Puts a list of data into the Ray object store in parallel using a thread pool. Args: data_list (List[Any]): A list of Python objects

Parse a PDF document using MinerU Python API (pipeline mode). Args: document_path: Path to PDF file inside config.docume

postprocess the output of a forward_backward_batch. output_lst is a list of dict containing outputs for each micro-batch reorder entropy and o

Recover left padding from result return result

Postprocess bshd sequences

_summary_ For fused kernels, the output is a dictionary with keys like 'log_probs', 'entropy', etc. This function post-processes each tensor i

Postprocess packed sequences

Prepare micro batches from data.

Remove left padding from input_ids, attention_mask and position_ids return new_input_ids, new_attention_mask, new_position_ids

Preprocess bshd sequences return "input_ids, attention_mask, position_ids"

Preprocess packed sequences CP splits sequence into CP*2 chunks, and each GPU gets 2 chunks (GPU0 gets first and last chunks, GPU1 gets s

Reduces a dictionary of metric lists by computing the mean of each list. Args: metrics: A dictionary mapping metric names to lists o

Reduce timing information across all processes. This function uses distributed communication to gather and sum the timing information from al

Register a new rollout replica type.

Reload.

Remove the LaTeX boxed command from a string. Args: s: String with format "\\boxed{content}" Returns: The content inside the

Note that this function only rename the key in the batch

This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch, num_key_value_heads, seqlen, he

This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch, num_key_value_heads, seqlen, he

This is the equivalent of torch.repeat_interleave(x, dim=2, repeats=n_rep). The hidden states go from (batch, seqlen, num_key_value_heads, he

Resize total and recalculate remain accordingly.

Resume GPU memory occupation. Args: tags (Optional[List[str]], optional): List of tags to specify which memory to resume.

Rewrite dataset based on evaluation results. Args: dataset: Original dataset evaluations: Evaluation results

Context switch hybridengine to rollout mode.

Context switch hybridengine to rollout mode.

Rotates half the hidden dims of the input.

Execute the main PPO training workflow. This method sets up the distributed training environment, initializes workers, datasets, and

Run the full evaluation pipeline. Returns: Evaluation results dictionary

Run GRPO training. Args: config: GRPO configuration log_callback: Optional callback for each log line

Run inference on dataset using the specified model. Args: dataset: Test dataset with 'input' field model: Mo

Run SFT training. Args: config: SFT configuration log_callback: Optional callback for each log line

Save all three categorized datasets to separate files. Args: solved: Dataset with solved samples (score=1.0)

Generate SSE events from a queue for SDG jobs. Args: event_queue: Queue containing (event_type, data) tuples Yields: SS

Set the dispatch and collect dp_ranks for all registered meshes. Args: mesh_name (str): Mesh name to set dispatch and collect dp_

Store the SDG service for later refinement.

Sets the target topk indices for replay.

Set ulysses sequence parallel process group.

Sleep each rollout server.

Slice the DataProto and return a new DataProto object. This is an improved version of direct slicing which returns a DataProtoItem.

Sort the placement groups by node ip, all bundles in a single placement group should be on the same node. FSDPCheckpointManager saves sharde

Spawn to a dictionary of worker groups, each with a subset of method with prefix. Args: prefix_set: Set of prefixes to create wor

Split the elements in a tuple/interval, while handling well-formatted commas in large numbers

Start a new phase.

Context manager for temporarily setting an environment variable. This context manager ensures that environment variables are properly set and res

Convert to verl trainer command arguments.

Calculate topk items take-up ratio, and min/max timing across all ranks.

Generate SSE events for training job by polling log list. Args: reporter: Training progress reporter with log buffer Yields:

Context manager entry for switching the engine and model into training mode. Usage: with engine.train_mode():

Translate a single text from source language to target language.

Unpad the data proto with pad_size. i.e. `data[:-pad_size]`

Update the policy with an iterator of DataProto Args: data (DataProto): an iterator over the DataProto that returns by

Validate critic configuration with runtime parameters. Args: n_gpus: Total number of GPUs available train_batch_size:

Compute entropy when the logits are sharded in tp ranks Args: vocab_parallel_logits: (total_nnz, vocab_size // tp_size) Returns: (to

Wake up all rollout replica instances.

Initialize the HTTP server engine adapter. Args: router_ip (Optional[str], optional): IP address of router for worker registratio

Initialize parser with configuration. Args: config: Parser configuration containing device, document_dir and other setti

Initialize the Rewriter

Initialize the TextDataGenerator.

Validate critic configuration parameters.

Perform post-processing operations on the output of each individual agent loop.