All pull requests

Pull requests

• [TRTLLM-17360][feat] Add Beam StairCase production components

  #21352·Mgluhovskoi opened Oct 30, 2026Collaborator·

  ·
  

• [TRTLLM-17356][feat] Consolidate Beam serving and first-100 QA regression

  #21351·Mgluhovskoi opened Oct 30, 2026Collaborator·

  ·
  

• [TRTLLM-16282][feat] Integrate sparse KV offload for DeepSeek-V4 Flash and GLM5.2 on KVCM2

  #21350·cascade812 opened Oct 30, 2026Collaborator·1 task·

  ·
  

• [None][test] Run session-reuse unit tests in CPU CI

  #21349·aswinvisva opened Oct 30, 2026Collaborator·1 task done·

  ·
  3

  

• [None][feat] Add LLM API example to measure startup time

  #21348·yijingl-nvidia opened Oct 30, 2026Collaborator·1 task·

  ·
  

• [https://nvbugs/6908161][fix] Bootstrap NCCL EP from the torch EP group under Ray

  #21347·nv-lschneider opened Oct 30, 2026Collaborator·1 task done·

  ·
  1

  1

  

• [None][fix] Qwen3.8-Flash-Next: mask FP8 PLE rows by bytes under attention DP

  #21346·Muchuan-Jin opened Oct 30, 2026·3 of 4 tasks·

  ·
  2

  

• [None][fix] Isolate modeling_v2 tests and scope shared test state

  #21345·litaotju opened Oct 30, 2026Collaborator·1 task done·

  ·
  11

  

• [None][fix] Warmup: no generation shapes on a context-only KV cache

  #21343·longlee0622 opened Oct 30, 2026Collaborator·

  ·
  1

  

• [None][fix] KDA Small-BH prefill: 64-bit state-pool row stride

  #21342·longlee0622 opened Oct 30, 2026Collaborator·

  ·
  2

  

• [#21318][fix] Drop streamed stop-word ids from the final streamed text

  #21341·spped2000 opened Oct 30, 2026·1 task done·

  ·
  1

  1

  

• [#21317][fix] trtllm-bench: fall back to system memory when NVML memory query is not supported

  #21340·spped2000 opened Oct 30, 2026·1 task done·

  ·
  1

  1

  

• [None][chore] Reject per-layer KV-head configs in the generic HF weight mapper

  #21339·lfr-0531 opened Oct 30, 2026Collaborator·1 task done·

  ·
  1

  

• [https://nvbugs/6226287][fix] Pick EOS handling from speculative decoding in trtllm-bench and benchmark_serving

  #21338·trtllm-agent opened Oct 30, 2026Collaborator·

  ·
  9

  

• [None][refactor] Move PrimTS attention cache updates into upstream source

  #21337·yuxianq opened Oct 30, 2026Collaborator·1 task done·

  ·
  5

  

• [TRTLLM-17515][feat] modeling_v2: GLM-5.3-Flash NVFP4 target with one-model MTP in the forward

  #21335·zhaoyangwang-nvidia opened Oct 30, 2026Collaborator·1 task done·

  ·
  

• [None][feat] GEN slot preemption: pause decode requests in place to admit at-risk waiters

  #21334·reasonsolo opened Oct 30, 2026Collaborator·1 task done·

  ·
  

• [None][perf] Enable breakable prefill CUDA graphs for multimodal requests

  #21333·yechank-nvidia opened Oct 30, 2026Collaborator·1 task done·

  ·
  

• [https://nvbugs/6910975][fix] Keep GDN decodes out of a bounded prefill state workspace

  #21332·chenfeiz0326 opened Oct 30, 2026Collaborator·1 task done·

  ·
  17

  

• [None][test] Add DGX Spark agentic and real-text speculative decoding perf cases

  #21331·JennyLiu-nv opened Oct 30, 2026Collaborator·1 task done·

  ·
  4

  

• Shawnwan/fp4 mla 4x

  #21330·lambw opened Oct 30, 2026·1 task·

  ·
  6

  

• DO NOT REVIEW

  #21329·bo-nv opened Oct 30, 2026Collaborator·1 task·

  ·
  7

  

• [None][fix] Preserve no-cache attention prompt lengths during async copies

  #21328·yechank-nvidia opened Oct 30, 2026Collaborator·1 task done·

  ·
  

• [None][infra] unify GB300 pre-merge and post-merge test lists

  ci: full pre-merge approved
  #21327·ZhanruiSunCh opened Oct 30, 2026Collaborator·1 task done·

  ·
  5

  

• [None][infra] improve CBTS hits for formatting and VisualGen changes

  #21326·crazydemo opened Oct 30, 2026Collaborator·

  ·
  8