[sync] vLLM 2026-07-12 — 766469a4c..83762b77b#1708

Merged

AlpinDale merged 49 commits into

mainfrom

sync/vllm-2026-07-12-83762b77b

Jul 12, 2026

Commits on Jul 12, 2026

[sync] [ROCm] Revert Part of `[ROCm] Fix pooling startup workspace lock` #47912 (#48154)

AlpinDale
and
micah-wil
committed
[sync] Correct model layer aliasing for Bert style models (#43896)

AlpinDale
and
ap9272
committed
[sync] update marlin M size for EP (#48144)

AlpinDale
and
gnovack
committed
[sync] [bugfix] bge-m3-sparse-plugin mismatch requests (#48112)

AlpinDale
and
staugust
committed
[sync] [CI/Build][AMD] Fix ROCm OOM in eagle_correctness_heavy by reserving CUDA graph memory (#47366)

AlpinDale
and
peizhang56
committed
[sync] [kv_offload] Emit tier-owned BlockStored events from FS/OBJ secondary tiers (#47923)

AlpinDale
and
Change72
committed
[sync] DCP supports hybrid attention (#40996)

AlpinDale
and
Yancey0623
committed
[sync] [Core][KV events] Report prefix-cache-reused blocks in full report mode (#45261)

AlpinDale
and
GongLei-HW
committed
[sync] [Feature][Parser] Support include_reasoning param for non-Harmony models (#44301)

AlpinDale
and
albertoperdomo2
committed
[sync] [Perf] fuse more rmsnorm and all-reduce in qwen3.5 (#46998)

AlpinDale
and
ZJY0516
committed
[sync] [Rust Frontend] Integrate MM video support (#47959)

AlpinDale
and
BugenZhao
committed
[sync] [Model][CI/Build] Cosmos3: enable registry tests and register Cosmos3-Super (#48211)

AlpinDale
and
guoriyue
committed
[sync] [CI] Add TORCH_NIGHTLY=1 build mode (run full suite on torch nightly) (#47180)

AlpinDale
and
atalman
committed
[sync] Deepstream video backend (#42424)

AlpinDale
and
ViranjanPagar
committed
[sync] [Rust Frontend] Add roundtrip fixtures for more chat parsers (#47883)

AlpinDale
and
reidliu41
committed
[sync] [Misc] Remove dead code in ViT functionality test (#48220)

AlpinDale
and
Isotr0py
committed
[sync] [Bugfix][Spec Decode] Fix DFlash draft/target layer-count mismatch (#48113)

AlpinDale
and
njhill
committed
[sync] [Model] Migrate MistralLarge3ForCausalLM to AutoWeightsLoader (#48153)

AlpinDale
and
Functionhx
committed
[sync] [Refactor] Remove unused rocm kernel `combine_topk_swa_indices_ragged` (#48158)

AlpinDale
and
yewentao256
committed
[sync] [Bugfix] Fix turboquant FP8 cast failure for BF16 models on Ampere GPUs (#39988)

AlpinDale
and
XuZhou26
committed
[sync] fix: correct load_weights track logic and enable weight integrity for… (#41811)

AlpinDale
and
MynameFelix
committed
[sync] [Build/CI] Build arm64 PR and postmerge image builds for Blackwell SM10x and SM110 (#48041)

AlpinDale
and
tlrmchlsmth
committed
[sync] [Model] Add LongCat-Flash-Lite (n-gram embedding) (#47857)

AlpinDale
and
mgoin
committed
[sync] [ROCm] Enable DeepSeek-V4 DSpark speculative decoding on AMD (MI350X / MI355X, gfx950) (#47419)

AlpinDale
and
larryli2-amd
committed
[sync] [Bugfix] Fix FlashMLA dense fp8 metadata crash (num_sm_parts clamp) (#48045)

AlpinDale
and
MatthewBonanni
committed
[sync] handle topk_ids padding in align sum kernel (#47785)

AlpinDale
and
gnovack
committed
[sync] [Bugfix][Test] Register Qwen/Qwen3.5-4B example model (#48276)

AlpinDale
and
njhill
committed
[sync] [Bugfix] Fix thinking_token_budget not enforced after natural </think> re-entry (#45984)

AlpinDale
and
ashwing
committed
[sync] Add VLLM_FLASHINFER_AUTOTUNE_SKIP_OPS and skip CuTeDSL fp4_gemm autotuning by default (#48268)

AlpinDale
and
mgoin
committed
[sync] [Quantization] Bound peak memory when repacking FP4 MoE weights for Marlin (#47851)

AlpinDale
and
joerowell
committed
[sync] [BugFix] weights processing peak memory reduction for nvfp4 MoE layers (#46276)

AlpinDale
and
thisisjimmyfb
committed
[sync] [BugFix] Fix packed HND KV cache reshape for FlashAttention (#47314)

AlpinDale
and
LucasWilkinson
committed
[sync] [Misc] Use meta tensor for KV cache stride calculation (#47316)

AlpinDale
and
LucasWilkinson
committed
[sync] [Logs] DP Supervisor Log Improvement (#48278)

AlpinDale
and
robertgshaw2-redhat
committed
[sync] [Revert] [Build] Update vllm ...builds FA3 with torch stable API (#48269)

AlpinDale
and
LucasWilkinson
committed
[sync] Bump Transformers version to 5.13.0 (#47867)

AlpinDale
and
hmellor
committed
[sync] [XPU]remove is_xxx from moe class and bump up kernels (#48079)

AlpinDale
and
mayuyuace
committed
[sync] [CI] Point CI at Transformers release rather than release branch (#48328)

AlpinDale
and
hmellor
committed
[sync] FP32 router GEMV optimization (#48335)

AlpinDale
and
jeejeelee
committed
[sync] [XPU][UT]Fix InternS1ProForConditionalGeneration AssertionError (#48232)

AlpinDale
and
Yejing-Lai
committed
[sync] [2/N][KV-Cache Layout Refactor] Pack K/V into the content dim across attention backends (#44455)

AlpinDale
and
LucasWilkinson
committed
[sync] fix(entrypoints): stop resolve_items leaking in-flight media fetch tasks on partial failure (#48333)

AlpinDale
and
ErenAta16
committed
[sync] fix(processor): route MiMo-V2-Omni media fetch through MediaConnector (#43117)

AlpinDale
and
ibondarenko1
committed
[sync] [Bugfix][LoRA] Support ark_linear base layer in _get_lora_device (#47690)

AlpinDale
and
AlejandroParedesLT
committed
[sync] [CI][CPU] Add Qwen2-VL multimodal tests for CPU backend and fix incompatibilities (#48072)

AlpinDale
and
zhejiangxiaomai
committed
[sync] [2/N][Core] support partial prefix cache hit for hybrid model (#46384)

AlpinDale
and
ZJY0516
committed
[sync] Runtime Draft Weight Update for Speculative Decoding (#46725)

AlpinDale
and
vx120
committed
[sync] [Perf][Qwen] Replace MOE all-reduce with reduce-scatter (#47006)

AlpinDale
and
gcanlin
committed
[sync] [Frontend] Add /abort_requests to the RLHF dev API router (#47173)

AlpinDale
and
aoshen02
committed