All pull requests

Pull requests

• fix: validate profile-band reward multipliers

  community-requestDocumentation
  #5012·Yushu-Li opened Nov 1, 2026·4 of 5 tasks·

  ·
  1

• refactor(checkpoint): move rollout checkpointing out of the single controller

  Documentation
  #5011·macandro96 opened Nov 1, 2026Contributor·

  ·
  1

• feat: Publish periodic vLLM and router inference telemetry

  community-requestDocumentation
  #5010·ekellbuch opened Nov 1, 2026·

  ·
  1

• docs(skills): add contributing testing rules, drop the tests they forbid

  #5009·yuki-97 opened Oct 31, 2026Contributor·

  ·
  1

• perf(grpo): enable Qwen3-30B async training MBS 4

  CI:Lfast
  #5007·youngeunkwon0405 opened Oct 31, 2026Contributor·4 of 5 tasks·

  ·
  2

• feat(loss): add sampler replay for top-k/top-p training incl score centering

  Documentation
  #5006·isomap opened Oct 31, 2026Contributor·

  ·
  1

• build: update Megatron-Bridge and its runtime dependencies

  CI:L1CI
  #5005·mxinO opened Oct 31, 2026Contributor·

  ·
  8

• ci: Super 3.5 RLVR SingleController smoke for the Gym-bump lane, with external judge pools

  #5003·yfw opened Oct 30, 2026Contributor·4 tasks done·

  ·
  1

• fix: index W&B performance history by trainer step

  CI:docs
  #5002·youngeunkwon0405 opened Oct 30, 2026Contributor·4 of 5 tasks·

  ·
  4

• perf(sc): opt-in switches that shorten startup and the first training step

  Documentation
  #5001·sarichardson-nv opened Oct 30, 2026·3 of 4 tasks·

  ·
  1

• chore(skills): consolidate local PR review on review-pr-team

  #4999·yuki-97 opened Oct 30, 2026Contributor·

  ·
  6

• docs(skills): move author-facing rules into contributing references

  CI:docs
  #4998·yuki-97 opened Oct 30, 2026Contributor·

  ·
  22

• feat: pair the Automodel fp32 LM head with vLLM's fp32 head

  Documentation
  #4995·akoumpa opened Oct 30, 2026·3 of 4 tasks·

  ·
  1

• perf(research/logra): Perf-LoGRA — port #4605 optimizations onto #4403

  community-request
  #4992·yl3469 opened Oct 30, 2026·2 of 3 tasks·

  ·
  1

• perf(distillation): defer fp32 upcast for masked local top-k

  community-requestwaiting-on-maintainers
  #4989·tianyi-zhang-02 opened Oct 30, 2026Contributor·

  ·
  1

• fix(docker): remove Windows launcher stubs from uv-managed Python

  cherry-pick-to-mainCI:L0
  #4986·kajalj22 opened Oct 29, 2026Contributor·4 tasks·

  ·
  7

• fix(sc): supervise failures while training drains

  community-requestwaiting-on-maintainers
  #4985·tianyi-zhang-02 opened Oct 29, 2026Contributor·3 of 4 tasks·

  ·
  1

• fix(algorithms): honor reward-side KL clamp configuration

  community-requestwaiting-on-maintainers
  #4984·tianyi-zhang-02 opened Oct 29, 2026Contributor·3 of 4 tasks·

  ·
  1

• chore: Bump mlflow and aiohttp for security fixes

  CI:L1
  #4981·kajalj22 opened Oct 29, 2026Contributor·2 of 3 tasks·

  ·
  3

• feat(dynamo): serve NeMo-Gym through the Dynamo frontend

  CI:L1CIDocumentation
  #4979·jthomson04 opened Oct 29, 2026Contributor·

  ·
  38

• refactor(vllm): share the per-sample rollout pipeline and chat token helpers

  CI:L1
  #4978·jthomson04 opened Oct 29, 2026Contributor·

  ·
  18

• feat: enable native FP32 LM-head logits for vLLM

  CI:LfastDocumentation
  #4976·joyang-nv opened Oct 29, 2026Member·4 tasks done·

  ·
  2

• ci: Bump Megatron-Bridge to e301c81

  CI:L1
  #4974·svcnvidia-nemo-ci opened Oct 29, 2026Contributor·

  ·
  3

• fix(model-utils): reuse the logits storage for the chunked logprob grad

  CI:Lfast
  #4973·bg51717 opened Oct 29, 2026Member·

  ·
  18

• test: load shipped configs once across validation checks

  CI:L0
  #4972·joyang-nv opened Oct 29, 2026Member·4 of 5 tasks·

  ·
  3