Commits vllm-project/vllm

[Rust Frontend] mock engine: decode add requests with auxiliary frames

BugenZhao committed 39 minutes ago

Verified d4abbbc7

[Perf][KV Offload] Split the shared Tiering region and skip the scheduler pre-fault (#66391)

Etelis committed 7 hours ago

Verified 9e6815ac

[Model] Remove TeleChat2 and TeleChat3 (#67035)

hmellor committed 7 hours ago

Verified d9690929

[Recipe] MiniMax-M3.1-Flash: sync the GB300 Dynamo AgentX presets to InferenceX #4998 (#67184)

Juntian777 committed 8 hours ago

Verified c6573eed

[Misc] Use the b12x package shipped with flashinfer-python (#64765)

stecasta committed 8 hours ago

Verified efe12305

[Bugfix] Use max_split_size_mb:21 so freed 1-10 MiB tensors are reused (#66561)

viktorkou committed 8 hours ago

Verified 6d3eba02

[ROCm][Disagg][CI] Set token budgets for the MiniMax-M3 wide-EP cells (#66646)

avininjamay8 committed 9 hours ago

Verified 7a8f1f9e

[Doc][Tests] Stop setting `gpu_memory_utilization` in most extensible KV cache cases (#67071)

njhill committed 9 hours ago

Verified 3b95523e

[Rust Frontend] Bump xgrammar-structural-tag to 0.5 and compose Lark answers (#66692)

BugenZhao committed 10 hours ago

Verified be47189f

[Frontend] Bump xgrammar to 0.2.9 and compose Lark answers as Lark (#66691)

BugenZhao committed 10 hours ago

Verified 4fee061e

[Bugfix][Frontend] Pool HF fast tokenizers in BaseRenderer so concurrent truncation stays per request (#65932)

haishuok0525 committed 10 hours ago

Verified cfe069c9

[ROCm][Attention] Add an mxfp4_mla KV cache for rope-free sparse MLA (read/write path) (#60321)

amd-dlimpus committed 11 hours ago

Verified dac51063

[Perf][Frontend] Trim per-chunk and per-request work in the completions stream (#66325)

davidteather committed 11 hours ago

Verified 54ba5698

[Bugfix] Remove compiled models' bytecode hooks at in-process shutdown (#67081)

njhill committed 11 hours ago

Verified ee548e7f

[ROCm][Bugfix] Route biased softmax/sigmoid top-k to AITER topk_gating (#67107)

rjrock committed 11 hours ago

Verified b5dcc877

[Bugfix][Multimodal] Cache multi-page Unlimited-OCR-2 requests in the processor cache (#66638)

subhashpolisetti committed 11 hours ago

Verified 2a14a5e4

[Perf][mHC] use FlashInfer mega mHC for MHCFusedPostPreOp (GLM-5.3-Fl… (#65521)

Ayu190505 committed 11 hours ago

Verified 3de0b48c

[KVConnector] Allocate the extensible KV cache exportable for EXPORTS_FD (LMCache MP) (#66674)

ApostaC committed 11 hours ago

Verified 6201c7d3

[Rust Frontend] Stream MiniMax M3 tool-call arguments (#66531)

BugenZhao committed 12 hours ago

Verified 8527d757

[Rust Frontend] Match MiniMax M3 markers by token identity (#66767)

BugenZhao committed 12 hours ago

Verified 9a4decde

[MoE] Replace routed-scale bool flags with a RoutedScaleMode enum (#66859)

nliu365 committed 12 hours ago

Verified c9c36486

[XPU][UT] Increase the prompt length in the late chunking test to allow prefix cache hits on XPU (#66978)

RyanMa29 committed 12 hours ago

Verified 52373885

[ROCm][MLA][DCP] Pad the AITER MLA decode head count to a native tile (#54899)

okorzh-amd committed 12 hours ago

Verified a547de8a

[Perf][GLM-5.3] DeepGEMM BF16 MegaMoE for unquantized checkpoints (#67040)

gcanlin committed 12 hours ago

Verified 20ad1db3

[ROCm][DSv4/4.1] Enable Tensor scale FP8 KV Cache for Sparse MLA (#66036)

cagrikymk committed 12 hours ago

Verified 0854d2f2

[ROCm][Perf] AITER FA: trim the sliding-window decode table with fewer ops (#63392)

okorzh-amd committed 13 hours ago

Verified 11ddd587

[Bugfix][XPU] Remove CUDA-only dependencies from Unlimited-OCR-2 DASH attention (#66495)

faaany committed 13 hours ago

Verified d838863e

[ROCm][Perf][DSv4.1] Mono: tuned K2 FFN seam at TP4 / TP8 (#64883)

okorzh-amd committed 13 hours ago

Verified 8f0a381f

[Bugfix] Exclude FlashInfer autotune peak from KV cache sizing (#66197)

janbernloehr committed 13 hours ago

Verified 318fca3f

[Attention][MiniMax-M3] Add FlashInfer MSA backend for SM90 and SM120/SM121 (#66018)

rmhaskarnvidia committed 13 hours ago

Verified d0eca76e