Upstream sync 2026-05-25 (158 commits) - #101

Draft

Clarit-AI wants to merge 160 commits into

mainfrom

upstream-sync-20260525

Commits on May 21, 2026

[MoE Refactor] deprecate forward_npu and NpuFuseEPMoE (#25678)

ch-wan
and
claude
authored
b765fae
Default MegaMoE to W4A8 for Max-Throughput recipe (#26004)
yhyang201
authored
81d686d
[Fix]: Restrict Kimi-K2.5 shared-experts fusion to Quark MXFP4 checkpoints (#25974)
Jiminator
authored
1a85586
[Docs] DeepSeek-V4: switch H200 FP4 Pro to flashinfer_mxfp4, Flash Balanced too (#25923)

zijiexia
and
claude
authored
17dadeb
[Spec] Polish FutureMap after #25879: rename callback, async guard, cleanup (#25962)
hnyls2002
authored
c9a0e55
[core] Unify output_tokens_buf in FutureMap (#25922)
hnyls2002
authored
44ec2ee
feat(model_runner): remove pool/backend refs from ForwardBatch via ForwardContext (#25983)

ch-wan
and
claude
authored
c5251a9
refactor(attn): init hisparse_coordinator before attn_backend; replace lazy property with init-time capture (#26012)

ch-wan
and
claude
authored
d765dfd
feat: support HybridLinearKVPool in chunked prefix cache handling (#25753)
imp2002
authored
7cf193f

Commits on May 22, 2026

Commits on May 23, 2026

Commits on May 24, 2026

Commits on May 25, 2026

Commits on May 26, 2026

[SRT] minor: reuse req input id array for unpadded ids (#26232)
mickqian
authored
8f2a4e7
[misc] Update permission (#26336)
Fridge003
authored
29e245e
[Intel GPU] Fix incorrect KV-cache page table for local attention when page_size > 1 (#23757)
ckvermaAI
authored
156d1af
Signal CUDA coredumps to tracker issue (#26338)
hnyls2002
authored
1953565
[diffusion] feat: layerwise NVTX markers for Nsight Systems profiling (#25683)

mispa-ms
and
mickqian
authored
3142278
Support DeepSeek V4 DeepEP Waterfill (#25391)
xutizhou
authored
59cad67
GLM-4.7-Flash: standalone MLA impl and MLA NextN/MTP (#26088)
zRzRzRzRzRzRzR
authored
7ef06bf
[kernel] reuse wna16 marlin moe workspace (#26112)
mickqian
authored
63a89bf
[Doc] Update pip install commands for Cuda12 (#26344)
Fridge003
authored
f0ba651
fix: fix fa3 cross-attention batched-decode for per-request varlen encoder (#25045)
zsj555
authored
7c0fbc8
[AMD] Dsv4/pr2 compressor opt (#26208)

authored
3f5e2c7
[Perf][Moe]improve cutlass_moe_fp4 performance by using apply_router_weight_on_i… (#19493)
chengchao23
authored
1371685
fix(tool_call): normalize non-standard JSON Schema types in tool params (#23476)
JustinTong0323
authored
64c7c68
[GDN] Support SM100 CuTeDSL GDN Prefill Kernel (#26200)

yuan-luo
and
luoyuan.luo
authored
d34d4d9
[EPD] Cross-request batching for image/audio encoder (#25964)
Abatom
authored
dabdd91
Revert "fix(tool_call): normalize non-standard JSON Schema types in tool params" (#26379)
hnyls2002
authored
7e6e5ef
[feat] Support `extra_buffer` in Mamba2-based models (#15829)
roikoren755
authored
e958f45
[AMD] Relaxing timeout for AMD CI (#26392)
yctseng0211
authored
d25a220
[HiCache]: Check return code of cudaHostRegister (#26301)
stepinto
authored
48f3264
Extract Scheduler init methods and add skills to enforce the splitting requirements (#26271)
fzyzcjy
authored
d9c8293
Revert "[perf][spec decoding] Skip full-vocab softmax in EAGLE draft when topk == 1 (#26235)" (#26358)
michaelzhang-ai
authored
9409969
fix(ci): enforce legacy docs/ gate in Lint workflow (#26322)
zijiexia
authored
a269131
[PP] Skip PP output communication for pure chunked prefill batches (#26148)
xu-yfei
authored
98eb844
[PD] Fix cross-rank queue divergence by gating metadata readiness before all-reduce (#26394)
ShangmingCai
authored
c8c1aed
[PD] Fix top logprobs crash in prefill path (#26299)
junliu-mde
authored
c47f0e7
[UnifiedRadixTree]: Support L3 HiStorage framework (#26062)
hzh0425
authored
38f32c3
[Bugfix] Fix flashinfer_cutlass MoE crash when intermediate_size_per_partition is not 16-aligned (#22627)

vguduruTT
and
vguduruTT
authored
6c81286
[Spec] Async-assert probes across EAGLE/MTP; zero `tgt_cache_loc` (#26335)
hnyls2002
authored
ec6f8d6
Sgl flashmla (#26132)
zcnrex
authored
b66f8e0
[docs] DeepSeek-V4 cookbook: note cu129 image for GB200 Pro DeepEP backend (#26413)

zijiexia
and
claude
authored
6afebc2
chore: bump sgl-kernel version to 0.4.3 (#26414)
sglang-bot
authored
0753182
[CPU Doc]Add Xeon CPU info in Qwen3 Cookbook (#25971)
ZailiWang
authored
47617cc
[RL] Fix FP8 skip matching for trailing-dot prefixes (#26287)
zianglih
authored
2b1e53c
chore(engram): close dangling KHA-390 marker

Sync Review Bot
and
claude
committed
61873ba
Merge upstream/main into engram (2026-05-25 sync, 158 commits)

Sync Review Bot
and
claude
committed
43b1bd3