All pull requests

Pull requests

• feat(rl): exclude invalid and masked trajectories from advantage calculation

  #2437·s-noghabi opened Sep 24, 2026Collaborator·

  ·
  4

  

• Add public helpers to SingleTypeKVCacheManager.

  ready-to-submit
  #2436·copybara-service[bot] opened Sep 24, 2026Bot·

  ·
  1

  

• [v7x] 397B 256 recipe: restore v7x trainer config after mlperf_base move

  #2435·khatwanimohit opened Sep 24, 2026Collaborator·

  ·
  3

  

• Add allocation to SingleTypeKVCacheManager.

  ready-to-submit
  #2434·copybara-service[bot] opened Sep 24, 2026Bot·

  ·
  1

  

• Add prefix cache lookup to SingleTypeKVCacheManager.

  ready-to-submit
  #2433·copybara-service[bot] opened Sep 24, 2026Bot·

  ·
  1

  

• Add state sync to SingleTypeKVCacheManager for v2 sampler.

  ready-to-submit
  #2432·copybara-service[bot] opened Sep 24, 2026Bot·

  ·
  1

  

• Wuhao/mlperf 397b

  #2431·knightwh opened Sep 24, 2026Collaborator·

  ·
  5

  

• Introduce SingleTypeKVCacheManager for experimental v2 sampler.

  ready-to-submit
  #2430·copybara-service[bot] opened Sep 24, 2026Bot·

  ·
  1

  

• Unify Tunix container builds (FROM python:3.12-slim), split-layer MaxText/Raiden caching, and multi-region Artifact Registry publishing (tunix/mlperf) with CI workflows.

  ready-to-submit
  #2428·copybara-service[bot] opened Sep 24, 2026Bot·

  ·
  

• Lewu/rcp logging eval

  #2425·uwlei opened Sep 24, 2026Collaborator·

  ·
  14

  

• [Tunix] Count parser-appended assistant suffix tokens toward the response budget.

  #2418·copybara-service[bot] opened Sep 24, 2026Bot·

  ·
  

• fix wrong validation checks on train_mb when sequence packing is enabled

  #2412·s-noghabi opened Sep 24, 2026Collaborator·

  ·
  5

  

• Auto-run TPU unit tests on same-repo pull requests and enable CI on MLPerf branches

  #2411·copybara-service[bot] opened Sep 24, 2026Bot·

  ·
  

• Set custom_mamba_cache_multiplier to 32 in mlperf_35b_256 recipe

  #2410·wenxindongwork opened Sep 23, 2026Collaborator·

  ·
  

• Update mlperf_35b_256_multi-host.sh

  #2409·knightwh opened Sep 23, 2026Collaborator·

  ·
  2

  

• [tunix] Enable specifying checkpoint_restore_step and checkpoint_overwrite along the RL orchestrator and worker execution path

  #2407·copybara-service[bot] opened Sep 23, 2026Bot·

  ·
  

• feat(orchestrator): add RolloutFaultToleranceManager with capacity-ga…

  #2405·s-noghabi opened Sep 23, 2026Collaborator·

  ·
  7

  

• Fail fast when assistant step logprobs are missing instead of falling back to zero

  #2403·wenxindongwork opened Sep 23, 2026Collaborator·

  ·
  3

  

• Remove worker_rpc_client parameter and encapsulate WorkerRpcClient in JobEntity

  #2402·copybara-service[bot] opened Sep 23, 2026Bot·

  ·
  1

  

• fix(distributed): dynamic slicing admission for single-host v7x rollouts and 35B recipe updates

  #2393·niting opened Sep 23, 2026Collaborator·

  ·
  5

  

• [DeepSWE] Fix sandbox pool registration, Kueue finalizers, pod resources, and GRPO fault tolerance

  #2392·jiantanfig opened Sep 23, 2026Collaborator·

  ·
  8

  

• feat(experimental): add RoutingActorPool dynamic membership, HRW/leas…

  #2387·s-noghabi opened Sep 23, 2026Collaborator·

  ·
  5

  

• [Tunix] Add Score Centering support to Agentic GRPO learner and vLLM rollout.

  #2386·copybara-service[bot] opened Sep 22, 2026Bot·

  ·
  

• feat(deepswe): opt-in concurrent warming in PrewarmDatasetIterator

  #2379·tomergee opened Sep 22, 2026·

  ·
  5

  

• [Do not Submit] Enable GSM8K Qwen3-0.6B training in ci_smoke_gsm8k_qwen3_0p6b_maxtext.sh and patch cl/983986955

  #2378·copybara-service[bot] opened Sep 22, 2026Bot·

  ·