All pull requests

Pull requests

• refactor(evals): migrate EvalConfig from TypedDict to BaseModel

  community-request
  #4909·lonexreb opened Oct 25, 2026Contributor·4 tasks done·

  ·
  2

• ci: Bump Megatron-Bridge to afcc7ee

  CI:L1
  #4908·svcnvidia-nemo-ci opened Oct 25, 2026Contributor·

  ·
  3

• fix(vllm): merge policy hf_config_overrides into generation hf_overrides

  community-requestDocumentation
  #4907·lonexreb opened Oct 25, 2026Contributor·

  ·
  1

  2

• fix(distributed): do not leak launcher env into checkpoint config.yaml

  community-request
  #4906·lonexreb opened Oct 25, 2026Contributor·

  ·
  1

  2

• feat(dynamo)!: serve any vLLM generation config through Dynamo's backend SDK

  CI:L1CIDocumentation
  #4905·jthomson04 opened Oct 24, 2026Contributor·

  ·
  47

• fix(sc): reinstall token capture on restarted vLLM shards

  Documentation
  #4904·zyzhou5 opened Oct 24, 2026Contributor·4 tasks·

  ·
  1

• feat(loss): add ipo probability-gap token mask

  CI:L0
  #4903·cmunley1 opened Oct 24, 2026Contributor·

  ·
  7

• fix: avoid softmax collapse in target logprobs

  community-request
  #4902·kashif opened Oct 24, 2026·3 of 4 tasks·

  ·
  1

• fix(router-replay): accept vLLM's full-layer route layout and hybrid MoE layer patterns

  #4901·linj-glitch opened Oct 24, 2026Contributor·4 tasks done·

  ·
  1

• fix(router-replay): treat routeless batches and token-free groups as native-routed, not schema errors

  #4900·linj-glitch opened Oct 24, 2026Contributor·3 of 4 tasks·

  ·
  1

• fix(rollout-manager): stamp a per-rollout id on Gym rows when the ledger supplies none

  #4899·linj-glitch opened Oct 24, 2026Contributor·2 of 4 tasks·

  ·
  1

• fix(nemo-gym-shards): give each shard instance its own uv_venv_dir

  #4898·linj-glitch opened Oct 24, 2026Contributor·4 tasks done·

  ·
  1

• fix(data): add full-conversation SFT assistant-span formatting

  community-requestDocumentation
  #4897·vigneshwaran opened Oct 24, 2026Contributor·3 tasks done·

  ·
  1

• fix(ray.sub): add NRL_SRUN_MPI so the Ray sruns can run with --mpi=none

  #4896·linj-glitch opened Oct 24, 2026Contributor·3 of 4 tasks·

  ·
  1

• fix(gym): make NeMo-Gym telemetry reach Gym's server processes

  Documentation
  #4894·rrs45 opened Oct 24, 2026Collaborator·3 of 4 tasks·

  ·
  1

• feat(grpo): Add kimi style length penalty

  Documentation
  #4893·smahdavi4 opened Oct 24, 2026Contributor·4 tasks done·

  ·
  1

• fix(external_gym_vllm): skip link-local addresses for the replica head IP

  CI:L0
  #4892·cmunley1 opened Oct 24, 2026Contributor·

  ·
  7

• test(ci): uv cache seed refresh/reuse/prune test for #4863 (DO NOT MERGE)

  CI:L0CIDocumentation
  #4891·kajalj22 opened Oct 23, 2026Contributor·

  ·
  2

• feat(sc): log every Gym rollout with its model calls and RL's decision

  Documentation
  #4890·pthombre opened Oct 23, 2026Contributor·

  ·
  1

• build: add STRIP_UV_CACHE to drop uv cache from the r0.8.0 release image

  CI:L1Documentation
  #4889·kajalj22 opened Oct 23, 2026Contributor·3 tasks·

  ·
  8

• docs(gym): add Megatron GRPO recipe and update integration flow

  CI:docsDocumentation
  #4887·zyzhou5 opened Oct 23, 2026Contributor·

  ·
  11

• fix(gym): reuse rollout clients to prevent capture timeouts

  #4886·aroshanghias-nvd opened Oct 23, 2026Contributor·

  ·
  3

• chore: Bump aiohttp, virtualenv, gitpython, pillow for security fixes (r0.8.0)

  CI:L1
  #4885·kajalj22 opened Oct 23, 2026Contributor·1 of 2 tasks·

  ·
  4

• fix(loss): reject non-finite active IS weights in seq-mask-tis

  community-request
  #4883·ot4562-glitch opened Oct 23, 2026·

  ·
  2

• feat(single-controller): support reward_shaping on the SC path

  CI:LfastDocumentation
  #4882·RayenTian opened Oct 23, 2026Contributor·4 tasks·

  ·
  1