All pull requests

Pull requests

• fix(transformers): pin config and weights to the resolved Hub snapshot

  #4019·akoumpa opened Sep 25, 2026Contributor·

  ·
  1

  8

• 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

• docs: update Super 3.5 VL checkpoint names for GA

  #4013·HuiyingLi opened Sep 24, 2026Contributor·

  ·
  4

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

  #4010·akoumpa opened Sep 24, 2026Contributor·

  ·
  4

• chore: CI proxy for #3999

  #4008·akoumpa opened Sep 24, 2026Contributor·

  ·
  3

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

  #4007·akoumpa opened Sep 24, 2026Contributor·

  ·
  3

• 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·

  ·
  2

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

  #4002·akoumpa opened Sep 23, 2026Contributor·

  ·
  2

• fix(nemotron_v3): accept singular embedding HF key alias

  community-request
  #3999·dundysm opened Sep 23, 2026·

  ·
  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·

  ·
  2

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

  community-request
  #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(recipes): use SDPA for GLM-5.3 sparse attention [AMINT-355]

  #3989·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(distributed): retain TE context-parallel token indices [AMINT-353]

  #3984·HuiyingLi opened Sep 22, 2026Contributor·

  ·
  3

• 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

• docs(fern): organize model coverage by provider and model ID

  #3974·akoumpa opened Sep 22, 2026Contributor·

  ·
  38

• 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