All pull requests

Pull requests

• [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·

  ·
  

• [TRTLLM-15863][fix] Raise ImportError with the cause when a lazily imported serve module raises AttributeError

  #20948·YihuiLu512 opened Oct 24, 2026Collaborator·1 task done·

  ·
  4

  

• [None][fix] EagerMoeWorkspace test and TestResourceManager

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

  ·
  8

  

• [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

  #20945·0z5a opened Oct 24, 2026·1 task done·

  ·
  1

  5

  

• [https://nvbugs/6621362][fix] Cherry-pick #20459: Unblock Python disagg transfers stalled by a context-side timeout

  #20943·bo-nv opened Oct 24, 2026Collaborator·1 task done·

  ·
  2

  

• [None][perf] FP8 block-scale low-M GEMM for decode projections (SM100/103, SM107 tactic tables)

  #20942·longlee0622 opened Oct 24, 2026Collaborator·1 task·

  ·
  1

  

• [None][perf] Kimi-K3 KDA prefill grid sizing and MTP-verify kernels; DWDP cross-step expert prefetch with configurable slots

  api-compatible
  #20941·longlee0622 opened Oct 24, 2026Collaborator·1 task·

  ·
  5

  

• [None][perf] One-model spec decode: vocab-parallel token selection, fused DSpark Markov chain and spec tail, drafter TP fusion

  #20940·longlee0622 opened Oct 24, 2026Collaborator·1 task·

  ·
  5

  

• [None][perf] Helix CP: MLA V up-projection before the all-to-all, O(prompt/cp) prompt partition, FIFO all-to-all sender fence, Helix test harness fix

  #20939·longlee0622 opened Oct 24, 2026Collaborator·1 task·

  ·
  2

  

• [None][perf] Serving runtime: disagg/KV-transfer/ADP fixes, KV cache manager V2 prefix wait, ConversationRouter prefix placement, KV-balanced ADP admission

  #20938·longlee0622 opened Oct 24, 2026Collaborator·1 task·

  ·
  7

  

• [TRTLLM-16460][perf] Qwen3.8-Flash-Next NVFP4 modeling_v2 target: GB300 Pareto optimizations for ISL 8K / OSL 1K

  #20937·Wanli-Jiang opened Oct 24, 2026Collaborator·5 tasks done·

  ·
  10

  

• [None][feat] Anthropic Messages: carry the engine request id in the message id

  #20936·JunyiXu-nv opened Oct 24, 2026Collaborator·1 task done·

  ·
  4

  

• [None][perf] Pick the full-attention decode KV split per launch in FlashInfer trtllm-gen FMHA

  #20934·Tabrizian opened Oct 24, 2026Member·1 task done·

  ·
  

• [https://nvbugs/6962175][fix] Pin GPT-OSS accuracy prompt dates and remove waivers

  #20933·dongfengy opened Oct 24, 2026Collaborator·5 tasks done·

  ·
  1

  

• [None][perf] Apply vision RoPE to Q and K in one launch

  #20932·yechank-nvidia opened Oct 24, 2026Collaborator·1 task·

  ·
  1

  

• [#20930][feat] Add request-level dynamic batching to VisualGen

  VisualGen
  #20931·yanxinzhangcs opened Oct 24, 2026Contributor·4 tasks done·

  ·
  1

  3

  

• [TRTLLM-14961][refactor] Give the CuteDSL grouped-GEMM MoE implementations canonical identities

  #20929·xxi-nv opened Oct 24, 2026Collaborator·1 task done·

  ·
  3

  

• [None][perf] Speed up the post-warmup KV cache NaN/Inf scan

  #20928·thorjohnsen opened Oct 24, 2026Collaborator·1 task done·

  ·
  7

  

• [None][feat] Import Comfy NVFP4 checkpoints for FLUX and MiniMax-H3

  VisualGen
  #20926·karljang opened Oct 24, 2026Collaborator·3 of 8 tasks·

  ·
  

• [TRTLLM-17125][feat] Preempt requests on KV pressure under aggregated…

  #20925·asfiyab-nvidia opened Oct 24, 2026Contributor·1 task done·

  ·
  

• [None][feat] Serve MiniMax-M3.1-Flash with its bundled DSpark drafter

  #20922·pranav-nvidia opened Oct 23, 2026Contributor·1 task done·

  ·
  5

  

• [None][infra] remove unused CUDA12 pytorch mirror

  #20919·tburt-nv opened Oct 23, 2026Collaborator·1 task done·

  ·
  9

  

• [https://nvbugs/6943029][fix] Size Whisper's cross-KV pool from its encoder length, not max_input_len

  #20918·trtllm-agent opened Oct 23, 2026Collaborator·2 tasks done·

  ·
  1