All pull requests

Pull requests

• [Tunix] Compute orchestrator advantages with numpy instead of jax.numpy.

  #3126·copybara-service[bot] opened Nov 3, 2026Bot·

  ·
  

• [Tunix] Move validate_packing_budget from rl.utils to rl.packing.

  #3125·copybara-service[bot] opened Nov 3, 2026Bot·

  ·
  

• Both vLLM rollout samplers already run in server mode no matter what callers

  #3124·copybara-service[bot] opened Nov 3, 2026Bot·

  ·
  1

  

• [Tunix] Gate DispatchWindow directly by in-flight prompt count.

  #3122·copybara-service[bot] opened Nov 3, 2026Bot·

  ·
  

• Write step files before metadata and use atomic POSIX file replacement in FileTrajectoryStore.

  #3120·copybara-service[bot] opened Nov 3, 2026Bot·

  ·
  1

  

• The previous example launchers and run_*_node.py mains export OpenTelemetry metrics when OTEL_EXPORTER_OTLP_ENDPOINT is set. Port that to the shared dist_rl/launch tooling, so the examples keep it once they switch:

  #3117·copybara-service[bot] opened Nov 3, 2026Bot·

  ·
  

• Extracts derive_kv_geometry, derive_cache_sharding, compute_page_limits, and create_kv_cache_group_managers out of kv_cache_manager.py into utils.py.

  #3116·copybara-service[bot] opened Nov 3, 2026Bot·

  ·
  1

  

• On cluster_config.target: k8s the pods did not resolve the same config, so the config_sha256 each node logs could not match across pods:

  #3114·copybara-service[bot] opened Nov 3, 2026Bot·

  ·
  

• Consolidates the v2 generation engine's configuration dataclasses (CacheConfig, SchedulerConfig, and ModelRunnerConfig) into a shared tunix.experimental.generate.configs module so lower-level modules can depend on them without circular imports.

  #3111·copybara-service[bot] opened Nov 2, 2026Bot·

  ·
  1

  

• Add GPU flash attention kernel to Tunix models.

  #3102·copybara-service[bot] opened Nov 2, 2026Bot·

  ·
  

• Enable GPU on legacy peft_trainer v1 and vanilla v1.

  #3101·copybara-service[bot] opened Nov 2, 2026Bot·

  ·
  

• Add CL Complexity Gate (scripts/check_cl_complexity.py and [tool.cl_complexity_gate] in pyproject.toml).

  #3100·copybara-service[bot] opened Nov 2, 2026Bot·

  ·
  

• Recipes: Add standalone trellis_scaling_run.sh recipe

  #3099·andytwigg opened Oct 31, 2026Collaborator·

  ·
  4

  

• DeepSWE: Add SWE-smith and SWE-Gym rollout setup and fresh-container evaluation

  #3098·andytwigg opened Oct 31, 2026Collaborator·

  ·
  6

  

• Orchestrator: DAPO dynamic sampling, dispatch windowing, memory release, and unfiltered reward metrics

  #3097·andytwigg opened Oct 31, 2026Collaborator·

  ·
  7

  

• feat(eval): merge distributed DeepSWE evaluation pipeline and MLPerf eval recipes (#2389, #2504, #2579, #2733, #2795)

  #3094·susanbao opened Oct 30, 2026Collaborator·

  ·
  7

  

• [Tunix] Batch PollResponses to drain ready worker results per RPC.

  #3093·copybara-service[bot] opened Oct 30, 2026Bot·

  ·
  

• Delete train_maxtext_nb.py

  #3089·niting opened Oct 30, 2026Collaborator·

  ·
  5

  

• Fix RLLearner resume skipping unconsumed input micro-batches when num_iterations > 1

  #3086·SUPR1M3 opened Oct 30, 2026·6 tasks done·

  ·
  1

  4

  

• Add documentation for RL metrics and per-role profiling with Google Cloud ML Diagnostics.

  #3085·copybara-service[bot] opened Oct 30, 2026Bot·

  ·
  1

  

• Add 3D CL Understandability & Complexity Gate (scripts/check_cl_complexity.py and [tool.cl_complexity_gate] in pyproject.toml).

  #3081·copybara-service[bot] opened Oct 30, 2026Bot·

  ·
  

• Stop serializing duplicate token and logprob arrays on TunixAgentStep.

  #3073·copybara-service[bot] opened Oct 29, 2026Bot·

  ·
  

• Add a standalone end-to-end verification script (google3/third_party/py/tunix/dist_rl/distributed/deployment/verify_otel_gke.sh) that renders jobset.otel.yaml, deploys the in-cluster OTel Collector service (${USER}-otel:4317), and launches 4 worker pods (rollout-0, rollout-1, trainer-0, orchestrator-0) to verify K8s Service DNS resolution, otel_setup.setup_metrics() initialization across all roles, and OTLP gRPC metric export.

  #3072·copybara-service[bot] opened Oct 29, 2026Bot·

  ·
  

• When converter.create_agent_step and converter.create_env_step translate an agent_types.Step into TunixAgentStep and TunixEnvStep, downstream calls to AsyncTrajectoryWriter.add_step and InMemoryTrajectoryStore.add_step invoke trajectory.deep_copy, which previously fell through to copy.deepcopy and duplicated every NumPy token, mask, logprob, and routed-expert buffer. Additionally, converter.create_env_step aliased step.info directly onto TunixEnvStep.extra instead of taking a shallow copy like create_agent_step.

  #3070·copybara-service[bot] opened Oct 29, 2026Bot·

  ·
  

• Remove the AlgorithmAdapter (GRPOAdapter / PPOAdapter) wrapper layer from tunix/dist_rl/orchestrator/ and unify distributed RL with tunix/rl/algo_core.py and tunix/rl/algorithm_config.py:

  #3063·copybara-service[bot] opened Oct 28, 2026Bot·

  ·