All pull requests

Pull requests

• perf(gtp): plan head and MTP gathers per use and drain reduce-scatters in issue order

  complexity: medium
  #9237·JF-D opened Nov 9, 2026Contributor·5 of 7 tasks·

  ·
  4

  

• Fail a decode handoff on arrival when it can never fit

  #9236·lmcafee-nvidia opened Nov 9, 2026Contributor·

  ·
  1

• Forget an aborted request when its decode reply arrives

  #9235·lmcafee-nvidia opened Nov 9, 2026Contributor·

  ·
  1

• Keep the MTP draft-KV commit pass on each request's own rows after async compaction

  #9234·lmcafee-nvidia opened Nov 9, 2026Contributor·

  ·
  2

• docs: highlight Blackwell performance gains and Q4 roadmap

  #9233·sbhavani opened Nov 9, 2026Contributor·2 of 6 tasks·

  ·
  1

• Speed up indexed dataset pointer construction

  community-request
  #9201·Hang-get opened Nov 9, 2026·4 of 7 tasks·

  ·
  1

  1

• Pairwise testing batch — 2026-11-09

  Batch PRP3
  #9198·lmcafee-nvidia opened Nov 9, 2026Contributor·

  ·
  3

• fix(optimizer): take the LayerWise all-gather's device/dtype from the first non-empty owner list

  #9133·linmuchuiyang opened Nov 9, 2026·4 of 7 tasks·

  ·
  1

  2

• [DRAFT] Support batch invariant for simplified DSA GQA train/inference

  #9128·xuantengh opened Nov 9, 2026Contributor·7 tasks·

  ·
  1

• feat(moe): support routing replay with global-batch quantile balancing

  #9127·RayenTian opened Nov 9, 2026·1 of 7 tasks·

  ·
  5

• Fix packed CSA compressor parameter readiness

  deepseekv4
  #9126·FDecaYed opened Nov 9, 2026Contributor·

  ·
  1

• refactor(training): finish GPT data-length runtime consumers

  #9125·cuichenx opened Nov 9, 2026Contributor·2 of 7 tasks·

  ·
  1

• fix(gtp): stop eager reduce-scatter tickets from sharing a pending ou…

  complexity: low
  #9124·JF-D opened Nov 9, 2026Contributor·5 of 7 tasks·

  ·
  3

  

• fix(inference): reserve recurrent-state dummy slot when tp_ep > 1

  complexity: low
  #9123·lauradang opened Nov 9, 2026Contributor·

  ·
  9

  

• [MFSDP] Replace gradient synchronization flags with three modes

  complexity: mediumRun NeMoRL tests
  #9122·wujingyue opened Nov 9, 2026Contributor·4 tasks done·

  ·
  1

• fix(inference): check free Mamba slots before admitting any hybrid request

  complexity: lowFinal Review
  #9121·lauradang opened Nov 9, 2026Contributor·

  ·
  4

  

• Fix disaggregated serving for hybrid MoE models with MTP

  #9120·santhnm2 opened Nov 9, 2026Contributor·

  ·
  1

• docs: fix duplicated word in speculative decoding guide

  community-request
  #9119·ddlw opened Nov 8, 2026·

  ·
  1

  2

• Scale TorchSampling logits in fp32 before filtering

  community-request
  #9117·suhas018 opened Nov 8, 2026·4 of 7 tasks·

  ·
  1

  2

• Support custom layers in HybridModel

  complexity: medium
  #9116·Phlip79 opened Nov 8, 2026Member·1 task done·

  ·
  6

• Keep MFSDPv1 gradient buckets until their reduction finishes

  Approvedcomplexity: low
  #9114·XiaodaNV opened Nov 8, 2026Contributor·1 task done·

  ·
  6

• [Dev] Integrate Megatron FSDP v2 with partial CUDA Graphs

  community-request
  #9112·WEIXILIN0307 opened Nov 8, 2026·3 of 6 tasks·

  ·
  1

• chore: AUT-2775 nightly sync main into dev (2026-11-06)

  determinism-exemptDisable selective unit testsRun CICDRun full unit testsRun functional testsRun MBridge testsRun NeMoRL testsRun tests
  #9111·svcnemo-autobot opened Nov 8, 2026Collaborator·

  ·
  6

• [dev][kimi-k3] Fix: Split the KDA unit tests out of the broad ssm CI bucket

  #9110·yuzhongw-nvidia opened Nov 8, 2026Contributor·

  ·
  1

• feat(moe): add optional ep-invariant expert combine

  complexity: high
  #9107·ko3n1g opened Nov 7, 2026Contributor·

  ·
  5