All pull requests

Pull requests

• Allow checkpoint_optimizer_interval_steps=0 for params-only checkpoints

  #2630·khatwanimohit opened Oct 2, 2026Collaborator·

  ·
  1

  

• Make RCP logging and manifest I/O failures non-fatal

  #2629·andytwigg opened Oct 2, 2026Collaborator·

  ·
  9

  

• [mlperf] Port orchestrator & batch packing memory optimizations (#2573, #2621, #2624)

  #2628·tianshub opened Oct 2, 2026Collaborator·

  ·
  4

  

• Jfacevedo/deepswe turn hints off

  #2626·entrpn opened Oct 2, 2026Collaborator·6 tasks·

  ·
  4

  

• Wuhao/deepswe nv openhands

  #2625·knightwh opened Oct 2, 2026Collaborator·6 tasks·

  ·
  9

  

• - Update _routed_experts_for in algorithm_adapter.py to return a zero-copy prefix view routed_arr[:seq_len] directly without allocating a per-trajectory UNSET_ROUTED_EXPERT padding array or calling np.concatenate, and avoid redundant np.concatenate([p_arr, c_arr]) in GRPOAdapter.create_trainer_payloads.

  #2624·copybara-service[bot] opened Oct 2, 2026Bot·

  ·
  

• fix(mlperf): Fix MLPerf 6.1 RCP compliance disclosures and rollout staleness lookahead

  #2623·uwlei opened Oct 2, 2026Collaborator·

  ·
  7

  

• Add CHECKPOINT_SKIP_IF_IN_PROGRESS for MaxText's skip_checkpoint_save_if_in_progress

  #2622·A9isha opened Oct 2, 2026Collaborator·6 tasks done·

  ·
  3

  

• feat(deepswe): align OpenHands prompts, tool schemas, parser, and execution semantics with MLPerf v6.1

  #2620·susanbao opened Oct 2, 2026Collaborator·

  ·
  16

  

• Fix orchestrator memory accumulation and extra routed_experts copies

  #2619·andytwigg opened Oct 2, 2026Collaborator·

  ·
  7

  

• Remove obsolete force_on_policy_ratio and log_sampler_trainer_agreement flags and unify GRPO in-loss agreement fusion across learners:

  #2616·copybara-service[bot] opened Oct 2, 2026Bot·

  ·
  

• Smoke test

  #2614·pv97 opened Oct 2, 2026·6 tasks·

  ·
  

• fix(deployment): wait for all nodes to join Ray cluster in mcjax ray templates

  #2611·andytwigg opened Oct 2, 2026Collaborator·

  ·
  4

  

• fix(distributed): verify worker_id on RPCs, hash JOB_PREFIX ports, and configure sandbox RBAC/NetworkPolicy

  #2606·andytwigg opened Oct 2, 2026Collaborator·

  ·
  4

  

• Feat(Deployment): Support colocated_python checkpointing

  #2603·YixuanWang-99 opened Oct 2, 2026Collaborator·

  ·
  8

  

• DeepSWE RL: non-blocking log fields for per-step performance attribution

  #2602·ultrons opened Oct 2, 2026·

  ·
  6

  

• Add the tunix/experimental/zero_tim_kernel package and Qwen3 canonical model benchmark suite (tunix/experimental/zero_tim_kernel/bench/qwen3):

  #2600·copybara-service[bot] opened Oct 1, 2026Bot·

  ·
  

• [MLPerf] Support GCS persistence and restore for JAX compilation cache in rollout and trainer

  #2599·wenxindongwork opened Oct 1, 2026Collaborator·

  ·
  

• Decouple trajectory store interfaces and implementations from TunixTrajectoryMetadata so stores can statically type, preserve, and rehydrate arbitrary TrajectoryMetadata subclasses.

  #2598·copybara-service[bot] opened Oct 1, 2026Bot·

  ·
  

• Update mlperf_base.sh for the dataset

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

  ·
  1

  

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

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

  ·
  7

  

• 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