All pull requests

Pull requests

• [Model] Nemotron-3.5-Transcribe: support ReplaySSM and the MTP drafter on Model Runner V1

  documentationspeculative-decoding
  #65520·DongjiGao opened Nov 3, 2026Contributor·6 tasks done·

  ·
  2

• [Perf][KV Offload] Bound copy descriptors submitted per step for CPU->GPU loads

  nixl
  #65519·zetxqx opened Nov 3, 2026·6 tasks done·

  ·
  1

  1

• [Build] Bump vllm-flash-attn to compress fatbins with --compress-mode=size

  ci/build
  #65518·mgoin opened Nov 3, 2026Member·6 tasks done·

  ·
  2

• [Bugfix][ROCm][MoE] Fix tkw1 metadata check, rename AITER QuantMethod

  bugdeepseekDSv4quantizationrocm
  #65517·divakar-amd opened Nov 3, 2026Contributor·6 tasks done·

  ·

• [Attention/TopK] Compact hit_mask candidates via __ffs and BLSR in sampled_topk

  #65516·hianova opened Nov 3, 2026·

  ·
  1

• [CI] Run PyTorch Compilation root tests in one pytest, with test_config.py's forked tests spawned

  ci/buildtorch.compile
  #65514·Thangnguyenvn98 opened Nov 3, 2026Contributor·2 of 6 tasks·

  ·
  2

• [WIP][ROCm][Perf][DeepSeek-V4.1] Mega-attention prefill kernel for gfx942

  ci/builddeepseekdocumentationDSv4.1needs-rebasenvidiaperformancerocm
  #65512·frida-andersson opened Nov 3, 2026Contributor·1 of 10 tasks·

  ·
  2

• [Build][CUDA] Raise CUDA compiler version gates for cooperative_topk (12.8) and DeepGEMM (12.9)

  ci/buildnvidia
  #65511·marlonferreira1800-hue opened Nov 3, 2026·2 tasks done·

  ·
  1

• [Mamba] Store ReplaySSM ring trackers in the Mamba page; put the dt ring last

  ci/buildcpudocumentationmrv2scheduler
  #65510·tomeras91 opened Nov 3, 2026Member·6 tasks done·

  ·
  1

• [ROCm][Perf] Capture the routed MoE chain before the shared-expert aux branch for small decode batches

  rocm
  #65509·Fangzhou-Ai opened Nov 3, 2026Collaborator·

  ·
  1

• [ROCm][Perf][MiniMax-M3.1] Emit the o_proj MXFP8 input from the NVFP4 sparse decode kernel

  minimaxreadyrocm
  #65508·Fangzhou-Ai opened Nov 3, 2026Collaborator·

  ·
  2

• [KVConnector] Enable LMCache MP connector for the extensible kv cache

  kv-connectorready-run-all-tests
  #65507·ApostaC opened Nov 3, 2026Contributor·4 of 6 tasks·

  ·

• [Bugfix][Frontend] Stop API workers and coordinator on parent loss

  bugci/builddocumentation
  #65506·soma1902 opened Nov 3, 2026·6 tasks done·

  ·
  2

• [Perf][Kernel][Attention] Binary-search mm_prefix spans in Triton unified attention

  performance
  #65505·Noezor opened Nov 3, 2026·

  ·
  3

• [Step5] Allow MegaMoE with pipeline parallelism

  #65504·xsxszab opened Nov 3, 2026Contributor·6 tasks done·

  ·

• [Bugfix][KV Connector][Mooncake] Fix transfer address overflow on Intel XPU

  bugintel-gpukv-connectormooncake
  #65501·slokesha opened Nov 3, 2026·5 of 6 tasks·

  ·
  1

• [ROCm][Perf] Try AITER attention backends before ROCM_ATTN (limited to MI300)

  ci/buildrocm
  #65500·alexm-redhat opened Nov 3, 2026Contributor·

  ·
  3

  

• [Benchmark] Add synthetic audio support to random-mm

  performance
  #65499·sahirema opened Nov 3, 2026Contributor·2 tasks done·

  ·
  1

• [Perf][MoE] Skip the combine staging copy in flashinfer_nvlink_one_sided

  nvidiaready
  #65498·vadiklyutiy opened Nov 3, 2026Member·6 tasks done·

  ·
  2

• [Perf][Kernel] Fuse RMSNorm + block FP8 quant for fp32 (Gemma-style) norm weights

  quantizationtorch.compile
  #65497·cheese-cakee opened Nov 3, 2026Contributor·5 of 6 tasks·

  ·

• Update DeepGEMM pin

  ci/build
  #65494·qqaatw opened Nov 3, 2026·6 tasks·

  ·
  5

• [Benchmark] Add --random-mm-video-encode-fps to random-mm

  performance
  #65493·sahirema opened Nov 3, 2026Contributor·2 tasks done·

  ·
  1

• [Perf] Pre-warm float32 greedy-with-penalties sampler and GDN RMSNorm buckets

  mrv2qwen
  #65492·arashb opened Nov 3, 2026·

  ·
  1

• [Bugfix] Omit inactive CP checkpointing kwargs in FlashInfer GDN prefill

  bugnvidia
  #65491·arashb opened Nov 3, 2026·

  ·
  1

• [Bugfix][Perf] Exclude draft-only buckets from min_num_layers in KV cache grouping

  bugkv-cache-manager
  #65490·arashb opened Nov 3, 2026·

  ·
  1