Commit Graph

Select branches

Hide Pull Requests

dsv41-pd

dsv41-pd-visioncp

main

Mono Color

• 6275114548 [NPU] [DOC] Update model names supported on Ascend NPU (#31316) amote-i 2026-07-16 14:45:59 +08:00
• e73f323464 [JIT] Reduce MoE fused gate CI test sweep (#31400) Xiaoyu Zhang 2026-07-16 14:42:45 +08:00
• 8c5e0cee18 [AMD] Bump MoRI to f7e6ac6 to fix ROCm install_dependency build break (#31406) YC Yen-Ching Tseng 2026-07-16 14:40:18 +08:00
• 0a64139c94 Fix --moe-a2a-backend silently ignored for LongCat-2.0 (moe_topk missing from gate) (#30975) 3 people 2026-07-16 14:40:01 +08:00
• 5af65d8542 [DeepSeek-V4] Support BF16 Compress State for Online C128 (#29609) Ryan Zzzandzhujunyu 2026-07-16 14:17:29 +08:00
• dc60f65661 chore: bump tokenspeed_mla to 0.1.8 (#31385) Khoa PhamandClaude Fable 5 2026-07-15 22:37:54 -07:00
• ac4fa6496c Skip no-op EAGLE sampling renormalization (#31294) weireweireandweireweire 2026-07-16 13:37:10 +08:00
• 40517b593b [docs] Inkling cookbook: LoRA cells require --disable-prefill-cuda-graph (#31418) Yanbin Jiang 2026-07-15 22:19:49 -07:00
• 0b372d03de [CI] Guard partition consumers against failed check-changes and degenerate fits (#31416) Liangsheng Yin 2026-07-15 22:09:40 -07:00
• 238448f4c8 ci: fix runner utilization report undercounting busy time ~25x (#31396) Alison Shao 2026-07-15 22:02:38 -07:00
• b871a509e6 [Fix] Wire the detokenizer soft watchdog into the multi-http-worker event loop (#31392) Liangsheng Yin 2026-07-15 21:55:46 -07:00
• bc525dcf90 [Cookbook][CPU]Update CPU model support info in Cookbook (#30520) Zaili Wangandzijiexia 2026-07-16 11:42:09 +08:00
• dee91c51cf perf(deepseek_v4): enable SGLANG_OPT_FP8_WO_A_GEMM on sm90 (Hopper) (#28983) guzekai01 2026-07-16 11:16:32 +08:00
• 7f9a902cd9 fix(humming): handle missing quant_method (#31185) guzekai01 2026-07-16 11:15:52 +08:00
• 22453ca63c docker: build HPC-Ops into the GPU image (#31390) Xiaoyu ZhangandClaude Fable 5 2026-07-16 11:07:01 +08:00
• f28ce5c420 [XPU]REPO cache dtype xpu align with cuda (#31140) Swift.Sun 2026-07-16 10:30:01 +08:00
• 3a8ddd3fd0 [XPU] Route topk_sigmoid and topk_softmax to AOT sgl-kernel-xpu symbols (#31038) 3 people 2026-07-16 07:59:19 +05:30
• 14095ef78a [AMD] Disable CUDA IPC multimodal transport on ROCm in MMMU VLM tests (#31342) YC Yen-Ching Tseng 2026-07-16 09:42:30 +08:00
• d9003dd452 fix: skip unsafe automatic prefill graph capture (#31204) Mick 2026-07-16 09:27:38 +08:00
• 1c4892d7bb [Mamba] Fix spec-v2 + extra_buffer crash (guard None mamba_next_track_idx) (#27998) Serge Panev 2026-07-15 18:23:14 -07:00
• 7647a9d260 Fuse the preprocess kernels of trtllm-gen attention (#29690) Brayden ZhongandBrayden Zhong 2026-07-15 18:21:34 -07:00
• 871c648203 [NPU]revert add scoring func for GLM 4.7 Flash (#31388) McZyWu 2026-07-16 09:16:57 +08:00
• edb2059139 Support Flashinfer one-sided A2A + CuteDSL MoE for Nemotron Ultra (#28309) Brayden ZhongandBrayden Zhong 2026-07-15 18:14:20 -07:00
• ac23be8d09 Skip MXFP8 autotune on dense GEMM, which causes IMA (#29669) Brayden ZhongandBrayden Zhong 2026-07-15 18:06:55 -07:00
• 0b04e9da83 Use fused A GEMM for fc1_latent_proj in NemotronH (#29692) Brayden ZhongandBrayden Zhong 2026-07-15 18:05:22 -07:00
• 1cc9493747 [Spec] Converge DP-attention spec width scaling onto num_tokens_per_req (#31244) Liangsheng Yin 2026-07-15 17:57:08 -07:00
• 5d004a20c5 Fix FlashInfer A2A top-k ID dtype (#29929) Po-Han Huang (NVIDIA) 2026-07-16 08:56:11 +08:00
• 34f5691ea1 docs: sync LMSYS SGLang blog cards (#31386) sglang-botandsglang-bot 2026-07-15 17:53:19 -07:00
• b0b2dfbda1 [Spec] Extract the shared draft() tail into build_eagle_verify_input (#31375) Liangsheng Yin 2026-07-15 15:59:31 -07:00
• 7a973c03a0 [Bugfix] Stamp capture-time num_tokens_per_req in multi-layer EAGLE; close jit_kernel CI filter gaps (#31367) 3 people 2026-07-15 15:24:32 -07:00
• 3101c1258c [DSv4] Use BF16 instead of FP32 for indexer score computation (#30012) Lewisand百麒 2026-07-16 06:06:55 +08:00
• 26cb0fcdda Empty _REQ_TYPES_WITH_OPAQUE_FIELDS on the msgpack IPC path (#29465 Task 4) (#30182) Jorge António 2026-07-15 22:55:06 +01:00
• 67148447a6 [AMD] Register 3 CPU-bound / triton unit + light-integration tests for AMD 1-GPU PR CI (#31088) Michael 2026-07-15 14:36:05 -07:00
• ec32590025 feat(moriep): add fp4 combine dtype (SGLANG_MORI_COMBINE_DTYPE=fp4) (#30706) karverma-amd 2026-07-15 16:21:41 -05:00
• e78051a419 [AMD] [Fix] Fix --attention-backend triton work for DeepSeek MLA on MI355 (null-K + decode dispatch + RoPE) (#30355) jacky.cheng 2026-07-16 05:19:23 +08:00
• e76cc75cfa [CI] Remove nightly registrations redundant with scheduled stage runs (#31371) Liangsheng Yin 2026-07-15 14:12:07 -07:00
• 5abec3fbf8 Fix MiMo-V2 on Blackwell: FA3 fallback and TP-aware audio weight loading (#31343) Yuhao Yang 2026-07-16 04:26:27 +08:00
• 18043aec20 [CI] Fix TRTLLM MHA graph metadata test fixture (#31332) Mohammad Miadh Angkad 2026-07-16 03:44:51 +08:00
• ab627e5d75 fix: load the right mtp lm head quantization (#30976) Shaun Kotek 2026-07-15 22:12:04 +03:00
• 7e7129acd7 [Bugfix] Release Mamba cache after PP dynamic chunk profiling (#31321) Xuwei 2026-07-16 02:32:07 +08:00
• 2d00e20a52 [Disagg][Qwen3.5] Fix heterogeneous attn-TP scatter transfer: GDN conv sub-block slice + GQA replicated-KV head map (#30997) YAMYandXuwei Li 2026-07-15 11:31:37 -07:00
• dd2e4cdc99 Add Inkling cookbook (#31360) Yuhao Yang 2026-07-16 02:24:31 +08:00
• 9d147fdca1 [Multimodal] Support n>1 outputs for GLM-Image generation (#31027) AuFlowandAuFlow 2026-07-16 01:55:19 +08:00
• dc078ddd2a [Spec] Extract stateless draft prepare helpers into eagle_worker_common (#31257) Liangsheng Yin 2026-07-15 10:54:31 -07:00
• d36e96ce23 [AMD] Enable mamba-extra-buffer for Qwen3.5 on ROCm (#30359) Bingxu Chenandntgiang71096 2026-07-16 01:18:12 +08:00
• a8b60433c2 [AMD] Fix DSV4 JIT build on rocm (#31131) YC Yen-Ching Tsengandkangwangamd 2026-07-16 00:58:12 +08:00
• c879f3da5c [diffusion] rl: support rl rollout for the wan pipeline via a per-request scheduler switch (#30036) Andy YeandClaude Fable 5 2026-07-15 07:30:51 -07:00
• d2b1243be0 docs: document CUDA crash dump output (#31333) Jun LiuandXinyuan Tong 2026-07-15 23:18:36 +09:00
• 495ae9aaa6 Fix Ministral3 accuracy issue by aligning YaRN RoPE scaling with Transformers implementation (#31232) 3 people 2026-07-15 15:23:04 +03:00
• 8ed82afcc8 [MoE Refactor] [NPU] Refactor Ascend MoE implementation to reduce code duplication and align with community design (#25663) Артем Савкинandronnie_zheng 2026-07-15 14:59:42 +03:00
• c9b17403e7 Fix image URL response for multiple outputs (#30621) 3 people 2026-07-15 19:49:04 +08:00
• 947a14d617 feat: unify multimodal feature transport (#30904) Mick 2026-07-15 17:42:38 +08:00
• f2c875d1c8 [PD] Route PD server warmup to every DP rank (#30748) weireweireandweireweire 2026-07-15 15:59:41 +08:00
• 5af670284e [CI] Lower GLM-5.2 NVFP4 MTP speed threshold (#31289) Mohammad Miadh Angkad 2026-07-15 15:39:58 +08:00
• fbcbe0a986 cookbook(deepseek-v4): add MORI disagg backend for AMD + bump MI355X image (#30651) 4 people 2026-07-15 08:33:05 +01:00
• 241937af87 [NPU] Determine the topk norm_type through scoring_func (#31107) 4 people 2026-07-15 15:26:44 +08:00
• dec0836302 Fix processor config loading for object-storage model paths (#31211) Sam ShleiferandAlex Nails 2026-07-15 00:21:31 -07:00
• 980acd6eca Fix MoE TP allreduce to use NCCL symmetric memory via in-pool output allocation (#29007) skyandBrayden Zhong 2026-07-15 15:06:37 +08:00
• 41e0b4b369 [CPU] add fused input proj for qwen3.5 (#31171) Ma Mingfei 2026-07-15 15:06:24 +08:00
• a649b5a9db [KDA] Add FlashInfer SM100 KDA decode + MTP (target_verify) backend (#30113) Yuan Luoandluoyuan.luo 2026-07-15 15:04:20 +08:00
• 1afab30577 Fix bookkeeping fields not encapsulated with real allocations in normal alloc, PD pre-alloc, DFlash and EAGLE (#29432) fzyzcjy 2026-07-15 14:52:21 +08:00
• e789ca24a7 Lightweight extract allocation logic from mem_cache/common.py to more clearly show nearly parallel variants (#29431) fzyzcjy 2026-07-15 14:49:37 +08:00
• c315df49bb Fix abusing presence of req.req_pool_idx to indicate the presence of req.kv resources (#29430) fzyzcjy 2026-07-15 14:48:39 +08:00
• 2d979f1d8c Let the presence of req.kv indicate the existence of owned kv resources (#29429) fzyzcjy 2026-07-15 14:47:16 +08:00
• 27256aee5b Let cache backend do not couple with owned committed kv details and avoid kv_committed_freed/kv_overallocated_freed fields (#29428) fzyzcjy 2026-07-15 14:43:10 +08:00
• d8d76c4d12 Introduce req.kv container for coupled owned kv field lifecycle (#29427) fzyzcjy 2026-07-15 14:40:38 +08:00
• 201ddeaba1 Avoid relaying per-step outputs through ScheduleBatch fields in disagg prefill and PP (#30677) fzyzcjy 2026-07-15 14:33:37 +08:00
• 01343d2759 Avoid implicit running_batch access in dllm and pdmux scheduling (#30676) fzyzcjy 2026-07-15 14:33:00 +08:00
• 21c62b9830 Rewrite pause_generation retract path as req-level release and requeue for clarity (#30675) fzyzcjy 2026-07-15 14:32:12 +08:00
• 1967b9ec99 Fix missed hisparse release and stale field cleanup in pause retract (#30674) fzyzcjy 2026-07-15 14:31:11 +08:00
• b6cc897fea Fix non-existent abort mode in Scheduler.pause_generation and inline retract_all (#30673) fzyzcjy 2026-07-15 14:27:48 +08:00
• 52a88fb212 Avoid mutating ScheduleBatch fields in place (#30672) fzyzcjy 2026-07-15 14:27:08 +08:00
• e77d95c3d5 Pass per-forward overrides to ForwardBatch.init_new as explicit arguments (#30670) fzyzcjy 2026-07-15 14:25:59 +08:00
• 861d97d24d Remove dead ScheduleBatch fields and avoid inplace seq_lens bump (#30669) fzyzcjy 2026-07-15 14:23:31 +08:00
• a3194d3585 [AMD] Remove ROCm page_first+kernel -> layer_first HiCache fallback (follow-up to #28534) (#30622) AMD-yanfeiwang 2026-07-15 13:48:16 +08:00
• 0832d856ca [Bugfix] fix quickreduce acc error in cudagraph mode (#29508) haoyangli0109 2026-07-15 13:16:02 +08:00
• 4aadf94146 [Kernel] Relocate vendored fla and mamba kernel trees to sglang.kernels (RFC #29630, Phase 2.5, 7/7) (#30795) Xiaoyu ZhangandClaude Fable 5 2026-07-15 12:52:15 +08:00
• 23f2b77d82 Make UTs compatible for XPU (#27106) ANSHUMAN TRIPATHY 2026-07-15 10:05:56 +05:30
• c00131ebaa [Kernel] Migrate linear-attention, MiniMax-sparse and diffusion kernels to sglang.kernels (RFC #29630, Phase 2.5, 6/7) (#30793) Xiaoyu ZhangandClaude Fable 5 2026-07-15 11:21:36 +08:00
• ba5be86d42 [Kernel] Migrate DSA + DSV4 attention kernels to sglang.kernels (RFC #29630, Phase 2.5, 5/7) (#30792) Xiaoyu ZhangandClaude Fable 5 2026-07-15 11:11:22 +08:00
• 4ae9cc3c81 Fix gate stride for 4D decode layouts (#31231) 3 people 2026-07-14 20:06:50 -07:00
• b4fdce3b63 Fix post-capture KV sizing for SWA pools (#31092) Lianmin Zheng 2026-07-14 20:06:15 -07:00
• 532cd337ed [Intel GPU] DeepSeek V4 12/N: use sgl-kernel implementation of silu_and_mul_clamp to run on XPU (#28428) Polisetty V R K Jyothendra Varma 2026-07-15 08:33:53 +05:30
• 46b675ce70 [Intel GPU] DeepSeek V4 11/N: support fp8_paged_mqa_logits_triton from sgl-kernel to run on XPU (#28059) Polisetty V R K Jyothendra VarmaandMa Mingfei 2026-07-15 08:33:35 +05:30
• aafa706f8f [AMD] Update qwen3.5 cookbook (#31258) Thomas Wang 2026-07-15 10:59:12 +08:00
• 43124cdd90 fix: fix image benchmark backend parity (#30867) Mick 2026-07-15 10:11:22 +08:00
• b22f20b660 [CI] Fix CUDA 12 NVIDIA wheel cleanup (#31035) Hank Han 2026-07-15 10:06:11 +08:00
• 76dc427806 [Spec] Single-source num_tokens_per_req derivation and access (#31013) Liangsheng Yin 2026-07-14 18:41:08 -07:00
• ca0ee3f1a8 [Spec] Consolidate spec-worker weight updates into BaseSpecWorker via draft_runners (#31078) Liangsheng Yin 2026-07-14 18:39:46 -07:00
• a9cf5e68e6 [DSV4] Remove per-step seqlen D2H from speculative to make overlap scheduler work (#30365) weireweireandweireweire 2026-07-15 09:01:22 +08:00
• b8a00e2ec8 docs: sync LMSYS SGLang blog cards (#31242) sglang-botandsglang-bot 2026-07-14 17:49:30 -07:00
• 90f10cbe26 [diffusion] post_training: Add LoRA IPC weight sync via lora_merge mode (#31029) WenhaoZhang 2026-07-15 08:42:54 +08:00
• 50d1edaa7f [misc] Move SchedulerRecvSkipper into scheduler_components (#31222) Liangsheng Yin 2026-07-14 16:35:04 -07:00
• 21a6d08557 ci: strip invisible Unicode format chars from slash-command input (#31234) Alison Shao 2026-07-14 16:33:45 -07:00
• 771e386332 Disable flaky DSV4-Flash FP4 BCG determinism test (nondeterminism from #30898 idle-rank dummy extend) (#31125) Alison Shao 2026-07-14 16:21:10 -07:00
• 463a3f4248 [Mamba] Support configurable conv-window layouts (#31059) paulzhang-tm 2026-07-14 14:41:10 -07:00
• 08c46e1f1a Add dummy forward batch preparation hook (#31070) paulzhang-tm 2026-07-14 14:30:31 -07:00
• 0d89564d27 Support scheduler_recv_interval (recv skipper) under DP-attention (#30457) 4 people 2026-07-14 16:02:58 -05:00
• bdc9848c25 [Doc]Standardize the names of PyTorch NPU-related software throughout the documentation by replacing them all with TorchNPU. (#29886) 3 people 2026-07-15 02:56:08 +08:00
• cb47a68717 [PD] Stride KV token->page indices on device before D2H copy (#31173) cctryandcctry 2026-07-14 10:08:30 -07:00