All pull requests

Pull requests

• [None][fix] Backport OpenMPI 5 ARM64 wait-sync barriers

  #19705·peihu-nv opened Sep 29, 2026Collaborator·

  ·
  20

  

• [None][fix] Keep exit telemetry independent of blocked heartbeats

  #19704·Mgluhovskoi opened Sep 29, 2026Collaborator·

  ·
  

• [https://nvbugs/6818710][fix] Added optional helix_owned_new_tokens to the base update_helix_param…

  #19703·trtllm-agent opened Sep 29, 2026Collaborator·

  ·
  1

  

• [TRTLLM-15824][feat] Enable EVS handoff for Nemotron-H in EPD

  #19702·eopXD opened Sep 29, 2026Collaborator·

  ·
  1

  

• [TRTLLM-16672][feat] MiniMax H3 Turbo LoRA Tensor Name Mapping

  VisualGen
  #19701·yibinl-nvidia opened Sep 29, 2026Collaborator·

  ·
  7

  

• [https://nvbugs/6843918][fix] Pick the expected steady-state ngen bucket for gen_only device step time

  #19700·chenfeiz0326 opened Sep 29, 2026Collaborator·

  ·
  3

  

• [None][fix] Diagnose semantic review replies and strengthen conflict evidence

  #19699·chzblych opened Sep 29, 2026Collaborator·

  ·
  40

  

• [None][fix] Initialize cu_kv_seqlens for packed-QKV context FMHA

  Community want to contribute
  #19698·awesome-pro opened Sep 29, 2026·

  ·
  2

  

• [https://nvbugs/6840916][fix] Clean up orphaned perf test workers on failure

  #19697·yufeiwu-nv opened Sep 29, 2026Collaborator·

  ·
  5

  

• [TRTLLM-15761][refactor] Parameterize decoder invocation state and separate decoder-only engine state

  #19696·lori-ren opened Sep 29, 2026Contributor·

  ·
  

• [None][perf] Skip MoE padding in breakable prefill graphs

  #19694·longlee0622 opened Sep 29, 2026Collaborator·

  ·
  4

  

• [None][fix] Increase DeepSeek V4 FP8 perf server startup timeout

  #19692·yufeiwu-nv opened Sep 29, 2026Collaborator·

  ·
  4

  

• [None][infra] Waive 3 failed cases for main in post-merge 2990

  ci: full pre-merge approved
  #19691·trtllm-agent opened Sep 29, 2026Collaborator·

  ·
  10

  

• [None][perf] MTP draft-step glue trim (TRTLLM_MTP_TAIL_TRIM, default on; "0" = original path)

  #19690·Tabrizian opened Sep 29, 2026Member·

  ·
  21

  

• [None][perf] Trim the decode step prologue (TRTLLM_STEP_PROLOGUE_TRIM; default on, "0" = original path, "verify" = on + self-check)

  #19688·Tabrizian opened Sep 29, 2026Member·

  ·
  6

  

• [None][perf] DSA indexer: fused Triton prologue (k_norm + RoPE + FP8 Q/K quant + weight scale)

  #19685·Tabrizian opened Sep 29, 2026Member·

  ·
  9

  

• [None][chore] Align the prefix-tokenization cache switch with main (enable_tokenization_cache)

  #19683·zheyuf opened Sep 29, 2026Collaborator·

  ·
  

• [None][perf] MoE routing: small-token single-block kernel for <= 8 tokens in the trtllm-gen custom routing

  #19682·Tabrizian opened Sep 29, 2026Member·

  ·
  5

  

• [None][perf] Remove IndexK duplication with TEP for MinimaxM3

  #19681·brb-nv opened Sep 29, 2026Collaborator·

  ·
  3

  

• [None][feat] Register Mooncake KV pools per GPU pool mapping

  #19680·brb-nv opened Sep 29, 2026Collaborator·

  ·
  4

  

• [TRTLLMINF-474][infra] Upgrade NGC PyTorch base image to 26.09

  ci: full pre-merge approved
  #19679·EmmaQiaoCh opened Sep 29, 2026Collaborator·

  ·
  16

  

• [None][chore] agent-flow: opt-in --debug-file for the bundled CLI

  #19678·hyukn opened Sep 29, 2026Collaborator·

  ·
  

• [None][perf] Warp-group-local write-back in the SM90 grouped FP8 block-scale DeepGEMM kernel

  Community want to contribute
  #19676·charon14 opened Sep 29, 2026·

  ·
  1

  

• [None][test] Add coverage for TritonMXFP4LinearMethod.apply

  ci: full pre-merge approved
  #19675·StanleySun639 opened Sep 29, 2026Collaborator·

  ·
  8

  

• [None][fix] Fail stop executor worlds on unproven KV retirement

  #19674·chienchunhung opened Sep 29, 2026Collaborator·

  ·
  1