All pull requests

Pull requests

• refactor(distributed): move parallelization into model sidecars

  #4032·akoumpa opened Sep 27, 2026Contributor·

  ·
  6

• fix(test): restore Step-3.7 HF PEFT checkpoint reload

  #4030·yuhezhang-ai opened Sep 26, 2026Contributor·

  ·
  2

• feat(diffusion): add Qwen-Image-2.1 text-to-image finetuning

  #4027·HuiyingLi opened Sep 26, 2026Contributor·

  ·
  5

• test(peft): close the remaining reload-suite coverage gaps

  community-request
  #4026·stanley1208 opened Sep 26, 2026Contributor·

  ·
  1

• chore: CI proxy for #2514

  #4025·akoumpa opened Sep 26, 2026Contributor·

  ·
  3

• perf(kimi-k3): add FA4 varlen backend for MLA attention

  #4023·HuiyingLi opened Sep 25, 2026Contributor·

  ·
  1

• chore: CI proxy for #3898

  #4017·HuiyingLi opened Sep 24, 2026Contributor·

  ·
  3

• fix: support native RADIO video projection in Nemotron Omni

  #4014·HuiyingLi opened Sep 24, 2026Contributor·

  ·
  2

• fix(moe): reduce peak expert MLP memory [AMINT-351]

  #4010·akoumpa opened Sep 24, 2026Contributor·

  ·
  5

• fix(distributed): prepare runtimes before pipeline traffic [AMINT-362]

  #4007·akoumpa opened Sep 24, 2026Contributor·

  ·
  4

• perf(qwen3-8-flash-next): replay QSA routes on activation-checkpoint recompute

  #4005·Agoniii opened Sep 24, 2026Collaborator·

  ·
  6

• fix(ci): accept identical short greedy outputs [AMINT-15]

  #4003·akoumpa opened Sep 23, 2026Contributor·

  ·
  3

• fix(nemotron-flash): support Transformers 5.15 layer types [AMINT-360]

  #4002·akoumpa opened Sep 23, 2026Contributor·

  ·
  3

• perf(mimo): add V2.6 FLOPs accounting and a mock benchmark

  #3995·HuiyingLi opened Sep 23, 2026Contributor·

  ·
  1

• feat(vlm): support fused linear cross entropy with pipeline parallelism

  #3994·HuiyingLi opened Sep 23, 2026Contributor·

  ·
  6

• fix(ci): normalize recipe paths and git decoding in coverage table sync

  community-requestwaiting-on-maintainers
  #3993·MohammedAlkindi opened Sep 23, 2026Contributor·

  ·
  3

• fix(qwen3.5): support MTP with VLM context parallelism [AMINT-352]

  #3990·HuiyingLi opened Sep 22, 2026Contributor·

  ·
  1

• fix(nemotron_omni): handle PEFT key conversion for adapter save and reload

  community-requestwaiting-on-customer
  #3986·taran-dev4u opened Sep 22, 2026Contributor·

  ·
  1

  6

• fix(init): support HF initialization of DTensor-sharded models

  #3983·yuhezhang-ai opened Sep 22, 2026Contributor·

  ·
  3

• fix(checkpoint): preserve nested HF key renames

  #3982·yuhezhang-ai opened Sep 22, 2026Contributor·

  ·
  2

• fix(ci): use accessible Super benchmark config

  #3980·athitten opened Sep 22, 2026Contributor·

  ·
  2

• fix(config): prefer cached files when resolving pretrained models

  #3976·janbernloehr opened Sep 22, 2026Contributor·

  ·
  1

  3

• fix: skip None modules during parameter freezing

  #3970·athitten opened Sep 21, 2026Contributor·

  ·
  1

• fix(docker): backport DeepEP NVSHMEM queue-pair compatibility

  #3968·janbernloehr opened Sep 21, 2026Contributor·

  ·
  1

• chore: CI proxy for #3942

  #3943·akoumpa opened Sep 19, 2026Contributor·

  ·
  3