All pull requests

Pull requests

• [TRTLLM-16624][fix] Compile the Helix NCCL reformat only for CUDA-graph batches

  #21685·zhaoyuanh-nvidia opened Nov 5, 2026Collaborator·1 task done·

  ·
  

• [None][fix] causal_conv1d: form conv-state slot offsets in 64 bits

  ci: full pre-merge approved
  #21684·Wanli-Jiang opened Nov 5, 2026Collaborator·1 task done·

  ·
  5

  

• [None][test] Waive 10 failed cases for main in QA CI

  #21683·trtllm-agent opened Nov 5, 2026Collaborator·

  ·
  3

  

• [None][test] Waive 8 failed cases for main in QA CI

  #21682·trtllm-agent opened Nov 5, 2026Collaborator·

  ·
  5

  

• [None][perf] Build MSA fmha_sm100 JIT kernels for the visible GPUs' target only

  #21681·chenfeiz0326 opened Nov 5, 2026Collaborator·1 task done·

  ·
  3

  

• [https://nvbugs/7028059][fix] Fix Qwen MoE native FP8 LoRA reference test configuration

  #21680·lori-ren opened Nov 5, 2026Contributor·1 task done·

  ·
  3

  

• [None][doc] Add WinoVAE technical blog

  #21679·ishovkun opened Nov 5, 2026Contributor·1 task·

  ·
  

• [https://nvbugs/6858534][fix] Hand beam-search state to the pipeline-parallel ranks that do not sample

  Agent Code Review: High Risk
  #21678·trtllm-agent opened Nov 5, 2026Collaborator·1 task done·

  ·
  16

  

• [None][fix] agent-flow: modeling-bringup resumes with the checkpointed task.yaml

  #21677·moraxu opened Nov 5, 2026Collaborator·1 task done·

  ·
  3

  

• [TRTLLM-17267][refactor] Move the chat handlers into a serving class

  VisualGen
  #21676·JunyiXu-nv opened Nov 4, 2026Collaborator·

  ·
  

• [https://nvbugs/6992761][fix] Bound decoded video memory per request in trtllm-serve

  #21675·aswinvisva opened Nov 4, 2026Collaborator·1 task done·

  ·
  1

  

• [None][fix] Count prompt tokens without materializing the prompt list

  #21674·jh-nv opened Nov 4, 2026·1 task done·

  ·
  3

  

• [TRTLLM-17648][fix] Recover command transport and detect remote worker loss

  api-compatiblehuman-review
  #21673·tanmayv25 opened Nov 4, 2026Collaborator·3 tasks done·

  ·
  9

  

• [https://nvbugs/7003528][fix] Restore ADP connector CPU test coverage

  #21672·allisonlim-nv opened Nov 4, 2026Contributor·2 tasks done·

  ·
  13

  

• [https://nvbugs/7028062][fix] Restore modeling v2 test coverage

  #21671·peihu-nv opened Nov 4, 2026Collaborator·1 task done·

  ·
  6

  

• [None][chore] promote flashinfer-prims-ts source from #20068

  #21670·yuxianq opened Nov 4, 2026Collaborator·3 tasks done·

  ·
  5

  

• [None][feat] Add Lamport to NVLinkOneSided All2All; Retire CFT counted-write

  #21668·zhangcl opened Nov 4, 2026Collaborator·1 task done·

  ·
  12

  

• [None][fix] Chat completions honor TRTLLM_KVCACHE_TIME_OUTPUT_PATH like completions and Responses

  #21667·JunyiXu-nv opened Nov 4, 2026Collaborator·1 task done·

  ·
  

• [TRTLLM-16624][fix] Do not specialize spec-decode Triton kernels on batch-dependent arguments

  #21666·zhaoyuanh-nvidia opened Nov 4, 2026Collaborator·1 task done·

  ·
  5

  

• [None][fix] Retire timed-out disaggregated KV receives while idle

  #21659·SimengLiu-nv opened Nov 4, 2026Collaborator·1 task done·

  ·
  10

  

• [None][fix] Report disaggregated stream failures before the first chunk

  #21658·SimengLiu-nv opened Nov 4, 2026Collaborator·1 task done·

  ·
  6

  

• [None][feat] Admission-time connector offers and work-aware ADP routing

  api-compatible
  #21657·SimengLiu-nv opened Nov 4, 2026Collaborator·1 task done·

  ·
  3

  

• [None][feat] Mooncake safe shared-pool reuse and asynchronous loading

  api-compatible
  #21656·SimengLiu-nv opened Nov 4, 2026Collaborator·1 task done·

  ·
  3

  

• [None][feat] KV connector and KVCacheManagerV2 reusable-prefix runtime contracts

  api-compatible
  #21655·SimengLiu-nv opened Nov 4, 2026Collaborator·1 task done·

  ·
  3

  

• [None][fix] qwen3_coder: recover a call a finished output ends inside in whole-text parsing

  #21654·jvpflum opened Nov 4, 2026·1 task done·

  ·