All pull requests

Pull requests

• [None][fix] Emit valid empty arguments for Qwen3-Coder tool calls

  #20528·dcaayushd opened Oct 18, 2026·

  ·
  1

  3

  

• [TRTLLM-17016][test] Add coverage-guided test cases

  #20527·fredricz-20070104 opened Oct 18, 2026Collaborator·

  ·
  

• [None][docs] Explain Nemotron's use of the Qwen3-Coder tool parser

  #20525·eopXD opened Oct 18, 2026Collaborator·

  ·
  4

  

• [TRTLLM-17096][feat] modeling_v2: GLM-5.2 NVFP4 low-latency decode targets for GB300

  #20524·xwang233 opened Oct 18, 2026Collaborator·6 tasks done·

  ·
  1

  

• [https://nvbugs/6948011][fix] Shut down attention test KV cache managers on failure paths

  #20522·HuiGao-NV opened Oct 18, 2026Collaborator·1 task done·

  ·
  14

  

• [#20515][fix] Pass the environment's nvcc to the trtllm-gen FMHA JIT

  #20521·lucifer1004 opened Oct 18, 2026Collaborator·5 tasks done·

  ·
  1

  4

  

• [#20506][fix] Size the generation-only attention workspace without the context requests

  #20520·lucifer1004 opened Oct 18, 2026Collaborator·5 tasks done·

  ·
  1

  4

  

• [#20504][fix] Keep the breakable CUDA graph shared output allocated

  #20519·lucifer1004 opened Oct 18, 2026Collaborator·5 tasks done·

  ·
  1

  1

  

• [#20507][fix] Number Sm107a as 5 to match the prebuilt trtllm-gen archives

  #20518·lucifer1004 opened Oct 18, 2026Collaborator·5 tasks done·

  ·
  1

  2

  

• [None][feat] Staircase: GLM-5.2 NVFP4 low-latency decode target for Rubin (modeling_v2, TP16)

  #20517·dongfengy opened Oct 17, 2026Collaborator·1 task done·

  ·
  6

  

• [None][perf] KV cache V2 scheduler: opt-in shortest-uncached-first context admission

  #20516·Tabrizian opened Oct 17, 2026Member·1 task done·

  ·
  

• [#20511][fix] DeepSeek-V4.1: use BF16 CSA2 projections off SM100/SM103

  Community want to contribute
  #20512·Smallfu666 opened Oct 17, 2026·1 task done·

  ·
  1

  1

  

• [https://nvbugs/6944606][fix] Recognize Rubin GPUs in VisualGen performance tests

  VisualGen
  #20501·daichu-nv opened Oct 16, 2026Contributor·

  ·
  1

  

• [None][feat] Add initial Beam component implementation

  #20499·Mgluhovskoi opened Oct 16, 2026Collaborator·5 tasks·

  ·
  

• [None][refactor][BREAKING] Use a string enum for checkpoint I/O policy

  api-breaking
  #20498·yijingl-nvidia opened Oct 16, 2026Collaborator·1 task·

  ·
  3

  

• [#20138][fix] Preserve Qwen3-VL video layout across E/PD

  Community want to contribute
  #20497·maheshrbapatu opened Oct 16, 2026·1 task done·

  ·
  1

  1

  

• [None][fix] Apply the drafter's declared logit transform in DFlash

  #20495·moraxu opened Oct 16, 2026Collaborator·1 task·

  ·
  

• [TRTLLM-17014][infra] Migrate build artifacts from On-prem to AWS Artifactory

  #20494·sharabeshj-nvda opened Oct 16, 2026Collaborator·1 task done·

  ·
  29

  

• [https://nvbugs/6942724][fix] Refresh serving configuration schema snapshots

  #20493·achartier opened Oct 16, 2026Collaborator·2 tasks done·

  ·
  24

  

• [None][perf] Find the common tool argument prefix with slice comparisons

  Community want to contribute
  #20492·deadstrobe5 opened Oct 16, 2026·1 task done·

  ·
  1

  

• [TRTLLM-17083][feat] Cancel paired GEN on CTX failure or cancellation

  #20491·chienchunhung opened Oct 16, 2026Collaborator·

  ·
  4

  

• [None][fix] Let --disagg_cluster_uri override the config file uri

  Community want to contribute
  #20490·spicysharkco opened Oct 16, 2026·1 task done·

  ·
  1

  

• [#20485][fix] Fix NGramPoolManager candidate pruning and single-match invariant

  Community want to contribute
  #20487·kunalsanga opened Oct 16, 2026·3 of 4 tasks·

  ·
  1

  1

  

• [TRTLLM-20434][feat] Dflash Unified KVCache

  #20486·allisonlim-nv opened Oct 16, 2026Contributor·1 task done·

  ·
  21

  

• [None][feat] Visual Gen token-sharded TP: gather FP8 block-scale inputs

  VisualGen
  #20483·chang-l opened Oct 16, 2026Collaborator·1 task done·

  ·
  6