A baseline implementation of applying RoPE for `thd` format. Args: t (Tensor): Input tensor T is of shape [t, h, d] cu_seqlens(Te

Applies the monkey patch for MoE Router Replay functionality. This patch dynamically adds the 'enable_routing_replay' attribute to Transforme

Convert arbitrary group labels to a contiguous 1-D torch.long tensor (0..G-1). Args: index: Any iterable of labels or tensor/ndarray

Assign a list with potentially nested structures (lists, dicts, etc.) to TensorDict. This function handles complex nested data structures like:

Build the memory references. The memory buffers are built using the build_memory_buffer API. This API will allocate a weight buffer pointer to the

Build DeepEval test cases from dataset and model outputs. Args: dataset: Test dataset with 'input' and 'output' (ground

Bulk reload group.

calculate rollout vs actor logprobs diff, for debugging purpose Args: data: DataProto the data batch to calculate

Calculate how many samples to extract from each dataset. Uses sequential score-based approach: extract from highest-scored datasets f

Phase 2: Calculate how many samples to extract from each dataset. Uses sequential score-based approach: extract from highest-scored d

Check the consistency of the DataProto. Mainly for batch and non_tensor_batch We expose this function as a public one so that user can call th

Checks the correctness of code generation using the remote sandbox API, processing test cases concurrently. Args: sandbox_fusion

Split text into chunks using recursive strategy. Args: text: Text content to chunk Returns: List of

Cleanup temporary files if needed.

Clears the recorded and target topk indices.

Clears the router replay action for this layer.

DataArc Synthetic Data Generation Toolkit.

Collates a list of samples into a single batch. Args: batch: A list of dictionary samples from the dataset. Ret

Compare two answers using the specified method. This is a convenience wrapper that routes to the appropriate comparison function

Mark job as completed.

Compute advantage estimates for policy optimization. This function computes advantage estimates using various estimators like GAE, GRPO, REINFORC

Computes various metrics from a batch of data for PPO training. This function calculates metrics related to scores, rewards, advantages, ret

Compute comprehensive metrics for truncated importance sampling weights. This function calculates statistics for truncated IS weights (TIS), usin

Compute policy loss with pure rollout correction (no PPO clipping). This function implements policy gradient with importance sampling correction

Computing reward given input_ids. The transformers should output a tensor with shape [batch_size, sequence_length], and the value at [EOS]

Compute rollout correction weights and apply rejection sampling. Computes importance sampling weights to correct for off-policy issues between

Compute importance sampling weights to correct for off-policy distribution shifts. This function calculates IS weights (π_train / π_rollout) usin

Compute rejection mask for outlier handling in off-policy RL training. This function identifies and masks outlier tokens/sequences using precompu

Compute comprehensive metrics for rejection sampling. This function calculates statistics for IS weights used in rejection sampling, balancin

Computes throughput metrics for PPO training. This function calculates performance metrics related to token processing speed, including

Computes timing metrics for different processing stages in PPO training. This function calculates both raw timing metrics (in seconds) and p

Compute the clipped value-function loss for PPO. Copied from https://github.com/huggingface/trl/blob/main/trl/trainer/ppo_trainer.py#L1151

Compute importance weights for resampling based on scores. Args: scores (torch.Tensor): Tensor of scores to compute weights from.

construct constraints for synthesis from the parsed_information. constraints like reference passages / web dataset fields / generatio

Load the model into /dev/shm to make the process of loading the model multiple times more efficient.

Count the number of trainable adapter parameters. Args: model: PyTorch model Returns: Tuple of (adapter_params, total_params

This function should return a class instance that delegates the calls to every cls in cls_dict

This function returns a FusedWorker class. `FusedWorker.{class_name}` -> FusedClass Use `class_name` as a param to directly access t

Create Answer Correctness G-Eval metric. Args: config: CorrectnessMetricConfig with criteria, steps, rubric, threshold judge

Create a device mesh for distributed training based on the world size and FSDP size. Args: world_size (int): Total number of process

Create Format Compliance G-Eval metric. Args: config: FormatComplianceMetricConfig with criteria, steps, rubric, threshold j

Args: model_name: override_config_kwargs: Returns:

Create Pairwise Preference G-Eval metric. Args: config: PairwiseMetricConfig with criteria and evaluation_steps judge_model:

Synthesizing / Filtering data from sources. Input: self.config self.sources: from self.prepare()

Compute the score for a given solution based on the data source. Args: data_source (str): The source dataset identifier which determines

Helper function for reproducibility in distributed training. See https://pytorch.org/docs/stable/notes/randomness.html for details.

Enables memory history recording for CUDA allocations. This function should be called before any large-scale CUDA allocations. For DDP or

Evaluate samples by running inference and comparing with ground truth. Args: samples: List of samples with 'input' and '

Evaluate samples with images by running VLM inference and comparing with ground truth. Args: samples: List of samples wi

Execute the tool. Args: instance_id: The instance id of the tool. parameters: The json string of the parameters of th

Execute the parallel processing Args: iterable_inputs: iterable inputs to be processed process_function: fun

Export chunks to JSONL format for passage dataset. Each chunk is converted to the format: {"text": "chunk_content"} Empty or whitespace-

Extract samples from a dataset, save images, and format them. Handles multiple image storage formats: - PIL Image (embedded

Phase 3: Extract samples from a dataset and format them.

Extract keywords for HuggingFace dataset search with domain-first priority. Strategy: 1. Domain field (if provided) - primar

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

Extract keywords for HuggingFace dataset search with domain-first priority. Strategy: 1. Domain field (if provided) - primar