All pull requests

Pull requests

• [None][fix] Cherry-pick #20393: save logs for stress tests

  #20979·bo-nv opened Oct 26, 2026Collaborator·1 task done·

  ·
  5

  

• [https://nvbugs/6765807][test] Unwaive deepseek_r1_v2_fp4 disagg stress test on B200

  ci: full pre-merge approved
  #20978·reasonsolo opened Oct 26, 2026Collaborator·1 task done·

  ·
  4

  

• [None][feat] configure input processor workers

  #20977·yechank-nvidia opened Oct 26, 2026Collaborator·1 task·

  ·
  

• [None][fix] install CBTS CLI dependencies before selection

  #20976·crazydemo opened Oct 26, 2026Collaborator·

  ·
  14

  

• [None][feat] Rows name their slots and regions in the backend contract

  #20975·Shixiaowei02 opened Oct 26, 2026Collaborator·1 task done·

  ·
  3

  

• [None][perf] Reduce padding in multimodal encoder CUDA graphs

  #20974·yechank-nvidia opened Oct 26, 2026Collaborator·1 task·

  ·
  

• [None][perf] Normalize Qwen3-VL uint8 image patches on the GPU

  #20973·yechank-nvidia opened Oct 26, 2026Collaborator·1 task·

  ·
  

• [None][refactor] Route per-step device copies through a model-declared copy plan

  #20972·WeiHaocheng opened Oct 26, 2026Collaborator·1 task·

  ·
  

• [TRTLLM-15944][feat] Enable SWA scratch reuse for Gemma3/4 with TRTLLM attention

  #20971·eopXD opened Oct 25, 2026Collaborator·

  ·
  7

  

• [https://nvbugs/6826000][fix] Size the DeepSeek-V4-Pro AgentX ctx KV pools for the compressed history

  #20970·chenfeiz0326 opened Oct 25, 2026Collaborator·1 task done·

  ·
  1

  

• [#20967][fix] Map the KVCM2 host-tier reservation PROT_NONE

  Community want to contribute
  #20968·Smallfu666 opened Oct 25, 2026·1 task done·

  ·
  1

  1

  

• [#20965][fix] Marlin MoE: take the gate from the second half of FC1

  Community want to contribute
  #20966·Smallfu666 opened Oct 25, 2026·1 task done·

  ·
  1

  1

  

• [None][test] Remove 75 closed-bug waive entries for main

  #20964·trtllm-agent opened Oct 25, 2026Collaborator·

  ·
  

• [#20962][fix] Accept the OpenAI safety_identifier field in Chat Completions requests

  Community want to contribute
  #20963·Sonica-B opened Oct 25, 2026·1 task done·

  ·
  1

  1

  

• [None][perf] AgentX: recover KV prefix reuse for DeepSeek-V4.1-Flash, Kimi-K3 and GLM-5.3-Flash

  #20961·chenfeiz0326 opened Oct 25, 2026Collaborator·1 task done·

  ·
  7

  

• [#19861][fix] Fix FP8 KV cache for cross attention on the PyTorch backend

  Community want to contribute
  #20958·lminer opened Oct 25, 2026·3 of 4 tasks·

  ·
  1

  6

  

• [#20955][fix] MiniMax-M2 tool parser: stop name patterns at the first >

  Community want to contribute
  #20957·LeizhenZhang opened Oct 25, 2026·1 task done·

  ·
  1

  4

  

• [#20954][fix] Qwen3 tool parser: keep plain answers that open with [ or { as content

  Community want to contribute
  #20956·LeizhenZhang opened Oct 25, 2026·1 task done·

  ·
  1

  9

  

• [None][feat] Extend Beam ModelingV2 inference and serving

  #20952·Mgluhovskoi opened Oct 25, 2026Collaborator·2 of 3 tasks·

  ·
  

• [#15178][fix] Account for reclaimable memory in V2 KV cache sizing on integrated GPUs

  #20951·janbernloehr opened Oct 24, 2026Collaborator·1 task done·

  ·
  4

  

• [None][fix] serialize CPU tensor sharing strategy changes

  #20950·yechank-nvidia opened Oct 24, 2026Collaborator·1 task done·

  ·
  

• [None][fix] preserve multimodal state across context and generation workers

  #20949·yechank-nvidia opened Oct 24, 2026Collaborator·1 task done·

  ·
  

• [None][fix] EagerMoeWorkspace test and LoRA tests

  #20947·Funatiq opened Oct 24, 2026Collaborator·1 task done·

  ·
  16

  

• [None][chore] Add a code-filtered mypy check for tensorrt_llm and fix the errors it found

  VisualGen
  #20946·HuiGao-NV opened Oct 24, 2026Collaborator·1 task·

  ·
  13

  

• [#20515][fix] Pass the selected nvcc path to the trtllm-gen FMHA JIT

  Community want to contribute
  #20945·0z5a opened Oct 24, 2026·1 task done·

  ·
  1

  5