All pull requests

Pull requests

• fix(sampling): make per-request single-token draws independent of prefill chunking

  #2380·sjqgogogogo opened Oct 15, 2026Contributor·

  ·

• fix(runtime): drain async work before closing the sync engine

  #2379·LYMDLUT opened Oct 15, 2026·

  ·

• fix(dsv41): derive image routing masks from multimodal spans

  #2378·chenht2022 opened Oct 15, 2026Member·

  ·
  3

• perf(dsv41): Reuse indexer metadata across attention layers

  #2377·chenht2022 opened Oct 15, 2026Member·3/3·

  ·
  2

• perf(qsa): reuse shared prefixes in CuTe decode

  #2373·tuanzhangCS opened Oct 15, 2026Contributor·

  ·
  2

• perf(npu): streamline DSA packed-query and value layouts

  #2371·zhuhaozhecool opened Oct 15, 2026·

  ·

• perf(npu): cache DSA selection-window constants

  #2370·zhuhaozhecool opened Oct 15, 2026·

  ·

• perf(npu): fuse DSA indexer RoPE

  #2369·zhuhaozhecool opened Oct 15, 2026·

  ·

• fix(cache): charge KVP tail copies to their source bucket at admission

  #2366·dongjiyingdjy opened Oct 15, 2026Contributor·

  ·
  2

• perf(sampling): reduce greedy argmax across vocab shards

  #2364·antiagainst opened Oct 15, 2026Member·

  ·
  7

• fix(l2): bound the Host memory share by the cgroup limit

  #2351·MartinHua opened Oct 14, 2026Contributor·

  ·

• Upgrade FlashInfer to 0.7.1

  #2338·lightseek-bot opened Oct 14, 2026Contributor·

  ·
  2

• perf(moe): write K3 NVFP4 inputs directly into the MegaMoE workspace

  #2337·Njuapp opened Oct 14, 2026Contributor·

  ·

• feat(comm): Iris across MI455X trays in one fabric domain

  #2335·drprajap opened Oct 14, 2026·

  ·
  5

• feat(serve): [WIP] add direct SMG ZMQ serving

  #2328·slin1237 opened Oct 14, 2026Contributor·

  ·
  6

• refactor(iris): select generic protocol codegen per architecture

  #2312·drprajap opened Oct 13, 2026·

  ·
  13

• gemm: add bounded H20 BF16 decode candidate

  #2304·LYMDLUT opened Oct 13, 2026·

  ·

• fix(fp8): restore cached MXFP8 scales for DSV41

  #2292·chenht2022 opened Oct 13, 2026Member·

  ·
  8

• perf(dsv41): Optimize sparse indexer scoring and Top-K preparation

  #2290·chenht2022 opened Oct 13, 2026Member·2/3·

  ·
  5

• perf(moe): reduce MegaMoE workspace and enable it for K3 EP16

  #2285·Njuapp opened Oct 13, 2026Contributor·

  ·
  6

• perf(qsa): bound score work and add length-aware top-k

  #2283·tuanzhangCS opened Oct 13, 2026Contributor·

  ·
  13

• Add Kimi K3 AgentX benchmarks with native InferenceX

  #2282·Xiangyi1996 opened Oct 13, 2026Collaborator·

  ·
  14

• feat(dsv41): context parallelism with QCP, KVP and DCP

  #2274·shineran96 opened Oct 13, 2026Contributor·3 of 9 tasks·

  ·

• feat(iris): K3 Gluon kernels on gfx1250

  #2273·drprajap opened Oct 13, 2026·

  ·

• feat(iris): Lamport all-reduce on CDNA5 TP2/TP4/TP8

  #2271·drprajap opened Oct 13, 2026·

  ·