All pull requests

Pull requests

• Resolve target-free MaxText rollout dtypes across MaxText revisions and pin CI wheel builds to ubuntu-24.04.

  #3035·copybara-service[bot] opened Oct 26, 2026Bot·

  ·
  

• Offload blocking TrainerWorker RPCs and share dummy MoE routed-experts resolution.

  #3034·copybara-service[bot] opened Oct 26, 2026Bot·

  ·
  

• Support JSON vLLM overrides, HBM utilization forwarding, and cache salt.

  #3033·copybara-service[bot] opened Oct 26, 2026Bot·

  ·
  

• Pathways Raiden weight sync and MaxText shims.

  #3032·copybara-service[bot] opened Oct 26, 2026Bot·

  ·
  

• Mask RL completions by rollout length so a generated EOS is trained on

  #3030·0xnithinmys opened Oct 25, 2026·5 of 6 tasks·

  ·
  1

  2

  

• perf(reshard): optionally batch the multi-controller weight publish

  #3028·lokic233 opened Oct 24, 2026·1 of 6 tasks·

  ·
  14

  

• perf(rl_learner): optional length-balanced per-host prompt shards

  #3027·lokic233 opened Oct 24, 2026·1 of 6 tasks·

  ·
  10

  

• perf(vllm_sampler): optionally defer post-processing submits

  #3026·lokic233 opened Oct 24, 2026·1 of 6 tasks·

  ·
  8

  

• perf(generate): optional vectorized sampled-token logprob extraction

  #3025·lokic233 opened Oct 24, 2026·1 of 6 tasks·

  ·
  8

  

• perf(vllm_sampler): optionally tokenize each prompt once

  #3024·lokic233 opened Oct 24, 2026·1 of 6 tasks·

  ·
  6

  

• perf(rl_learner): optionally freeze Python GC after warm-up

  #3023·lokic233 opened Oct 24, 2026·1 of 6 tasks·

  ·
  8

  

• perf(rl_cluster): optionally run metric flushes off the train loop

  #3022·lokic233 opened Oct 24, 2026·1 of 6 tasks·

  ·
  6

  

• feat(rl): opt-in multi-controller colocated RL without Pathways (host-local vLLM engines, per-host learner)

  #3021·lokic233 opened Oct 24, 2026·1 of 6 tasks·

  ·
  13

  

• docs: clarify GRPO training entry points

  #3020·shreya-mishra325 opened Oct 24, 2026·5 of 6 tasks·

  ·
  3

  

• Make AbstractTrainer's arguments explicit instead of **kwargs

  #3019·igorts-git opened Oct 24, 2026Collaborator·

  ·
  9

  

• Recipes: Add standalone trellis_scaling_run.sh recipe

  #3018·andytwigg opened Oct 23, 2026Collaborator·

  ·
  4

  

• DeepSWE: Add SWE-smith and SWE-Gym rollout setup and fresh-container evaluation

  #3017·andytwigg opened Oct 23, 2026Collaborator·

  ·
  6

  

• Orchestrator: DAPO dynamic sampling, dispatch windowing, memory release, and unfiltered reward metrics

  #3016·andytwigg opened Oct 23, 2026Collaborator·

  ·
  8

  

• Add the shared bash launch drivers and dry-run integration tests under tunix/dist_rl/launch/common/. This is additive; the existing tunix/dist_rl/examples/ launchers are unchanged:

  #3015·copybara-service[bot] opened Oct 23, 2026Bot·

  ·
  

• Add async_weight_sync pipelining to StandardRLProgram and DistributedRLEngine.

  #3013·copybara-service[bot] opened Oct 23, 2026Bot·

  ·
  

• Support host-staged source self-release in TrainerWorker and WeightSyncCoordinator.

  #3012·copybara-service[bot] opened Oct 23, 2026Bot·

  ·
  

• MLPerf: optional GCSFuse (Rapid bucket) checkpoint mount for colocated_python

  #3011·igorts-git opened Oct 23, 2026Collaborator·6 tasks done·

  ·
  6

  

• - Update ShowCommand.trajectory_id to accept one or more trajectory IDs (list[str] with nargs="+").

  #3003·copybara-service[bot] opened Oct 23, 2026Bot·

  ·
  

• The trajectory collector wrote only completion_tokens into each agent step's ATIF metrics, so stored Tunix rollouts carried no input token counts and the Trajectory Explorer could only show them as N completion.

  #3001·copybara-service[bot] opened Oct 22, 2026Bot·

  ·
  

• experimental: emit structured events at the twinned log sites; launcher forwarding (stacked on #2997)

  #2998·ultrons opened Oct 22, 2026·

  ·
  9