All pull requests

Pull requests

• perf(qwen3.5): fuse the input add + Gemma RMSNorm with the static-FP8 quant

  #2020·nperrin-fr opened Oct 6, 2026Collaborator·

  ·

• feat(cli): leave the gateway parsers to a declared response_template

  #2019·yechank-nvidia opened Oct 6, 2026Collaborator·6 of 7 tasks·

  ·

• refactor(comm): use upstream FlashInfer HT for K3

  #2018·Njuapp opened Oct 6, 2026Contributor·

  ·

• Fix CuTe DSL warnings in runtime and kernel tests

  #2017·lightseek-bot opened Oct 6, 2026Contributor·

  ·
  5

• fix(grammar): admit grammar requests in lockstep across attention TP ranks

  #2015·yechank-nvidia opened Oct 6, 2026Collaborator·5 tasks done·

  ·
  1

• refactor(scheduler): classify cache groups by AttnKind

  #2011·dongjiyingdjy opened Oct 6, 2026Contributor·

  ·
  2

• refactor(kernel): remove top-level operator exports

  #2004·borontion opened Oct 5, 2026Contributor·

  ·
  1

• feat(cache): add sparse KV offloading with fixed GPU hot pools and cross-layer prefetch

  #1950·cnStevenYu opened Oct 3, 2026·

  ·
  1

• perf(amd): tune packed prefill with upstream gfx1250 controls

  #1941·nithinsubbiah opened Oct 3, 2026Member·

  ·

• perf(memory): reserve the startup residue, restore the tuning stack limit, drop the multimodal 0.9 clamp

  #1940·nperrin-fr opened Oct 2, 2026Collaborator·

  ·
  1

• feat(runtime): emulate rank 0 of a parallel layout on one GPU

  #1938·Yu-Zhewen opened Oct 2, 2026Contributor·

  ·
  12

• fix(grammar): answer rejected constrained requests with 400

  #1933·yechank-nvidia opened Oct 2, 2026Collaborator·5 tasks done·

  ·
  1

• perf(gemm): add FP32 row-CTA decode GEMV leaves for small-M projections

  #1931·yechank-nvidia opened Oct 2, 2026Collaborator·

  ·

• fix(entrypoints): stop schedulers of a non-zero-rank node with the node

  #1921·yechank-nvidia opened Oct 1, 2026Collaborator·3 of 4 tasks·

  ·

• fix(kernel): keep the caller's shape of TRT-LLM per-token FP8 scales

  #1920·yechank-nvidia opened Oct 1, 2026Collaborator·4 of 5 tasks·

  ·

• fix(moe): gather Triton BF16 MoE rows by pointer below sm_100

  #1917·yechank-nvidia opened Oct 1, 2026Collaborator·4 of 5 tasks·

  ·

• feat(moe): opt-in FP32 correction bias for TRT-LLM DeepSeekV3 routing

  #1891·yechank-nvidia opened Sep 30, 2026Collaborator·

  ·
  1

• feat(dsa): support Hopper FP8 KV with FlashMLA

  #1887·qimcis opened Sep 30, 2026Contributor·

  ·

• feat(runtime): add Score API for decision-style typed output

  #1880·stmatengss opened Sep 29, 2026Contributor·

  ·
  14

• feat(moe): run TRT-LLM BF16 MoE at 64-aligned intermediate sizes

  #1878·yechank-nvidia opened Sep 29, 2026Collaborator·

  ·
  1

• Perf(dsa): skip empty Triton indexer scoring tiles in dcp

  #1877·JamesLim-sy opened Sep 29, 2026Contributor·

  ·
  4

  

• fix(rl): dispatch update_weights_from_disk in the scheduler

  #1874·vivekkhandelwal1 opened Sep 29, 2026Contributor·

  ·
  2

• fix(scheduler): pass a promotion boundary the chunk budget cannot reach

  #1861·yechank-nvidia opened Sep 29, 2026Collaborator·

  ·
  

• fix(scheduler): admit a remote prefill as the whole prompt

  #1860·yechank-nvidia opened Sep 29, 2026Collaborator·

  ·
  

• perf(dsv41): fuse block maxima and bound candidate selection

  #1857·shineran96 opened Sep 29, 2026Contributor·

  ·
  2