuhao YangandClaude 2026-08-14 10:42:27 +08:00
• 456c5551cc [XPU][test] Add cache_salt=None to _make_req in test_lmcache_radix_cache.py (#34728) ashwini rathiandarathi-hlab 2026-08-14 07:57:15 +05:30
• f2b2b567aa perf(jit_kernel/deepseek_v4): optimize paged_mqa_metadata (#25855) SII-yangdianandSII-yangdian 2026-08-14 10:00:22 +08:00
• 704e512836 [GDN] Honor configured linear-attn verify backend in the kernel dispatcher (#34592) Xiaoyu ZhangandClaude Fable 5 2026-08-14 08:59:52 +08:00
• 34219ed9a7 [AMD] CI: pin antlr4-python3-runtime back after lmms-eval (unblocks ROCm 7.2 stage-b evals) (#34768) Michael 2026-08-13 17:41:16 -07:00
• 9d34c2809f [FlashInfer v0.6.16] Support FlashInfer CuTe DSL NVFP4 MoE quantization (#28354) Ziang Li 2026-08-13 17:33:46 -07:00
• c4271c3fe1 [DSpark] Fix EP1 decode performance regression (#34759) ormandj 2026-08-13 19:00:54 -05:00
• 54c44fef73 [Perf] Skip trivial DSV4 nonpaged indexer logits (#33857) weireweireandweireweire 2026-08-14 07:21:30 +08:00
• 151a314829 [Fix] Make the DSpark draft num_token_non_padded host-to-device copy non-blocking (#34782) Liangsheng Yin 2026-08-13 16:10:54 -07:00
• abdef3c38e [Qwen] Update Docker image tag for MI300X to v0.5.17-rocm700-mi30x-20… (#34770) zijiexia 2026-08-13 15:59:42 -07:00
• 9720922671 [AMD][CI] Restore gfx942 Grok-1 INT4 and Grok-2 schedules (#34761) MichaelandMichael 2026-08-13 15:57:54 -07:00
• f3beb2c529 [CI] Disable the prefill CUDA graph on the P worker of test_kimi_linear_pd_dcp4 (#34779) Khoa PhamandClaude Opus 5 2026-08-13 15:26:21 -07:00
• 81fe452810 [DCP] Share one pack kernel between both a2a backends (#34651) Khoa PhamandClaude Opus 5 2026-08-13 15:06:02 -07:00
• 8bbca87780 [Core] Organize environment variable registry (#34730) Lianmin Zheng 2026-08-13 14:25:37 -07:00
• 652a2709d1 [Docs] Add decode context parallelism to advanced features (#34654) Khoa PhamandCursor 2026-08-13 13:51:57 -07:00
• 903439044a perf(kv-events): coalesce cache events (#31479) jthomson04 2026-08-13 13:36:53 -07:00
• 8554d9a5bc [Fix] Carry the backend on Kimi-K3 deferred preprocessing configs (#34766) Liangsheng Yin 2026-08-13 13:30:33 -07:00
• 8ad04a9bee docs(nixl): document OBJ throughput target (#30405) Lukas Humbel 2026-08-13 22:04:13 +02:00
• 29b067245b [AMD] CI: drop the spaces from SGL_EVAL_SPEC (fixes ROCm 7.2 stage-a sgl-eval install) (#34689) kangwangamdandBingxu Chen 2026-08-14 03:48:58 +08:00
• 96db53ec70 [CI] Fix test_resolution_is_reproducible after cuda_ipc became opt-in (#34746) Mohammad Miadh Angkad 2026-08-14 03:47:03 +08:00
• 6b94d39f13 [Model Loading] Overlap checkpoint staging with CUDA graph capture during startup (#32017) 3 people 2026-08-13 12:26:25 -07:00
• 0772e79ee7 [CI][PD] Pin nccl rendezvous port per side to fix flaky disaggregation tests (#34755) Shangming Cai 2026-08-14 01:44:20 +08:00
• a82f8e1777 [PD] Add the missing Prefill bootstrap timeout for NIXL (#34692) jambow0320 2026-08-14 01:08:09 +08:00
• 69a31ce342 fix: make Cache-DiT actually cache on MiniMax-H3 (#33827) 4 people 2026-08-14 00:39:58 +08:00
• c255fbc4fe [Diffusion] Add @triple-mu as a code owner (#34748) Xiaoyu Zhang 2026-08-14 00:37:49 +08:00
• ea7a6e0e99 fix(diffusion): unshard FSDP root group for custom encoder entry points (#34575) triple-mu 2026-08-13 23:24:00 +08:00
• ebca0bbde4 [Diffusion][ERNIE] Fuse QKNorm with full-width RoPE (#34620) Xiaoyu Zhang 2026-08-13 23:23:21 +08:00
• 82f7afb881 [Diffusion] Make auto residency decisions component-scoped (#34615) Xiaoyu Zhang 2026-08-13 23:20:55 +08:00
• 07821e9d56 [AMD][CI] Stop scheduling Grok-1 and Grok-2 on MI30x (#34643) 3 people 2026-08-13 07:22:42 -07:00
• c1142677a8 [do not merge] add new cookbooks (#34658) 3 people 2026-08-13 21:35:15 +08:00
• 74c0322342 [Diffusion][FLUX.2] Fuse eager AdaLN and packed SwiGLU (#34616) Xiaoyu Zhang 2026-08-13 19:55:53 +08:00
• 3c1791a7df [Diffusion][HunyuanVideo] Fuse eager QKV packing and high-quality QKNorm (#34617) Xiaoyu Zhang 2026-08-13 19:54:00 +08:00
• 993e24df75 [diffusion] Add --dit-layerwise-residency-policy for strided DiT residency (#34534) triple-muandXiaoyu Zhang 2026-08-13 19:49:56 +08:00
• b764194e81 [diffusion] model: support LongCat-Image (#23274) Liu ZhenlongandXiaoyu Zhang 2026-08-13 19:48:26 +08:00
• eea2e5d6e5 Optimize delayed sample and mrope position computation (#32637) Estrella-xx 2026-08-13 19:15:47 +08:00
• 9c0d4cba3f [CI] Split Kimi K2.5 performance batches by config (#34669) Baizhou Zhang 2026-08-13 03:20:50 -07:00
• a23670ddbf [diffusion] Wan2.2-TI2V: fuse per-token adaLN table add into contiguous slices + hoist rope cache (denoise -13.1% H100 / -12.6% H200, bit-exact; eager beats compile) (#34584) Xiaoyu ZhangandClaude Fable 5 2026-08-13 17:26:46 +08:00
• ba23846ccf move the PD bootstrap registry under api_server::disaggregation (#33895) Rain Jiang 2026-08-13 02:13:43 -07:00
• 34206c0017 [AMD] Run V4 MTP target-verify through the decode kernel (#34597) 4 people 2026-08-13 17:00:45 +08:00
• fd1e04d952 refactor error responses into shared utils::response helpers (#33894) Rain Jiang 2026-08-13 01:30:29 -07:00
• dbebc1deb4 [Perf] Occupancy tuning for DSA indexer fp8-quant Q kernel (#32755) Rainchar9119 2026-08-13 16:15:27 +08:00
• fad376d3ee [CPU][QUANT] add amx cpu support for auto-round (#29593) 4 people 2026-08-13 15:50:59 +08:00
• c034120cb8 [AMD] Enable draft-extend CUDA graph and reduce bubble for MTP (#29202) Xinyi SongandThomas Wang 2026-08-13 00:50:19 -07:00
• a34f81251f fix(hicache/umbp): support DeepSeek-V4 hybrid HostPoolGroup (multi-po… (#30762) AMD-yanfeiwangandZhangheng 2026-08-13 15:40:07 +08:00
• b7f87a2513 [AMD][Perf] Fuse GatedDeltaNet QKVZBA split/reshape/cat into a single Triton kernel for Qwen3.5-architecture MoE on HIP (#34421) jacky.cheng 2026-08-13 15:18:35 +08:00
• 5a5c3d309b Drop the mmlu case from the unified radix cache kit (#34667) Ke Bao 2026-08-13 15:16:05 +08:00
• 969921b32d [diffusion] chore: track minimax-h3 in the nightly diffusion benchmark (#34655) MickandClaude Opus 5 2026-08-13 15:12:26 +08:00
• 3f6ef01322 Optimize MiniMax-M2.7 on CPU (#31956) Xinguo Zhu 2026-08-13 15:04:16 +08:00
• 889c2f31aa feat(attention): add architecture-owned SM12x FA4 kernels (#32991) Gabriel Wu 2026-08-13 14:58:42 +08:00
• ebc144ce3f Enable unified cache out-of-window slot freeing by default (#34653) Ke Bao 2026-08-13 14:31:42 +08:00
• e8c7dddfa0 [VLM] add content-addressed preprocessing cache infrastructure (#34398) Mick 2026-08-13 14:09:07 +08:00
• 035c622a14 Reenable breakable CUDA graph for NemotronH (#34538) elvischenv 2026-08-13 13:58:10 +08:00
• ef7208d41d [kernel] add triton moe TMA up support (#33559) 5 people 2026-08-13 13:29:32 +08:00
• 69bf601e3c fix: restore VLM nightly regression coverage (#34662) Mick 2026-08-13 12:47:16 +08:00
• 1286a50bb9 Add New Intel members into CI permission list (#34671) Zaili Wang 2026-08-13 12:28:00 +08:00
• 8761b971f1 [CI] Fix nightly test failures (#34637) Liangsheng Yin 2026-08-12 20:53:44 -07:00
• 26627e999d [Fix] Snapshot req.prefix_indices when the prefix cache is disabled (#34644) Liangsheng Yin 2026-08-12 20:50:46 -07:00
• a318e16956 [diffusion] feat: publish an index of nightly comparison runs (#34652) MickandClaude Opus 5 2026-08-13 11:50:10 +08:00
• bca8ed4afc [minimax m3][npu]Adaptation of Minimax M3(w8a8) for NPU platforms [1/2] (#32941) +7686 people 2026-08-13 11:23:14 +08:00
• d44c836cfd [XPU][CI] disable SYCL_CACHE_PERSISTENT to fix topk segfault (#34567) ashwini rathiandClaude Opus 4.7 2026-08-13 08:49:39 +05:30
• 5efe9b104d Record both architectures in the bit-exact guard docstrings (#34656) Ke Bao 2026-08-13 10:55:35 +08:00
• 26579d893e [Diffusion][GLM-Image] Retune QK head LayerNorm for SM103 (#34619) Xiaoyu Zhang 2026-08-13 10:49:13 +08:00
• 50cc1aa241 [CI] Route mmlu and GB300 MMMU-Pro evals through sgl-eval (#34477) Liangsheng Yin 2026-08-12 19:33:21 -07:00
• cfdfd31826 Add thanhhao98 to CI_PERMISSIONS.json (#34649) Khoa PhamandClaude Opus 5 2026-08-12 19:02:03 -07:00
• 6a5a9eccaa add flashinfer cute-dsl backend for mxfp8 gemm (#34042) Carrie ChenandBrayden Zhong 2026-08-12 20:50:01 -04:00
• 40eaf34428 fix: make automatic NUMA binding configurable (#30394) Lukas Humbel 2026-08-13 02:40:28 +02:00
• aefe2d0207 Revert "[Kimi K3] Fuse MLA gate projection into QKV-A GEMM" (#34642) Mohammad Miadh Angkad 2026-08-13 08:35:38 +08:00
• bbda7f32b1 [AMD][CI] CI: fix AMD 2-GPU multimodal-gen partition-count abort (#34328) 3 people 2026-08-12 17:32:33 -07:00
• 273d978bed [CI] Use default installer for B300 tests (#34635) Baizhou Zhang 2026-08-12 16:21:40 -07:00
• 385903b0ac feat: add cache salt support to KV cache events (#30827) jthomson04 2026-08-12 16:14:04 -07:00
• 198b7e9240 [Docs] Use Meta's canonical Muse Glimmer GGUF filename (#34626) Jimmy Shong 2026-08-12 16:07:55 -07:00
• 6c6294b7be [DCP] Fuse the a2a pack/unpack copies in the MLA LSE reduce (#34614) Khoa PhamandClaude Opus 5 2026-08-12 16:04:46 -07:00