All pull requests

Pull requests

• Keep profile_options as None when set_profile_options and enable_continuous_profiling flags are disabled.

  #2582·copybara-service[bot] opened Oct 1, 2026Bot·

  ·
  

• feat(weight_sync): overlap Raiden H2H transfer with rollout sampler p…

  #2581·knightwh opened Oct 1, 2026Collaborator·

  ·
  7

  

• Update rollout cluster settings in mlperf_397b_v7x_eval.sh to match mlperf_397b_256_v7x.sh

  #2579·susanbao opened Oct 1, 2026Collaborator·

  ·
  4

  

• Add VeRL-style orthogonal rollout importance-sampling (sampler_is) and rejection-sampling (sampler_rs) controls on AlgorithmConfig, unified inside compute_sampler_trainer_agreement_jax and grpo_loss_fn:

  #2578·copybara-service[bot] opened Oct 1, 2026Bot·

  ·
  

• feat(rollout): support partial rollout across weight syncs for vLLM sampler

  #2576·jiangyangmu opened Oct 1, 2026Collaborator·

  ·
  21

  

• feat(mlperf): automatic offline eval loop for MLPerf RL checkpoints

  #2574·uwlei opened Oct 1, 2026Collaborator·

  ·
  16

  

• Router replay: pack routed_experts contiguously and defer its padding past the trainer RPC

  #2572·lc5211 opened Sep 30, 2026Collaborator·

  ·
  3

  

• fix(utils): prevent trajectory drops and parallelize per-step GCS writes in AsyncTrajectoryLogger

  #2571·niting opened Sep 30, 2026Collaborator·

  ·
  5

  

• [experimental] orchestrator: fail the run when a worker rejoins (preemption)

  #2569·s-noghabi opened Sep 30, 2026Collaborator·

  ·
  8

  

• feat(orchestrator): support elastic rollout worker leave, re-registration, and recovery

  #2568·andytwigg opened Sep 30, 2026Collaborator·

  ·
  6

  

• - Convert GrpcRemoteActorHandle._ensure_async_channel into an async def method that captures any existing self._channel, binds the new channel and RPC stubs to the current event loop, and closes the stale channel (await old_channel.close()) when the event loop changes or is closed.

  #2567·copybara-service[bot] opened Sep 30, 2026Bot·

  ·
  

• Sync remote_execution.py and remote_execution_test.py with main (support large gRPC payload)

  #2564·tianshub opened Sep 30, 2026Collaborator·

  ·
  10

  

• Overlap weight sync with the next train step (--async_weight_sync)

  #2559·khatwanimohit opened Sep 30, 2026Collaborator·

  ·
  13

  

• feat(recipes): add FP8_MOE environment variable directly to mlperf recipes

  #2548·wenxindongwork opened Sep 30, 2026Collaborator·

  ·
  4

  

• Add create_trajectory to TrajectoryMetadata and make Trajectory generic over step types.

  #2544·copybara-service[bot] opened Sep 29, 2026Bot·

  ·
  

• [tunix] Add prepared diffusion policy scoring

  #2543·ethannnnnn opened Sep 29, 2026Contributor·

  ·
  4

  

• [tunix] Add prepared diffusion distillation batches

  #2542·ethannnnnn opened Sep 29, 2026Contributor·

  ·
  5

  

• Consolidates static rollout configuration onto RolloutConfig across the rollout stack:

  #2540·copybara-service[bot] opened Sep 29, 2026Bot·

  ·
  

• enable RPC retry

  #2539·copybara-service[bot] opened Sep 29, 2026Bot·

  ·
  

• Extends the progressive load recovery benchmark harness (data_generator.py, benchmark_lib.py, and run_benchmark.py) to simulate multi-worker concurrent trajectory appends and measure write throughput across both shared-writer and dedicated per-worker writer topologies.

  #2536·copybara-service[bot] opened Sep 29, 2026Bot·

  ·
  1

  

• Integrates SqlTrajectoryStore into the progressive load recovery benchmark harness (run_benchmark.py and benchmark_lib_test.py).

  #2535·copybara-service[bot] opened Sep 29, 2026Bot·

  ·
  1

  

• Plumb multi-host RL profile configuration from MaxText into Tunix core.

  #2530·copybara-service[bot] opened Sep 29, 2026Bot·

  ·
  1

  

• chore(recipes): set RAIDEN_BROADCAST_HOST_RATIO and RAIDEN_BROADCAST_PIPELINE_STAGES in mlperf_397b_256_v7x.sh

  #2529·justinlu13 opened Sep 29, 2026·

  ·
  3

  

• refactor(recipes): parameterize cluster config and add strict validation

  #2526·knightwh opened Sep 29, 2026Collaborator·

  ·
  13

  

• Implements TrajectoryStore configuration methods (_from_config and to_config) on SqlTrajectoryStore and registers the sql backend in TrajectoryStore.from_config.

  #2525·copybara-service[bot] opened Sep 28, 2026Bot·

  ·
  1