Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.

Extract surrounding text context for each image from pdf_info.json. Args: pdf_info_path: Path to pdf_info.json file c

Retrieves sharding information (device_mesh, mesh_dim_names) from a DTensor in the state_dict. If no DTensor is found, infers a simpl

Extract images from PDFs using MinerU parser with context. Args: reporter: Optional progress reporter Returns:

fetch gate_up tensor in tp shards

fetch gate_up tensor in tp shards

fetch tensor in tp shards across mp_group

fetch tensor in tp shards

fetch tensor in tp shards

Find the correct path for an image filename in a HuggingFace repo. Args: image_data: The image filename/path from the da

Find the latest checkpoint folder for model merging. verl checkpoint structures: SFT: ./checkpoints/global_step_X/ ├──

Find the latest checkpoint in the default local directory

Find the latest checkpoint in the default local directory

Character-level split with overlap (last resort).

Generate QA samples for a batch of images.

Generate a pattern summary from demonstration examples. Args: task_instruction: The task description input_i

Generate a single batch of samples. Args: batch_config: Dict containing batch configuration: - batch_idx

Get total step count for generation phase based on task type. Local task steps: 1. loading - Loading Documents 2. parsing - P

Create a MLATransformerConfig with common parameters across different model architectures. This is specifically for MLA models like DeepseekV

Get passages (from cache or load fresh). Returns: List of passage texts

Get differences between full prompt and current prompt with surrounding context. This function helps debug tokenization mismatches by showing

Get full qualified name including module and class (if any).

get response from model Args: prompt: str, the prompt kwargs: Arguments for LLM inference - tempe

Get response from VLM model with image inputs. Args: prompts: Single prompt string or list of prompts images

Recursively get type information for nested structures

_summary_ From FSDP json config file, extract the world size. Returns: int: world size

Handle the generating state: generate model response and check for tool calls.

Handle the interacting state: get user input from interaction.

Handle the pending state: prepare the prompt and start generation.

Handle the processing tools state: execute tool calls and prepare tool responses.

Initialize checkpoint manager with proper configuration

Initialize enabled metrics.

Initialize singleton rate limiter.

Initialize singleton rate limiter.

Initialize the worker group by creating new workers from a resource pool. Args: resource_pool: Resource pool for worker allocatio

Initialize the worker group by creating new workers from a resource pool or sub resource pool. Args: resource_pool: Resource pool

Initialize worker engine.

Initialize interactions from configuration. Returns: dict[str, BaseInteraction]: A dictionary mapping interaction names to interac

Inject demo_examples_path into config.

Inject document_dir into config for text.local or image.local task.

Inject image_dir into config for image.local task.

Automatically make a mixed number evalable e.g. 7 3/4 => 7+3/4

Check if a worker is alive. Must be implemented by derived classes.

Load dataloader state from checkpoint

Load dataloader state from checkpoint

Load image data into PIL Image based on detected type. Args: image_data: The image data from the dataset dat

Load image paths from a directory. Args: image_dir: Path to directory containing images Returns: List of abs

Load translation model on first use (lazy loading).

Load all passages from JSONL files in passages directory. Args: passages_dir: Directory path containing JSONL files