Contents Menu Expand Light mode Dark mode Auto light/dark, in light mode Auto light/dark, in dark mode Skip to content
dockyard_rl
dockyard_rl
  • Getting started
    • Installation
    • Quickstart — GRPO on SWE-bench
    • Configuration
  • Architecture
    • Architecture overview
    • Cluster bootstrap, fleets, and placement
    • Weight synchronization
    • The experience data plane
  • Design docs
    • GRPO and the clipped policy-gradient loss
    • Rewards and integrity
    • The sandbox task executor
    • Structured tool-use protocol
    • Mixture-of-Experts
    • Quantized generation
    • The JAX trainer backend
    • Preference optimization (DPO)
    • Distillation
  • Environments
    • SWE — coding-agent patch scoring
    • ProgramBench — cleanroom reconstruction
    • Terminal-Bench — terminal episodes
    • OSWorld — the computer-use agent
    • GDPval — file-producing deliverables
    • Verifier evals — HLE and math
  • Deployment
    • Slurm
    • Kubernetes
    • Hardware-deferred validation
  • Glossary
  • API Reference
    • dockyard_rl
      • dockyard_rl.distributed
        • dockyard_rl.distributed.virtual_cluster
        • dockyard_rl.distributed.model_utils
        • dockyard_rl.distributed.stateless_process_group
        • dockyard_rl.distributed.named_sharding
        • dockyard_rl.distributed.worker_groups
        • dockyard_rl.distributed.batched_data_dict
        • dockyard_rl.distributed.worker_groups_utils
      • dockyard_rl.environments
        • dockyard_rl.environments.code_jaccard_environment
        • dockyard_rl.environments.hle_environment
        • dockyard_rl.environments.vlm_environment
        • dockyard_rl.environments.utils
        • dockyard_rl.environments.rewards
        • dockyard_rl.environments.reward_model_environment
        • dockyard_rl.environments.gdpval_environment
        • dockyard_rl.environments.retriever
        • dockyard_rl.environments.program_bench_environment
        • dockyard_rl.environments.dapo_math_verifier
        • dockyard_rl.environments.terminal_bench_environment
        • dockyard_rl.environments.interfaces
        • dockyard_rl.environments.math_environment
        • dockyard_rl.environments.metrics
        • dockyard_rl.environments.multi_turn_session_env
        • dockyard_rl.environments.gdpval_agentic_environment
        • dockyard_rl.environments.code_environment
      • dockyard_rl.models
        • dockyard_rl.models.generation
          • dockyard_rl.models.generation.sglang
            • dockyard_rl.models.generation.sglang.sglang_worker
            • dockyard_rl.models.generation.sglang.utils
            • dockyard_rl.models.generation.sglang.config
            • dockyard_rl.models.generation.sglang.sglang_generation
            • dockyard_rl.models.generation.sglang.sglang_copied_utils
          • dockyard_rl.models.generation.vllm
            • dockyard_rl.models.generation.vllm.quantization
              • dockyard_rl.models.generation.vllm.quantization.fp8
              • dockyard_rl.models.generation.vllm.quantization.fp8_train_utils
            • dockyard_rl.models.generation.vllm.utils
            • dockyard_rl.models.generation.vllm.vllm_worker_async
            • dockyard_rl.models.generation.vllm.config
            • dockyard_rl.models.generation.vllm.router_capture
            • dockyard_rl.models.generation.vllm.vllm_backend
            • dockyard_rl.models.generation.vllm.vllm_generation
            • dockyard_rl.models.generation.vllm.vllm_worker
          • dockyard_rl.models.generation.interfaces
          • dockyard_rl.models.generation.constants
        • dockyard_rl.models.policy
          • dockyard_rl.models.policy.workers
            • dockyard_rl.models.policy.workers.dtensor_policy_worker_v2
            • dockyard_rl.models.policy.workers.dtensor_policy_worker
            • dockyard_rl.models.policy.workers.base_policy_worker
          • dockyard_rl.models.policy.tq_policy
          • dockyard_rl.models.policy.utils
          • dockyard_rl.models.policy.teacher_worker_group
          • dockyard_rl.models.policy.teacher_scorer
          • dockyard_rl.models.policy.interfaces
          • dockyard_rl.models.policy.lm_policy
        • dockyard_rl.models.dtensor
          • dockyard_rl.models.dtensor.moe
            • dockyard_rl.models.dtensor.moe.detect
            • dockyard_rl.models.dtensor.moe.convert
            • dockyard_rl.models.dtensor.moe.router
            • dockyard_rl.models.dtensor.moe.surgery
            • dockyard_rl.models.dtensor.moe.block
            • dockyard_rl.models.dtensor.moe.config
            • dockyard_rl.models.dtensor.moe.experts
            • dockyard_rl.models.dtensor.moe.sharding
            • dockyard_rl.models.dtensor.moe.dispatch
            • dockyard_rl.models.dtensor.moe.mesh
            • dockyard_rl.models.dtensor.moe.refit
            • dockyard_rl.models.dtensor.moe.load_balance
            • dockyard_rl.models.dtensor.moe.router_replay
            • dockyard_rl.models.dtensor.moe.parallelize_moe
          • dockyard_rl.models.dtensor.parallelize
        • dockyard_rl.models.huggingface
          • dockyard_rl.models.huggingface.common
        • dockyard_rl.models.jax
          • dockyard_rl.models.jax.moe
            • dockyard_rl.models.jax.moe.router
            • dockyard_rl.models.jax.moe.block
            • dockyard_rl.models.jax.moe.experts
            • dockyard_rl.models.jax.moe.dispatch
            • dockyard_rl.models.jax.moe.qwen3_moe
            • dockyard_rl.models.jax.moe.refit
            • dockyard_rl.models.jax.moe.alltoall
            • dockyard_rl.models.jax.moe.load_balance
            • dockyard_rl.models.jax.moe.weights
          • dockyard_rl.models.jax.models
            • dockyard_rl.models.jax.models.qwen3
            • dockyard_rl.models.jax.models.gemma4_refit
            • dockyard_rl.models.jax.models.qwen3_next
            • dockyard_rl.models.jax.models.gemma4
          • dockyard_rl.models.jax.linear_attn
            • dockyard_rl.models.jax.linear_attn.gated_deltanet
            • dockyard_rl.models.jax.linear_attn.delta_rule
            • dockyard_rl.models.jax.linear_attn.refit
            • dockyard_rl.models.jax.linear_attn.weights
          • dockyard_rl.models.jax.convert
          • dockyard_rl.models.jax.layers
          • dockyard_rl.models.jax.config
          • dockyard_rl.models.jax.optimizer
          • dockyard_rl.models.jax.sharding
          • dockyard_rl.models.jax.loss
          • dockyard_rl.models.jax.refit
          • dockyard_rl.models.jax.policy_worker
          • dockyard_rl.models.jax.train_step
          • dockyard_rl.models.jax.checkpoint
          • dockyard_rl.models.jax.weights
      • dockyard_rl.experience
        • dockyard_rl.experience.cua
          • dockyard_rl.experience.cua.vendor
            • dockyard_rl.experience.cua.vendor.desktop_env_eval
              • dockyard_rl.experience.cua.vendor.desktop_env_eval.metrics
                • dockyard_rl.experience.cua.vendor.desktop_env_eval.metrics.utils
                • dockyard_rl.experience.cua.vendor.desktop_env_eval.metrics.table
                • dockyard_rl.experience.cua.vendor.desktop_env_eval.metrics.basic_os
                • dockyard_rl.experience.cua.vendor.desktop_env_eval.metrics.general
                • dockyard_rl.experience.cua.vendor.desktop_env_eval.metrics.pdf
              • dockyard_rl.experience.cua.vendor.desktop_env_eval.getters
                • dockyard_rl.experience.cua.vendor.desktop_env_eval.getters.info
                • dockyard_rl.experience.cua.vendor.desktop_env_eval.getters.misc
                • dockyard_rl.experience.cua.vendor.desktop_env_eval.getters.file
                • dockyard_rl.experience.cua.vendor.desktop_env_eval.getters.general
              • dockyard_rl.experience.cua.vendor.desktop_env_eval.grading_env
              • dockyard_rl.experience.cua.vendor.desktop_env_eval.dispatch
          • dockyard_rl.experience.cua.actions
            • dockyard_rl.experience.cua.actions.pyautogui_parse
            • dockyard_rl.experience.cua.actions.cc_transpile
          • dockyard_rl.experience.cua.datasets
            • dockyard_rl.experience.cua.datasets.osworld
          • dockyard_rl.experience.cua.tasks
            • dockyard_rl.experience.cua.tasks.validate
          • dockyard_rl.experience.cua.backends
            • dockyard_rl.experience.cua.backends.windows
            • dockyard_rl.experience.cua.backends.official
            • dockyard_rl.experience.cua.backends.interface
            • dockyard_rl.experience.cua.backends.control_center
            • dockyard_rl.experience.cua.backends.macos
          • dockyard_rl.experience.cua.clients
            • dockyard_rl.experience.cua.clients.proto
            • dockyard_rl.experience.cua.clients.guest_server
            • dockyard_rl.experience.cua.clients.control_center
          • dockyard_rl.experience.cua.grading
            • dockyard_rl.experience.cua.grading.builtins
            • dockyard_rl.experience.cua.grading.dispatch
            • dockyard_rl.experience.cua.grading.registry
            • dockyard_rl.experience.cua.grading.factory
          • dockyard_rl.experience.cua.rollout
          • dockyard_rl.experience.cua.vision
          • dockyard_rl.experience.cua.environment
          • dockyard_rl.experience.cua.provisioner
        • dockyard_rl.experience.rollouts
      • dockyard_rl.modelopt
        • dockyard_rl.modelopt.utils
        • dockyard_rl.modelopt.registry
      • dockyard_rl.tool_protocol
        • dockyard_rl.tool_protocol.schema
        • dockyard_rl.tool_protocol.registry
        • dockyard_rl.tool_protocol.protocol
        • dockyard_rl.tool_protocol.hermes
      • dockyard_rl.data
        • dockyard_rl.data.datasets
          • dockyard_rl.data.datasets.response_datasets
            • dockyard_rl.data.datasets.response_datasets.tulu3
            • dockyard_rl.data.datasets.response_datasets.dapo_math
            • dockyard_rl.data.datasets.response_datasets.daily_omni
            • dockyard_rl.data.datasets.response_datasets.avqa
            • dockyard_rl.data.datasets.response_datasets.response_dataset
            • dockyard_rl.data.datasets.response_datasets.helpsteer3
            • dockyard_rl.data.datasets.response_datasets.gsm8k
            • dockyard_rl.data.datasets.response_datasets.general_conversation_dataset
            • dockyard_rl.data.datasets.response_datasets.openmathinstruct2
            • dockyard_rl.data.datasets.response_datasets.aime24
            • dockyard_rl.data.datasets.response_datasets.oai_format_dataset
            • dockyard_rl.data.datasets.response_datasets.refcoco
            • dockyard_rl.data.datasets.response_datasets.squad
            • dockyard_rl.data.datasets.response_datasets.deepscaler
            • dockyard_rl.data.datasets.response_datasets.gemometry3k
            • dockyard_rl.data.datasets.response_datasets.oasst
            • dockyard_rl.data.datasets.response_datasets.clevr
          • dockyard_rl.data.datasets.eval_datasets
            • dockyard_rl.data.datasets.eval_datasets.gpqa
            • dockyard_rl.data.datasets.eval_datasets.local_math_dataset
            • dockyard_rl.data.datasets.eval_datasets.aime
            • dockyard_rl.data.datasets.eval_datasets.mmau
            • dockyard_rl.data.datasets.eval_datasets.math
            • dockyard_rl.data.datasets.eval_datasets.mmlu
            • dockyard_rl.data.datasets.eval_datasets.mmlu_pro
          • dockyard_rl.data.datasets.preference_datasets
            • dockyard_rl.data.datasets.preference_datasets.binary_preference_dataset
            • dockyard_rl.data.datasets.preference_datasets.tulu3
            • dockyard_rl.data.datasets.preference_datasets.helpsteer3
            • dockyard_rl.data.datasets.preference_datasets.preference_dataset
          • dockyard_rl.data.datasets.utils
          • dockyard_rl.data.datasets.terminal_bench
          • dockyard_rl.data.datasets.raw_dataset
          • dockyard_rl.data.datasets.swe_bench
          • dockyard_rl.data.datasets.swe_bench_pro
          • dockyard_rl.data.datasets.processed_dataset
          • dockyard_rl.data.datasets.gdpval
          • dockyard_rl.data.datasets.hle
          • dockyard_rl.data.datasets.program_bench
        • dockyard_rl.data.packing
          • dockyard_rl.data.packing.packing
          • dockyard_rl.data.packing.metrics
          • dockyard_rl.data.packing.algorithms
        • dockyard_rl.data.utils
        • dockyard_rl.data.multimodal_utils
        • dockyard_rl.data.cross_tokenizer_collate
        • dockyard_rl.data.chat_templates
        • dockyard_rl.data.collate_fn
        • dockyard_rl.data.interfaces
        • dockyard_rl.data.dataloader
        • dockyard_rl.data.llm_message_utils
        • dockyard_rl.data.processors
      • dockyard_rl.data_plane
        • dockyard_rl.data_plane.adapters
          • dockyard_rl.data_plane.adapters.noop
          • dockyard_rl.data_plane.adapters.transfer_queue
        • dockyard_rl.data_plane.schema
        • dockyard_rl.data_plane.column_io
        • dockyard_rl.data_plane.observability
        • dockyard_rl.data_plane.worker_mixin
        • dockyard_rl.data_plane.interfaces
        • dockyard_rl.data_plane.codec
        • dockyard_rl.data_plane.rdma
        • dockyard_rl.data_plane.factory
        • dockyard_rl.data_plane.preshard
      • dockyard_rl.utils
        • dockyard_rl.utils.packed_tensor
        • dockyard_rl.utils.device_query
        • dockyard_rl.utils.timer
        • dockyard_rl.utils.gguf_export
        • dockyard_rl.utils.config
        • dockyard_rl.utils.native_checkpoint
        • dockyard_rl.utils.memory_tracker
        • dockyard_rl.utils.device_capability
        • dockyard_rl.utils.checkpoint
        • dockyard_rl.utils.nsys
        • dockyard_rl.utils.logger
      • dockyard_rl.sandbox
        • dockyard_rl.sandbox.task_executor_client
      • dockyard_rl.cluster
        • dockyard_rl.cluster.fleet
        • dockyard_rl.cluster.placement
        • dockyard_rl.cluster.bootstrap
      • dockyard_rl.algorithms
        • dockyard_rl.algorithms.loss
          • dockyard_rl.algorithms.loss.utils
          • dockyard_rl.algorithms.loss.interfaces
          • dockyard_rl.algorithms.loss.wrapper
          • dockyard_rl.algorithms.loss.loss_functions
        • dockyard_rl.algorithms.x_token
          • dockyard_rl.algorithms.x_token.utils
          • dockyard_rl.algorithms.x_token.token_aligner
          • dockyard_rl.algorithms.x_token.loss_utils
        • dockyard_rl.algorithms.async_utils
          • dockyard_rl.algorithms.async_utils.replay_buffer
          • dockyard_rl.algorithms.async_utils.interfaces
          • dockyard_rl.algorithms.async_utils.trajectory_collector
        • dockyard_rl.algorithms.rm
        • dockyard_rl.algorithms.utils
        • dockyard_rl.algorithms.distillation
        • dockyard_rl.algorithms.opd
        • dockyard_rl.algorithms.rlaif
        • dockyard_rl.algorithms.grpo
        • dockyard_rl.algorithms.routing_diversity
        • dockyard_rl.algorithms.config_access
        • dockyard_rl.algorithms.reward_functions
        • dockyard_rl.algorithms.online_dpo
        • dockyard_rl.algorithms.effort_reward
        • dockyard_rl.algorithms.teacher_panel
        • dockyard_rl.algorithms.distillation_memory
        • dockyard_rl.algorithms.advantage_estimator
        • dockyard_rl.algorithms.logits_sampling_utils
        • dockyard_rl.algorithms.dpo_evals
        • dockyard_rl.algorithms.rollout_allocation
        • dockyard_rl.algorithms.sft
        • dockyard_rl.algorithms.dpo
      • dockyard_rl.weight_sync
        • dockyard_rl.weight_sync.http_weight_synchronizer
        • dockyard_rl.weight_sync.ipc_weight_synchronizer
        • dockyard_rl.weight_sync.collective_weight_synchronizer
        • dockyard_rl.weight_sync.interfaces
        • dockyard_rl.weight_sync.factory
      • dockyard_rl.rewards
        • dockyard_rl.rewards.integrity
        • dockyard_rl.rewards.terminal_bench
        • dockyard_rl.rewards.invalid_action
        • dockyard_rl.rewards.interfaces
        • dockyard_rl.rewards.hle_grader
        • dockyard_rl.rewards.swe_bench_pro
        • dockyard_rl.rewards.test_runner
        • dockyard_rl.rewards.compiler
        • dockyard_rl.rewards.diff_utils
        • dockyard_rl.rewards.gdpval_rubric
        • dockyard_rl.rewards.program_bench
      • dockyard_rl.evals
        • dockyard_rl.evals.answer_parsing
Back to top
Copyright © 2026, Kartik A (NullVoider)