The response from a tool execution.

Token and Time counter for counting and estimating usage.

A DataProto is a data structure that aims to provide a standard protocol for data exchange between functions. It contains a batch (TensorDict

A wrapper class for Ray actors with initialization arguments. This class extends ClassWithInitArgs to provide additional functionality for co

Single rubric scoring range for G-Eval metrics

A class to represent a dataset consisting of samples, where each sample contains messages with specific roles.

A dispatcher that delegates to specific profilers based on config.tool. Supported tools: - nsys: NsightSystemsProfiler - npu: NPUProfiler

Worker profiler config. The inheritance from BaseConfig provides omegaconf.DictConfig-like interface for a dataclass config. Args: d

Used to count mfu during training loop Example: flops_counter = FlopsCounter(config) flops_achieved, flops_promised = flops_

Manage FSDP checkpointing in SPMD training. - Saves/loads per-rank sharded model & optimizer states - Persists full lr_scheduler and RNG

Sharding manager to support data resharding when using FSDP + Ulysses

Configuration for Mcore module wrapper.

A group of Ray workers that can be managed collectively. This class extends WorkerGroup to provide Ray-specific functionality for creating an

A unified tracking interface for logging experiment data to multiple backends. This class provides a centralized way to log experiment metrics, p

Extracts domain keywords using LLM based on task instruction and demo examples.

Checkpoint manager for Megatron-LM distributed training. This class manages the saving and loading of model checkpoints in a Megatron-LM

Configuration for model merger operations. Args: operation (str): Operation type - 'merge' or 'test'. backend (str): Backend type

Launch training jobs using verl framework.

TrainingWorker provides a Tinker-like API (https://thinkingmachines.ai/tinker/) as a RayWorkerGroup to a single controller. Currently, we onl

Agent loop output.

Additional arguments for extract_answers

Async token bucket for rate limiting with variable token consumption. The token bucket algorithm is a classic rate limiting technique that allows

Checkpoint handler handles the path, global_step of a checkpoint folder. Currently, it only works with a single model. We can expand it t

FSDP DataParallel PPO Actor or Ref worker Args: config (ActorConfig): Actor config actor_module (nn.Module): Actor or ref module

DataProtoFuture aims to eliminate actual data fetching on driver. By doing so, the driver doesn't have to wait for data so that asynchronous

Loads PDF documents from specified directories.

Model merger for FSDP (Fully Sharded Data Parallel) checkpoints. This class handles the conversion of FSDP distributed checkpoints into Hugg

A simple wrapper client for interacting with Hugging Face datasets and metadata.

MinerU parser for PDF document processing using pipeline mode. Uses MinerU Python API directly for better control and features. Workflow

Class for directly use original model without any postprocess method

A PyTorch profiler wrapper class for collecting performance metrics. TODO(haibin.lin): this should implement the DistProfiler interface, and the

Define a resource pool specification. Resource pool will be initialized first.

Reward model manager.

A custom collate_fn that handles batching of sequences. 1. for variable-length sequences, convert them into NestedTensors. 2. for fixed-l

Adaptive KL controller described in the paper: https://arxiv.org/pdf/1909.08593.pdf

Encapsulates all state variables for the agent loop.

Agent loop manager that manages a group of agent loop workers.

Answer extraction configuration.

Translator using Hala-style causal LM models with chat templates.

Compared to synchronize, this uses more memory because of the buffer but achieves better performance due to the overlapping. D2h and h2d copying a

Asynchronous HTTP-based adapter for SGLang engines. This class inherits from HttpServerAdapter and adds async capabilities for non-blocking H

A class to manage multiple OpenAI compatible LLM servers. This class provides - Load balance: least requests load balancing - Sticky sess

BM25-based retriever for finding relevant documents. Workflow: 1. Load documents from corpus 2. Rank documents using BM25 3. Ret

Configuration for efficient entropy kernel operations. Args: _backward (BackwardEnum): Backward computation method. Defaults to BackwardE

Configuration for Answer Correctness metric using G-Eval

Context-manager that offloads/recovers tensors through an offload hander. The hook just offloads/recovers the tensor object to the handler throug

Configuration for evaluation dataset

Main evaluator class for DeepEval-based post-training evaluation. Handles: - Loading test dataset - Running inference on models

class for rewrite synthetic data by adjust their difficulty

Evaluator for assessing synthetic data quality with the working model. This class handles: 1. Batch inference on synthetic data 2. A

Configuration for FSDP checkpointing. Args: FSDP_version (int): Version of FSDP being used. world_size (int): Number of processes

Configuration for FSDP (Fully Sharded Data Parallel). The inheritance from BaseConfig provides omegaconf.DictConfig-like interface for a dataclas