All pull requests

Pull requests

• Skip the KPool FP8 Hadamard reference test without fast_hadamard_transform

  community-request
  #8806·0z5a opened Oct 31, 2026·1 task done·

  ·
  1

• Skip unfused DSA parity cases that do not fit in GPU memory

  community-request
  #8805·0z5a opened Oct 31, 2026·1 task done·

  ·
  1

• Handle an empty local expert assignment in fused permute and unpermute

  community-request
  #8804·0z5a opened Oct 31, 2026·1 task done·

  ·
  1

  1

• Enable MFSDP v2 recompute weight planes in the MCore adapter

  #8803·shjwudp opened Oct 31, 2026Contributor·

  ·
  1

• ci: validate Testmon selection for a workflow-only change (do not merge)

  Run selective unit tests
  #8801·balasaajay opened Oct 31, 2026Contributor·

  ·
  2

• Add shared-parameter looped decoder execution for training

  community-request
  #8800·Levius-Fubuki opened Oct 31, 2026·

  ·
  1

  2

• Support multiple MTP layers with EP A2A overlap

  complexity: low
  #8799·guihong-nv opened Oct 31, 2026Contributor·2 of 7 tasks·

  ·
  6

• [refactor][training] Migrate rerun and straggler configuration

  #8798·cuichenx opened Oct 30, 2026Contributor·3 of 7 tasks·

  ·
  1

• refactor(training): remove redundant model type arguments

  complexity: medium
  #8797·cuichenx opened Oct 30, 2026Contributor·5 of 7 tasks·

  ·
  1

• refactor(training): centralize cross-config validation and finalization

  complexity: medium
  #8796·cuichenx opened Oct 30, 2026Contributor·3 of 7 tasks·

  ·
  2

• Fix decoding of configured byte-level special tokens

  community-request
  #8795·deavgprime opened Oct 30, 2026·4 of 5 tasks·

  ·
  1

  1

• Update fp8 default

  #8794·maanug-nv opened Oct 30, 2026Contributor·7 tasks·

  ·
  1

• build: update Transformer Engine to release_v2.21 tip on core_r0.20.0

  Disable selective unit testsRun functional tests
  #8792·balasaajay opened Oct 30, 2026Contributor·

  ·
  5

• Move portable unit-test coverage to four-GPU GB200 runners

  Disable selective unit tests
  #8791·Phlip79 opened Oct 30, 2026Member·1 task·

  ·
  3

• Start a chunked prefill only when its whole prompt fits

  #8790·lmcafee-nvidia opened Oct 30, 2026Contributor·

  ·
  2

• Handle aborts while a RECOMPUTE suspend has freed request state

  #8789·lmcafee-nvidia opened Oct 30, 2026Contributor·

  ·
  2

• Megatron-FSDP: compute params norm without resetting in-flight param all-gathers

  complexity: lowdeterminism-exempt
  #8788·rapatel opened Oct 30, 2026Contributor·4 of 7 tasks·

  ·
  8

  

• Megatron-FSDP: end the pipelined HFSDP prefetch window on an FSDP unit boundary.

  complexity: low
  #8787·rapatel opened Oct 30, 2026Contributor·4 of 7 tasks·

  ·
  10

  

• [DCPxMIMO] Disagg MIMO Encoder with Dynamic Context Parallel

  #8786·zhongbozhu opened Oct 30, 2026Contributor·7 tasks·

  ·
  1

• Make all HybridStackSubmodules specs optional

  complexity: low
  #8785·janEbert opened Oct 30, 2026Contributor·1 task done·

  ·
  4

• Fix mHC unit boundaries for hook-driven MFSDP v2

  #8784·shjwudp opened Oct 30, 2026Contributor·

  ·
  1

• [DCPxMIMO] Colocate MIMO Encoder with Dynamic Context Parallel

  #8783·zhongbozhu opened Oct 30, 2026Contributor·7 tasks·

  ·
  1

• Split/mfsdp muon step latency

  #8782·Autumn1998 opened Oct 30, 2026Contributor·7 tasks·

  ·
  2

• [CI] Limit test queue to two active PRs per user per branch

  #8781·balasaajay opened Oct 30, 2026Contributor·1 task·

  ·
  1

• Fix/preprocess data split sentences

  community-request
  #8780·Anshtyagi1729 opened Oct 30, 2026·3 of 7 tasks·

  ·
  1