All pull requests

Pull requests

• Offload synchronous dataset iteration from the asyncio event loop in tunix.experimental.orchestrator:

  #2597·copybara-service[bot] opened Oct 1, 2026Bot·

  ·
  

• Align mlperf_base.sh step and reward timeouts to 60s

  #2596·susanbao opened Oct 1, 2026Collaborator·

  ·
  1

  

• Update mlperf_base.sh for the dataset

  #2595·knightwh opened Oct 1, 2026Collaborator·6 tasks·

  ·
  1

  

• Prioritize the current step's rollouts in vLLM under max_staleness

  #2593·s-noghabi opened Oct 1, 2026Collaborator·

  ·
  3

  

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

  #2581·knightwh opened Oct 1, 2026Collaborator·6 tasks·

  ·
  7

  

• fix(evals): support mesh_dp and dp attention in eval worker and update MLPerf eval recipes

  #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(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

  

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

  #2548·wenxindongwork opened Sep 30, 2026Collaborator·

  ·
  4

  

• [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·6 tasks·

  ·
  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

  

• Cherry-pick PR #2466 and PR #2518 (fused sampler-trainer agreement & TIS) onto atwigg/mlperf

  #2524·tianshub opened Sep 28, 2026Collaborator·

  ·
  10

  

• fix(deepswe): match active claims and pods via warmpool ref and creator labels in cleanup script

  #2523·niting opened Sep 28, 2026Collaborator·

  ·
  5