Build prompt from sample input and instructions. Args: sample_input: The input/question from sample Returns:

Clear specific buffer subdirectories matching the patterns.

Compute loss with optional sequence parallelism and remove padding features

Computes information about prompts and responses from a batch. This is an internal helper function that extracts masks and lengths for promp

Recursively performs a deep comparison between two Python objects. - Handles NaN values correctly (NaN == NaN evaluates to True). - Handl

Dump rollout/validation samples as JSONL.

Evaluate a single prediction against ground truth. Args: sample_input: Original input/question ground_truth:

Extract text content from a block. Args: block: Block dict from pdf_info Returns: Extracted text or empty st

fetch tensor in tp shards across mp_group

Returns: entropy: # (bs, response_len) log_probs: # (bs, response_len)

Get master addr and port for this worker group

Get rank-specific checkpoint path following Megatron's convention. Returns path like: base_path/mp_rank_{tp:02d}_{pp:03d}_{ep:03d}/ Args:

Context manager for safe session access with proper connection pooling. Yields: aiohttp.ClientSession: Session instance for makin

Helper function containing the loading hf model logic

Log samples to wandb as a table

Normalize answer expressions.

Parse LLM response to extract single {'input': ..., 'output': ...}. Args: response: Raw LLM response string Ret

Process the padded outputs from _run_agent_loop and combine them into a batch.

Args: output: a dictionary containing loss, model_outputs and metrics Returns:

Remove generation prompt IDs from input tensors if they are present at the end.

Resolve device according to policy described in the module docstring. Priority: 1) explicit argument 2) VERL_FORCE_DEVICE env

Save metric evaluation results to JSON file.

Save an uploaded file to the target path.

Validate the bounding box dimensions and aspect ratio.

Validate a JSONL file contains required keys. Args: file_path: Path to the JSONL file prompt_key: Required key for prompts

Acquire tokens from the bucket, waiting if necessary. This method will block (using asyncio.sleep) until sufficient tokens are availa

Reroute to the appropriate apply_rotary_pos_emb function depending on bshd (conventional) / thd (packed seq) format In Qwen2-VL, the sha

Performs bootstrap resampling to estimate statistics of metrics. This function uses bootstrap resampling to estimate the mean and standard d

Build the torchrun command for SFT training. Args: config: SFT configuration train_parquet: Path to training

Build the memory buffer given weight_buffer_meta Args: weight_buffer_meta: contains mapping from name to a dictionary containing shape an

Bulk offload group.

Calculate the workload for a dense transformer block based on sequence length. FLOPs = 12 * hidden_size^2 * seqlen + 2 * hidden_size * seqlen

Calls the remote sandbox API to execute code with retry logic for Gateway Timeout, using increasing delay between retries. Logs internal call

Calls the remote search API to perform retrieval with retry logic for various errors, using increasing delay between retries. Logs internal c

Cancel current job.

Categorize samples into three datasets based on evaluation scores. Args: scores: List of scores (0.0 to 1.0) correspondi

A helper method to check if the passed module's key name matches any of the exclude modules in the adapter_config. Adapted from https://githu

A helper method to check if the passed module's key name matches any of the target modules in the adapter_config. Adapted from https://github

Clears the recorded and target topk indices in all instances.

Clears the router replay action for all router instances.

Mark current phase as complete.

Compute logits given a batch of data. Args: data (DataProto): a batch of data represented by DataProto. It must contain key ```in

Compute off-policy diagnostic metrics (helper function). This helper function operates on raw tensors and is used internally by: - compute_ro

Compute the attention mask for the response part of the sequence. This function extracts the portion of the attention mask that corresponds to th

Compute rollout correction metrics from log probabilities during training. This function is used in the actor to compute metrics using the CURREN

Unified interface for computing IS weights and rejection masks. This function combines IS weight calculation (truncated) and rejection sampling (

Concatenates tensordicts into a single tensordict on dim zero. Support nested tensor

Deprecated. Please use copy_to_local instead.

Create a new job.

Create Ray resource pools for distributed training. Initializes resource pools based on the resource pool specification, with each po

Create a sampler for the dataset. Arguments: data_config: The data config. dataset (Dataset): The dataset. Returns:

Create a dataset.

Create a dataset.

Create a dataset.

name: name of the parameter train_params: training parameters infer_params (Iterable[torch.Tensor]): a iterator towards list of parameter