All pull requests

Pull requests

• [Bugfix] Remove 100-connection httpx pool limit in disagg proxy server

  #3842·bhuvanpkaruturi opened Oct 17, 2026Collaborator·3 tasks done·

  ·
  1

  1

• [Benchmark] Qwen3.8-2.4T daily: C=384, fold routing into dispatch, skip mamba scheduler blocks

  ready
  #3841·lxhfirenking opened Oct 17, 2026Collaborator·3 tasks done·

  ·
  2

• [Pathways] Transfer real weights to TPU in their native dtype

  #3840·angelakewang opened Oct 17, 2026·

  ·

• [Bugfix][P/D] Fix decode hang when KV pull fails; lift proxy connecti…

  ready
  #3839·charlesli640 opened Oct 16, 2026Collaborator·3 tasks done·

  ·
  1

• [Bugfix] Compute logprobs when zero alternatives are requested

  #3838·pujitha24 opened Oct 16, 2026·

  ·
  1

• [Pluggable Attention Backend] Add --attention-backend flywheel

  #3837·lk-chen opened Oct 16, 2026Collaborator·

  ·
  1

  25

• [Bugfix] Don't treat a timed-out KV pull as received in the TPU P/D connector

  #3836·ksapru opened Oct 16, 2026·

  ·

• [CI] Release tooling: vllm/vllm-tpu release images and a release s…

  #3834·Jeff-Tseng-dev opened Oct 16, 2026Collaborator·2 of 3 tasks·

  ·
  1

• Use sc_info.vmem_capacity_bytes for SparseCore VMEM capacity

  #3826·pritha90 opened Oct 15, 2026Contributor·

  ·

• feat(gemma4): hybrid sliding-window KV cache for mixed head dims, XLA HBM headroom, and RPA v3 window clamping

  #3819·lucianommartins opened Oct 14, 2026·

  ·
  10

• Support continue decode for gdn models

  ready
  #3815·pv97 opened Oct 14, 2026Collaborator·

  ·

• [Fix] Fix Qwen2.5-VL vision precompile dtype and eliminate eager JAX overhead in multimodal path

  #3814·damadei-g opened Oct 13, 2026·3 tasks done·

  ·

• [GDN] Run the fused conv1d+GDN kernel on the touched conv-state slots only

  #3812·wenxindongwork opened Oct 13, 2026Collaborator·

  ·
  1

• [MoE] Let fused_moe_func fold an external partial sum into its all-reduce

  #3811·wenxindongwork opened Oct 13, 2026Collaborator·

  ·
  1

• [CI] Publish the nightly image from a Python release bootstrap [skip ci]

  #3809·theminghuang opened Oct 13, 2026Collaborator·

  ·

• [Fix] Apply temperature before filtering in rejection sampler

  #3808·junuxyz opened Oct 13, 2026·3 tasks done·

  ·

• [MoE] Wrap router_topk in jax.shard_map when mesh is provided

  #3803·andytwigg opened Oct 12, 2026Collaborator·

  ·
  1

• [DCP] Support 256-token prefix cache hit granularity under DCP via CoW block copying

  #3796·wenxindongwork opened Oct 11, 2026Collaborator·

  ·

• [RPA v3] Cap the decode KV block at the sliding window

  #3786·jerviscz opened Oct 10, 2026Contributor·3 tasks done·

  ·
  1

• fix: [Bug]: concurrent requests with structured inputs fail and terminate wit...

  #3784·Roshan931 opened Oct 10, 2026·

  ·
  1

• [Runner] Keep token ids for spec decode on multimodal steps

  #3777·Ishannaik opened Oct 9, 2026·3 tasks done·

  ·

• [MLPerf][MoE] Port Attention-to-MoE collective overlap fusion

  #3774·wenxindongwork opened Oct 8, 2026Collaborator·

  ·
  1

• [BugFix] Update test_routed_experts_slots according to vllm PR#59197

  oncall-fixready
  #3770·charlesli640 opened Oct 8, 2026Collaborator·

  ·
  1

• [Benchmarking] Add interactive trajectory and turn waterfall visualizer for agentic RL

  ready
  #3767·wenxindongwork opened Oct 8, 2026Collaborator·

  ·
  1

• [Benchmarking] Support staggered batch scheduling and trajectory metrics in agentic benchmark

  ready
  #3766·wenxindongwork opened Oct 8, 2026Collaborator·

  ·
  1