Commit Graph
15252 Commits
Author SHA1 Message Date
Mohammad Miadh Angkad 5af670284e [CI] Lower GLM-5.2 NVFP4 MTP speed threshold (#31289) 2026-07-15 00:39:58 -07:00
fbcbe0a986 cookbook(deepseek-v4): add MORI disagg backend for AMD + bump MI355X image (#30651)
Co-authored-by: Zijie Xia <zijie.xia@radixark.ai>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: zijiexia <37504505+zijiexia@users.noreply.github.com>
2026-07-15 07:33:05 +00:00
241937af87 [NPU] Determine the topk norm_type through scoring_func (#31107)
Co-authored-by: iridiumine <42236072+iridiumine@users.noreply.github.com>
Co-authored-by: zhaozx-cn <59479021+zhaozx-cn@users.noreply.github.com>
Co-authored-by: sglang-npu-bot <sglangnpu@163.com>
2026-07-15 10:26:44 +03:00
Sam ShleiferandAlex Nails dec0836302 Fix processor config loading for object-storage model paths (#31211)
Co-authored-by: Alex Nails <alex.nails@radixark.ai>
2026-07-15 00:21:31 -07:00
skyandBrayden Zhong 980acd6eca Fix MoE TP allreduce to use NCCL symmetric memory via in-pool output allocation (#29007)
Signed-off-by: wangfakang <fakangwang@gmail.com>
Co-authored-by: Brayden Zhong <b8zhong@uwaterloo.ca>
2026-07-15 15:06:37 +08:00
Ma Mingfei 41e0b4b369 [CPU] add fused input proj for qwen3.5 (#31171) 2026-07-15 15:06:24 +08:00
Yuan Luoandluoyuan.luo a649b5a9db [KDA] Add FlashInfer SM100 KDA decode + MTP (target_verify) backend (#30113)
Co-authored-by: luoyuan.luo <luoyuan.luo@antgroup.com>
2026-07-15 15:04:20 +08:00
fzyzcjy 1afab30577 Fix bookkeeping fields not encapsulated with real allocations in normal alloc, PD pre-alloc, DFlash and EAGLE (#29432) 2026-07-15 14:52:21 +08:00
fzyzcjy e789ca24a7 Lightweight extract allocation logic from mem_cache/common.py to more clearly show nearly parallel variants (#29431) 2026-07-15 14:49:37 +08:00
fzyzcjy c315df49bb Fix abusing presence of req.req_pool_idx to indicate the presence of req.kv resources (#29430) 2026-07-15 14:48:39 +08:00
fzyzcjy 2d979f1d8c Let the presence of req.kv indicate the existence of owned kv resources (#29429) 2026-07-15 14:47:16 +08:00
fzyzcjy 27256aee5b Let cache backend do not couple with owned committed kv details and avoid kv_committed_freed/kv_overallocated_freed fields (#29428) 2026-07-15 14:43:10 +08:00
fzyzcjy d8d76c4d12 Introduce req.kv container for coupled owned kv field lifecycle (#29427) 2026-07-15 14:40:38 +08:00
fzyzcjy 201ddeaba1 Avoid relaying per-step outputs through ScheduleBatch fields in disagg prefill and PP (#30677) 2026-07-15 14:33:37 +08:00
fzyzcjy 01343d2759 Avoid implicit running_batch access in dllm and pdmux scheduling (#30676) 2026-07-15 14:33:00 +08:00
fzyzcjy 21c62b9830 Rewrite pause_generation retract path as req-level release and requeue for clarity (#30675) 2026-07-15 14:32:12 +08:00
fzyzcjy 1967b9ec99 Fix missed hisparse release and stale field cleanup in pause retract (#30674) 2026-07-15 14:31:11 +08:00
fzyzcjy b6cc897fea Fix non-existent abort mode in Scheduler.pause_generation and inline retract_all (#30673) 2026-07-15 14:27:48 +08:00
fzyzcjy 52a88fb212 Avoid mutating ScheduleBatch fields in place (#30672) 2026-07-15 14:27:08 +08:00
fzyzcjy e77d95c3d5 Pass per-forward overrides to ForwardBatch.init_new as explicit arguments (#30670) 2026-07-15 14:25:59 +08:00
fzyzcjy 861d97d24d Remove dead ScheduleBatch fields and avoid inplace seq_lens bump (#30669) 2026-07-15 14:23:31 +08:00
AMD-yanfeiwang a3194d3585 [AMD] Remove ROCm page_first+kernel -> layer_first HiCache fallback (follow-up to #28534) (#30622) 2026-07-14 22:48:16 -07:00
haoyangli0109 0832d856ca [Bugfix] fix quickreduce acc error in cudagraph mode (#29508)
Signed-off-by: Haoyang Li <lihaoyang0109@gmail.com>
2026-07-14 22:16:02 -07:00
Xiaoyu ZhangandClaude Fable 5 4aadf94146 [Kernel] Relocate vendored fla and mamba kernel trees to sglang.kernels (RFC #29630, Phase 2.5, 7/7) (#30795)
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
2026-07-15 12:52:15 +08:00
ANSHUMAN TRIPATHY 23f2b77d82 Make UTs compatible for XPU (#27106) 2026-07-15 12:35:56 +08:00
Xiaoyu ZhangandClaude Fable 5 c00131ebaa [Kernel] Migrate linear-attention, MiniMax-sparse and diffusion kernels to sglang.kernels (RFC #29630, Phase 2.5, 6/7) (#30793)
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
2026-07-15 11:21:36 +08:00
Xiaoyu ZhangandClaude Fable 5 ba5be86d42 [Kernel] Migrate DSA + DSV4 attention kernels to sglang.kernels (RFC #29630, Phase 2.5, 5/7) (#30792)
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
2026-07-15 11:11:22 +08:00
4ae9cc3c81 Fix gate stride for 4D decode layouts (#31231)
Co-authored-by: lmzheng <lmzheng@fb.com>
Co-authored-by: michael604work <michael604@meta.com>
2026-07-14 20:06:50 -07:00
Lianmin Zheng b4fdce3b63 Fix post-capture KV sizing for SWA pools (#31092) 2026-07-14 20:06:15 -07:00
Polisetty V R K Jyothendra Varma 532cd337ed [Intel GPU] DeepSeek V4 12/N: use sgl-kernel implementation of silu_and_mul_clamp to run on XPU (#28428)
Signed-off-by: P V R K Jyothendra Varma <polisetty.v.r.k.jyothendra.varma@intel.com>
2026-07-15 11:03:53 +08:00
Polisetty V R K Jyothendra VarmaandMa Mingfei 46b675ce70 [Intel GPU] DeepSeek V4 11/N: support fp8_paged_mqa_logits_triton from sgl-kernel to run on XPU (#28059)
Signed-off-by: P V R K Jyothendra Varma <polisettyvarma@gmail.com>
Co-authored-by: Ma Mingfei <mingfei.ma@intel.com>
2026-07-15 11:03:35 +08:00
Thomas Wang aafa706f8f [AMD] Update qwen3.5 cookbook (#31258) 2026-07-15 10:59:12 +08:00
Mick 43124cdd90 fix: fix image benchmark backend parity (#30867) 2026-07-15 10:11:22 +08:00
Hank Han b22f20b660 [CI] Fix CUDA 12 NVIDIA wheel cleanup (#31035)
Signed-off-by: Hank Han <hanhan7630@outlook.com>
2026-07-15 10:06:11 +08:00
Liangsheng Yin 76dc427806 [Spec] Single-source num_tokens_per_req derivation and access (#31013) 2026-07-14 18:41:08 -07:00
Liangsheng Yin ca0ee3f1a8 [Spec] Consolidate spec-worker weight updates into BaseSpecWorker via draft_runners (#31078) 2026-07-14 18:39:46 -07:00
weireweireandweireweire a9cf5e68e6 [DSV4] Remove per-step seqlen D2H from speculative to make overlap scheduler work (#30365)
Co-authored-by: weireweire <20922698+weireweire@users.noreply.github.com>
2026-07-14 18:01:22 -07:00
sglang-botandsglang-bot b8a00e2ec8 docs: sync LMSYS SGLang blog cards (#31242)
Co-authored-by: sglang-bot <sglang-bot@users.noreply.github.com>
2026-07-15 00:49:30 +00:00
WenhaoZhang 90f10cbe26 [diffusion] post_training: Add LoRA IPC weight sync via lora_merge mode (#31029) 2026-07-15 08:42:54 +08:00
Liangsheng Yin 50d1edaa7f [misc] Move SchedulerRecvSkipper into scheduler_components (#31222) 2026-07-14 16:35:04 -07:00
Alison Shao 21a6d08557 ci: strip invisible Unicode format chars from slash-command input (#31234) 2026-07-14 16:33:45 -07:00
Alison Shao 771e386332 Disable flaky DSV4-Flash FP4 BCG determinism test (nondeterminism from #30898 idle-rank dummy extend) (#31125) 2026-07-14 16:21:10 -07:00
paulzhang-tm 463a3f4248 [Mamba] Support configurable conv-window layouts (#31059) 2026-07-14 14:41:10 -07:00
paulzhang-tm 08c46e1f1a Add dummy forward batch preparation hook (#31070) 2026-07-14 14:30:31 -07:00
0d89564d27 Support scheduler_recv_interval (recv skipper) under DP-attention (#30457)
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: hnyls2002 <lsyincs@gmail.com>
Co-authored-by: Liangsheng Yin <hnyls2002@gmail.com>
2026-07-14 14:02:58 -07:00
bdc9848c25 [Doc]Standardize the names of PyTorch NPU-related software throughout the documentation by replacing them all with TorchNPU. (#29886)
Signed-off-by: a60124901 <anyuxin4@h-partners.com>
Signed-off-by: axx-ty911 <anyuxin4@h-partners.com>
Co-authored-by: sglang-npu-bot <sglangnpu@163.com>
Co-authored-by: ronnie_zheng <zl19940307@163.com>
2026-07-14 21:56:08 +03:00
cctryandcctry cb47a68717 [PD] Stride KV token->page indices on device before D2H copy (#31173)
Co-authored-by: cctry <cctry@fb.com>
2026-07-14 10:08:30 -07:00
hunhokimandHun-ho Kim a5c3e0283f [Bug fix] Account for KV replication fan-out in transfer-byte metrics (#30351)
Co-authored-by: Hun-ho Kim <hunho.kim@samsung.com>
2026-07-14 10:08:24 -07:00
Mohammad Miadh Angkad 271e5ef5c3 [CI] Fix Flash MLA SM120 test import path (#31199) 2026-07-14 09:01:23 -07:00
Mick 04af94d150 fix: avoid tilelang cuda runtime pollution (#30870) 2026-07-14 22:30:27 +08:00