All pull requests

Pull requests

• fix(evals): reduce warmpool size and align CPU nodepools in eval_checkpoint_397B_v7.sh

  #2681·niting opened Oct 5, 2026Collaborator·

  ·
  1

  

• feat(evals): add eval_checkpoint_397B_v5p.sh for single-checkpoint

  #2680·uwlei opened Oct 5, 2026Collaborator·6 tasks done·

  ·
  

• Use the dense MoE fallback on every non-TPU platform in Qwen3

  #2676·dkling-it opened Oct 5, 2026·6 tasks done·

  ·
  4

  

• fix(deepswe): defer MLPerf RCP init_start/init_print until _on_train_start and bump mlcommons/logging to 6.1.0-rc2

  #2675·andytwigg opened Oct 5, 2026Collaborator·

  ·
  3

  

• Fix MLPerf RCP startup timing and ahead-of-time worker compilation

  #2672·andytwigg opened Oct 4, 2026Collaborator·

  ·
  8

  

• perf(deepswe_dist): parallelize rollout launch, batch Kueue workload recycle, and overlap prewarm with worker registration

  #2669·niting opened Oct 4, 2026Collaborator·

  ·
  3

  

• fix(cache): never fork for JAX cache GCS transfers; skip no-op uploads

  #2667·s-noghabi opened Oct 4, 2026Collaborator·

  ·
  4

  

• refactor(cache): derive rollout JAX cache bucket from recipe BUCKET (stacked on #2599)

  #2660·s-noghabi opened Oct 4, 2026Collaborator·

  ·
  3

  

• fix(deepswe): resolve toolcall failures and submission drops in OpenHands harness

  #2659·andytwigg opened Oct 4, 2026Collaborator·

  ·
  8

  

• Compile the trainer before the first batch, not inside it

  #2652·A9isha opened Oct 4, 2026Collaborator·

  ·
  10

  

• Fix low_var_kl going negative near equal log probabilities

  #2638·Shr3yash opened Oct 3, 2026·6 tasks done·

  ·
  1

  4

  

• [mlperf] Add intra-prompt rollout over-generation, straggler cancellation, and 35B recipe fixes

  #2633·tianshub opened Oct 3, 2026Collaborator·

  ·
  1

  

• Log actual restored init_checkpoint_step and fail fast on non-zero resume under RCP logging

  #2632·uwlei opened Oct 3, 2026Collaborator·

  ·
  19

  

• 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

  

• Wuhao/deepswe nv openhands

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

  ·
  10

  

• 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

  

• Smoke test

  #2614·pv97 opened Oct 2, 2026·

  ·
  

• 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

  

• 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·

  ·
  

• 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·

  ·
  

• 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