All pull requests

Pull requests

• cp: use GitHub App authentication for code freeze

  Run CICD
  #4418·svcnemo-autobot opened Oct 18, 2026Contributor·

  ·
  2

• ci: proxy CI run for PR #4404

  #4415·HuiyingLi opened Oct 18, 2026Contributor·

  ·
  3

• feat(decision): add CLM, CLEF and D1 post-training

  community-request
  #4414·zlapp opened Oct 18, 2026·3 tasks done·

  ·
  1

• perf(deepseek-v41): sparse-index plan per selection, replayed in AC recompute

  #4413·yisongbetter opened Oct 18, 2026Contributor·4 tasks done·

  ·
  4

• chore: CI proxy for #4179

  #4412·HuiyingLi opened Oct 18, 2026Contributor·

  ·
  3

• chore: CI proxy for #4405

  #4411·HuiyingLi opened Oct 17, 2026Contributor·

  ·
  3

• chore: CI proxy for #4395

  #4410·HuiyingLi opened Oct 17, 2026Contributor·

  ·
  3

• chore: CI proxy for #4177

  #4406·HuiyingLi opened Oct 17, 2026Contributor·

  ·
  3

• perf(moe): reduce-scatter the torch-dispatcher EP combine

  community-requestwaiting-on-customer
  #4405·ischencheng opened Oct 17, 2026·2 of 3 tasks·

  ·
  1

• fix(vlm): align MiMo-V2 image preprocessing with the official sglang stack

  community-requestwaiting-on-customer
  #4404·haowen-han opened Oct 17, 2026Contributor·3 tasks done·

  ·
  8

• fix(moe): release per-dispatch state from the flex token-dispatch managers

  #4402·HuiyingLi opened Oct 17, 2026Contributor·

  ·
  1

  4

• feat(vlm): add Nemotron 3.5 Bolt LoRA recipe

  #4396·HuiyingLi opened Oct 16, 2026Contributor·3 tasks done·

  ·
  15

• perf(moe): serve the MoE combine output from the forward in the AC recompute

  #4395·yisongbetter opened Oct 16, 2026Contributor·4 tasks done·

  ·
  15

• refactor(init): preserve storage dtypes explicitly during initialization

  #4389·akoumpa opened Oct 16, 2026Contributor·4 of 5 tasks·

  ·
  1

• feat(moe): add experimental NCCL-EP dispatcher

  #4384·akoumpa opened Oct 15, 2026Contributor·4 of 7 tasks·

  ·
  2

• feat(precision): preserve exact parameter update surfaces before loading

  #4377·vmendelev opened Oct 15, 2026·3 tasks done·

  ·
  8

• perf(kimi): make reference expert-count reductions contiguous

  #4372·akoumpa opened Oct 15, 2026Contributor·

  ·
  1

• fix(peft): standardize LoRA storage and compute precision

  #4359·akoumpa opened Oct 14, 2026Contributor·3 tasks done·

  ·
  24

• perf(models): re-enable TE RoPE fusion

  #4358·HuiyingLi opened Oct 14, 2026Contributor·3 of 6 tasks·

  ·
  1

• perf(moe): use native HybridEP ragged slots

  #4356·akoumpa opened Oct 14, 2026Contributor·

  ·
  6

• feat(telemetry): export OpenTelemetry spans and metrics through nemo-lens

  #4354·ahmadki opened Oct 14, 2026Member·3 tasks done·

  ·
  1

• feat(kd): add Qwen MoE distillation examples and validation logs

  #4351·akoumpa opened Oct 14, 2026Contributor·3 tasks done·

  ·
  1

  3

• chore: CI proxy for #4187

  #4340·akoumpa opened Oct 13, 2026Contributor·

  ·
  4

• fix(docker): backport HybridEP DMA-BUF registration fallback

  #4335·janbernloehr opened Oct 13, 2026Contributor·2 of 3 tasks·

  ·
  2

• perf(attention): pack sparse route membership for Qwen and MiniMax

  #4334·HuiyingLi opened Oct 13, 2026Contributor·3 tasks done·

  ·
  1