All pull requests

Pull requests

• Internal

  #2994·copybara-service[bot] opened Oct 22, 2026Bot·

  ·
  1

  

• Add export_trajectories_df reader helper and remove legacy TrajectoryLogger and _trajectory_logger_ext.cc.

  #2993·copybara-service[bot] opened Oct 22, 2026Bot·

  ·
  

• Support configurable pre-queue array filtering and deduplicate Step metrics in Tunix TrajectoryStore.

  #2992·copybara-service[bot] opened Oct 22, 2026Bot·

  ·
  

• Migrate GRPOLearner and qwen_agentic_maxtext from AsyncTrajectoryLogger to TrajectoryStore.

  #2991·copybara-service[bot] opened Oct 22, 2026Bot·

  ·
  

• AgenticSequenceRewardManager already supports environment-only rewards with reward_fns=None, but it gets there by skipping SequenceRewardManager.__init__, so its worker-pool state is never set up and close() relies on a getattr fallback. SequenceRewardManager(reward_fns=None) is worse: it wraps None into [None], passes the non-empty check, and only fails later with an unrelated TypeError when it scores. The experimental StandardRLProgram is about to build reward managers from algo_config.reward_manager, so a misconfigured manager should fail when it is created.

  #2990·copybara-service[bot] opened Oct 22, 2026Bot·

  ·
  

• Migrate StandardRLProgram from AsyncTrajectoryLogger to TrajectoryStore.

  #2989·copybara-service[bot] opened Oct 21, 2026Bot·

  ·
  

• Populate ATIF FinalMetrics at episode end in converter.py and TrajectoryCollectEngine.

  #2987·copybara-service[bot] opened Oct 21, 2026Bot·

  ·
  

• - Add --limit flag to SummaryCommand to bound trajectory loading and output table size.

  #2986·copybara-service[bot] opened Oct 21, 2026Bot·

  ·
  

• Implement TrajectoryMetrics persistence, shallow-merge upsert, and reader in SqlTrajectoryStore.

  #2985·copybara-service[bot] opened Oct 21, 2026Bot·

  ·
  

• Implement TrajectoryMetrics persistence and parallel batch reading in FileTrajectoryStore.

  #2984·copybara-service[bot] opened Oct 21, 2026Bot·

  ·
  

• Add TrajectoryMetrics task support and 3-tier queue admission policy to AsyncWriter.

  #2983·copybara-service[bot] opened Oct 21, 2026Bot·

  ·
  

• Implement TrajectoryMetrics in InMemoryTrajectoryStore, base protocols, and shared store_testing contract.

  #2982·copybara-service[bot] opened Oct 21, 2026Bot·

  ·
  

• [MLPerf RL 397B] Add e2e-mb-stage microbatch staging + sync mlperf_submission_standalone.sh

  #2980·uwlei opened Oct 21, 2026Collaborator·2/2·6 tasks done·

  ·
  18

  

• [MLPerf RL 397B] Merge #2884 + #2948 trainer recipe stack & pin MaxText #5885

  #2979·uwlei opened Oct 21, 2026Collaborator·1/2·6 tasks done·

  ·
  8

  

• StandardRLProgram scored trajectories with its own loop: it called each reward function as fn(completion, metadata) -> float per trajectory, or fell back to the environment's trajectory_reward when no reward functions were given. This duplicated SequenceRewardManager and AgenticSequenceRewardManager, used a different reward-function contract from the rest of Tunix, and ran reward functions inline on the event loop.

  #2971·copybara-service[bot] opened Oct 20, 2026Bot·

  ·
  

• Eliminate layout mismatch copies in Gemma 4 MoERagged._run_ffw_and_routing.

  #2968·copybara-service[bot] opened Oct 20, 2026Bot·

  ·
  1

  

• Remove vLLM sampler adapters and unify Raiden weight sync.

  #2963·copybara-service[bot] opened Oct 19, 2026Bot·

  ·
  1

  

• Generate PPO eval rollouts with the eval rollout config

  #2961·vineethsaivs opened Oct 18, 2026·

  ·
  3

  

• [Tunix] (v2 Sampler) Add ragged paged attention with TPU, GPU and pure-JAX backends.

  #2956·copybara-service[bot] opened Oct 16, 2026Bot·

  ·
  1

  

• [Tunix] (v2 sampler) Let ragged_paged_attention take RPAMetadata and shard itself over a mesh.

  #2955·copybara-service[bot] opened Oct 16, 2026Bot·

  ·
  1

  

• [Tunix] (v2 sampler) Run gemma4 through ragged paged attention when given RPAMetadata.

  #2954·copybara-service[bot] opened Oct 16, 2026Bot·

  ·
  1

  

• Use main RL rollout worker setup (jobset.tpu.yaml, 128x 2x2x1, FP8) and inherit args in offline eval scripts

  #2951·andytwigg opened Oct 16, 2026Collaborator·

  ·
  3

  

• Exclude padded positions from the logit distillation losses

  #2949·vineethsaivs opened Oct 16, 2026·

  ·
  3

  

• Qwen3.5-397B RL trainer on 128 TPU v7x: turn on the maxtext#5799 speedups in the tuned recipe; residual16 masters; per-chunk log-prob head gradient reduce-scatter

  #2948·parambole opened Oct 16, 2026Collaborator·

  ·
  

• Defer trajectory timing metadata sync to episode end

  #2944·joshua-goog opened Oct 16, 2026·

  ·
  6