vllm-project/vllm

[Model Loader] Make dummy weight init dtype-aware

[Bugfix][Frontend] Honor default penalties in transcription and translation requests bug frontend

[Perf][MoE] Batch per-expert token count GPU sync into single transfer

[Core] Expose prefix reuse policy in KV cache group metadata

[Perf][Attention] Cache FlashAttention KV cache transpose+split views

[Bugfix][Rust Frontend] Read the text config under `thinker_config` bug rust

[Bugfix][CPU] Reject -1 placeholder drafts in the C++ rejection sampler bug ci/build cpu

[Attention] Remove cascade attention documentation performance rocm intel-gpu speculative-decoding ready torch.compile qwen deepseek cpu kv-connector nvidia mrv2 DSv4 dflash kimi k3 scheduler kv-cache-manager minimax inkling DSv4.1

[Bugfix] Bump tokenspeed-mla to 0.2.21 to fix nanobind conflict with Triton 3.9 bug ci/build

[SketchSSM] Add SketchSSM decode for Mamba-2 with Triton kernels documentation performance ci/build mrv2

[Bugfix] Reject dynamic speculative decoding with ngram_gpu bug

[Bugfix] Preserve empty tensor metadata in packed IPC bug

[KV Connector][NIXL] Transfer DeepSeek-V4.1 packed MLA rows whole deepseek kv-connector DSv4.1

[Bugfix] Don't force a second </think> after a natural end in V1 spec decode bug

[ROCm][Perf][MiniMax-M3] Defer the FFN all-reduce into the next layer's norm rocm minimax

[Bugfix] Disable sequence-parallel MoE under batch invariance bug needs-rebase

[EPLB] Make migration peer limit configurable documentation

[Model] Support EAGLE3/DFlash aux hidden states for GraniteMoe and GraniteMoeShared dflash

[ROCm][Perf] Island all-reduce for TP8 on 2x4 XGMI-island nodes (MI210) documentation rocm ci/build nvidia

[Fix][Compile] Declare group_shape in the static FP8 quant fusion … bug ready torch.compile quantization

[Fusion] Fuse Qwen MoE shared MLP activation quantization torch.compile qwen quantization

[Test] Compare NVFP4 batch-invariance logprobs bitwise

[Test] Make the determinism tests runnable for MLA models

[Bugfix][Model] Fix EXAONE-4.5 image inputs: ViT received un-normalised pixels bug

[Docs][Distributed] Document GLOO_SOCKET_IFNAME for docker pipeline parallel deployments (#59109) documentation

[Bugfix][Distributed] Prevent silent split-brain in stateless DP on port collision (#61049) bug

[Bugfix][Spec Decode] Support dynamic speculative decoding in ngram_gpu (#60981) bug speculative-decoding

WIP: [Bugfix][ROCm] Preserve 64-bit QuickReduce phase offsets bug rocm

WIP: [Bugfix][Model] Preserve all Kimi-Audio embeddings in batched requests bug multi-modality kimi

[Bugfix][Multimodal] Use expanded prompt bounds for routed-expert outputs bug