All pull requests

Pull requests

• [Bugfix][KV Offload] Add VLLM_KV_OFFLOAD_DISABLE_BATCH_MEMCPY to bypass cuMemcpyBatchAsync

  bugnvidia
  #61054·jerry102039 opened Oct 11, 2026·

  ·
  1

  3

• [Bugfix] Bound CoW block-copy transient to avoid OOM near the memory limit

  bug
  #61051·michael-gpu opened Oct 11, 2026·6 tasks done·

  ·
  2

• [Bugfix][KV Offload] Preserve running request prefixes across LRU store batches

  bug
  #61050·ztysth opened Oct 11, 2026·6 tasks done·

  ·
  1

• [Core] Let TP-replicated Mamba state share a sharded state KV cache group

  kv-cache-manager
  #61048·kimseunghyun-kr opened Oct 11, 2026·4 of 6 tasks·

  ·
  1

• [Bugfix][Structured Output] Drain compilation tasks before releasing the backend

  bugstructured-output
  #61047·DimensionSTP opened Oct 11, 2026Contributor·6 tasks done·

  ·
  1

  1

• [Sleep] Retain draft-private weights with existing parameter selector

  documentationready
  #61046·aoshen02 opened Oct 11, 2026Contributor·6 tasks done·

  ·
  5

• Create cuBLAS handles before memory profiling to avoid OOM

  #61044·laithsakka opened Oct 11, 2026Contributor·

  ·
  1

  1

• [Frontend] Explain disabled scale-out endpoints with opt-in hints

  ci/builddocumentationfrontendreadyrust
  #61042·aoshen02 opened Oct 11, 2026Contributor·

  ·
  8

• [Bugfix][DSv4.1] Match decoder-replay input_ids dtype for FlashInfer MegaMoE

  bugdeepseekDSv4.1nvidia
  #61041·wangshangsam opened Oct 11, 2026Collaborator·6 tasks done·

  ·
  1

  

• [Perf][DSv4.1] Length-aware candidate block selection in the decode indexer

  DSv4.1
  #61039·Juntian777 opened Oct 11, 2026Contributor·

  ·

• [Bugfix] Keep legacy WNA16 MoE dispatch on Triton under batch invariance

  bug
  #61038·yifanFengg opened Oct 11, 2026Contributor·

  ·

• [Bugfix] Exchange MiniMax Lamport IPC handles through their pointer

  bugminimax
  #61037·yifanFengg opened Oct 11, 2026Contributor·

  ·

• [Bugfix] Skip the DeepSeek fused A GEMM kernel under batch invariance

  bugdeepseek
  #61036·yifanFengg opened Oct 11, 2026Contributor·

  ·

• [Bugfix] Pad unaligned custom all-reduce inputs under batch invariance

  bug
  #61035·yifanFengg opened Oct 11, 2026Contributor·

  ·

• [Perf][GDN] Cut host op calls in the eager prefill path of mixed steps

  #61034·kimseunghyun-kr opened Oct 11, 2026·4 of 6 tasks·

  ·
  1

• [Model][LoRA] Enable LoRA support for BERT and RoBERTa embedding models

  documentationpooling
  #61033·hoaivoke opened Oct 11, 2026·2 of 6 tasks·

  ·
  2

  

• [NIXL] Prepare KV transfer descriptors as NIXL 1.5 strided lists

  ci/buildkv-connectorrocm
  #61032·njhill opened Oct 10, 2026Member·

  ·
  8

• [Bugfix] Apply session.update language and prompt to Qwen3-ASR realtime

  bugdocumentationfrontendmistralqwen
  #61031·tamerrkanak opened Oct 10, 2026·6 tasks done·

  ·
  1

  2

• [Bugfix][Kernel] Mask out-of-range page-table loads in Triton unified attention

  bug
  #61028·ToukoUrsin opened Oct 10, 2026·6 tasks done·

  ·
  1

• [Bugfix][Model Runner V2] Profile the batch-sharded all-to-all before sizing the KV cache

  bugmrv2
  #61025·kimseunghyun-kr opened Oct 10, 2026·6 tasks done·

  ·
  1

• [Attention] Make the sparse MLA prefill workspace size configurable

  ready
  #61023·vMaroon opened Oct 10, 2026Contributor·6 tasks done·

  ·
  4

• [Docs] Fix broken llm-d.ai links on the llm-d integration page

  documentation
  #61022·almogtavor opened Oct 10, 2026Contributor·

  ·
  2

• [ROCm][Perf] Fuse add-RMSNorm into AITER FlyDSL INT4 QuickReduce

  rocm
  #61021·mjkvaak-amd opened Oct 10, 2026Contributor·6 tasks done·

  ·

• [ROCm][Perf] Quantize to MXFP4 in the FlyDSL QuickReduce + RMSNorm epilogue

  ci/buildquantizationrocmtorch.compile
  #61020·mjkvaak-amd opened Oct 10, 2026Contributor·6 tasks done·

  ·

• [Bugfix][KV Connector] Serve and store OffloadingConnector partial tails with EAGLE/MTP groups

  bugkv-connector
  #61019·kimseunghyun-kr opened Oct 10, 2026·4 of 6 tasks·

  ·
  1