scripts/split_dataset.py

import os
import sys
import argparse
import math

sys.path.append(
    os.path.abspath(os.path.join(os.path.dirname(__file__), os.path.pardir))
)

from megatron.core.datasets.indexed_dataset import (
    IndexedDataset,
    IndexedDatasetBuilder,
    get_bin_path,
    get_idx_path,
)


def get_args():
    parser = argparse.ArgumentParser()

    group = parser.add_argument_group(title="input data")
    group.add_argument(
        "--input-prefix",
        type=str,
        required=True,
        help="Path to binary input file without suffix",
    )

    group = parser.add_argument_group(title="output data")
    group.add_argument(
        "--output-dir",
        type=str,
        required=True,
        help="Directory to output split files",
    )
    group.add_argument(
        "--output-prefix",
        type=str,
        default="split",
        help="Prefix for output files (default: split)",
    )

    group = parser.add_argument_group(title="split options")
    group.add_argument(
        "--num-splits",
        type=int,
        default=None,
        help="Number of splits to create. If not provided, will be determined by max-split-size-gb",
    )
    group.add_argument(
        "--max-split-size-gb",
        type=float,
        default=40.0,
        help="Maximum size of each split in GB (default: 40.0)",
    )
    group.add_argument(
        "--split-by-documents",
        action="store_true",
        help="Split by documents instead of sequences (default: split by sequences)",
    )

    group = parser.add_argument_group(title="miscellaneous")
    group.add_argument(
        "--multimodal",
        action="store_true",
        help="Whether the dataset is assumed to be multimodal"
    )

    args = parser.parse_args()

    # Check input file exists
    bin_path = get_bin_path(args.input_prefix)
    idx_path = get_idx_path(args.input_prefix)
    assert os.path.isfile(bin_path), f"ERROR: {bin_path} does not exist"
    assert os.path.isfile(idx_path), f"ERROR: {idx_path} does not exist"

    # Check output directory exists
    assert os.path.isdir(args.output_dir), f"ERROR: {args.output_dir} is not a directory or does not exist"

    return args


def split_by_sequences(dataset, output_dir, output_prefix, multimodal, max_split_size_bytes, num_splits=None):
    """Split dataset by sequences, respecting max_split_size_bytes."""
    total_sequences = len(dataset)
    if total_sequences == 0:
        print("Warning: No sequences found in dataset")
        return

    print(f"Total sequences: {total_sequences}")
    
    split_idx = 0
    start_seq_idx = 0

    while start_seq_idx < total_sequences:
        print(f"Creating split {split_idx + 1}...")
        
        # Create output paths
        split_prefix = os.path.join(output_dir, f"{output_prefix}_{split_idx:03d}")
        bin_path = get_bin_path(split_prefix)
        idx_path = get_idx_path(split_prefix)
        
        # Create builder
        builder = IndexedDatasetBuilder(bin_path, dtype=dataset.index.dtype, multimodal=multimodal)
        
        current_split_size = 0
        sequences_in_split = 0
        
        # Determine target number of sequences for this split if num_splits is provided
        if num_splits is not None:
            sequences_per_split = math.ceil(total_sequences / num_splits)
            end_seq_idx_target = min(start_seq_idx + sequences_per_split, total_sequences)
        else:
            end_seq_idx_target = total_sequences

        for seq_idx in range(start_seq_idx, end_seq_idx_target):
            sequence_pointer, sequence_length, sequence_mode = dataset.index[seq_idx]
            sequence_size = sequence_length * dataset.index.dtype_size
            
            if sequences_in_split > 0 and current_split_size + sequence_size > max_split_size_bytes:
                break
                
            sequence = dataset.bin_reader.read(
                dtype=dataset.index.dtype, count=sequence_length, offset=sequence_pointer
            )
            
            import torch
            tensor = torch.from_numpy(sequence.copy())
            mode = sequence_mode if multimodal else 0
            builder.add_item(tensor, mode)
            
            current_split_size += sequence_size
            sequences_in_split += 1
        
        # Finalize the split
        builder.finalize(idx_path)
        end_seq_idx = start_seq_idx + sequences_in_split
        print(f"Split {split_idx + 1} completed: sequences {start_seq_idx} to {end_seq_idx - 1} ({sequences_in_split} sequences), size: {current_split_size / (1024**3):.2f} GB")
        
        start_seq_idx = end_seq_idx
        split_idx += 1


def split_by_documents(dataset, output_dir, output_prefix, multimodal, max_split_size_bytes, num_splits=None):
    """Split dataset by documents, respecting max_split_size_bytes."""
    document_indices = dataset.document_indices
    total_documents = len(document_indices) - 1
    
    if total_documents == 0:
        print("Warning: No documents found in dataset")
        return
        
    print(f"Total documents: {total_documents}")
    
    split_idx = 0
    start_doc_idx = 0

    while start_doc_idx < total_documents:
        print(f"Creating split {split_idx + 1}...")
        
        split_prefix = os.path.join(output_dir, f"{output_prefix}_{split_idx:03d}")
        bin_path = get_bin_path(split_prefix)
        idx_path = get_idx_path(split_prefix)
        
        builder = IndexedDatasetBuilder(bin_path, dtype=dataset.index.dtype, multimodal=multimodal)
        
        current_split_size = 0
        documents_in_split = 0
        
        if num_splits is not None:
            docs_per_split = math.ceil(total_documents / num_splits)
            end_doc_idx_target = min(start_doc_idx + docs_per_split, total_documents)
        else:
            end_doc_idx_target = total_documents
            
        for doc_idx in range(start_doc_idx, end_doc_idx_target):
            doc_start_seq = document_indices[doc_idx]
            doc_end_seq = document_indices[doc_idx + 1]
            
            doc_size = 0
            for seq_idx in range(doc_start_seq, doc_end_seq):
                _, sequence_length, _ = dataset.index[seq_idx]
                doc_size += sequence_length * dataset.index.dtype_size
            
            if documents_in_split > 0 and current_split_size + doc_size > max_split_size_bytes:
                break

            for seq_idx in range(doc_start_seq, doc_end_seq):
                sequence_pointer, sequence_length, sequence_mode = dataset.index[seq_idx]
                sequence = dataset.bin_reader.read(
                    dtype=dataset.index.dtype, count=sequence_length, offset=sequence_pointer
                )
                
                import torch
                tensor = torch.from_numpy(sequence.copy())
                mode = sequence_mode if multimodal else 0
                builder.add_item(tensor, mode)
            
            builder.end_document()
            current_split_size += doc_size
            documents_in_split += 1
            
        builder.finalize(idx_path)
        end_doc_idx = start_doc_idx + documents_in_split
        print(f"Split {split_idx + 1} completed: documents {start_doc_idx} to {end_doc_idx - 1} ({documents_in_split} documents), size: {current_split_size / (1024**3):.2f} GB")
        
        start_doc_idx = end_doc_idx
        split_idx += 1


def main():
    args = get_args()
    
    print(f"Loading dataset from {args.input_prefix}")
    dataset = IndexedDataset(args.input_prefix, multimodal=args.multimodal)
    
    print(f"Dataset loaded: {len(dataset)} sequences")
    if args.multimodal:
        print(f"Multimodal dataset with {len(dataset.document_indices) - 1} documents")
    else:
        print(f"Standard dataset with {len(dataset.document_indices) - 1} documents")

    max_split_size_bytes = args.max_split_size_gb * 1024 * 1024 * 1024
    
    # If num_splits is provided, check if it respects the max size.
    if args.num_splits is not None:
        input_bin_path = get_bin_path(args.input_prefix)
        total_size_bytes = os.path.getsize(input_bin_path)
        size_per_split = total_size_bytes / args.num_splits
        if size_per_split > max_split_size_bytes:
            print(f"Warning: With {args.num_splits} splits, the average split size would be {size_per_split / (1024**3):.2f} GB, which is larger than the specified max of {args.max_split_size_gb} GB.")
            print("The script will create more splits if necessary to respect the size limit.")
            
    if args.split_by_documents:
        split_by_documents(dataset, args.output_dir, args.output_prefix, args.multimodal, max_split_size_bytes, args.num_splits)
    else:
        split_by_sequences(dataset, args.output_dir, args.output_prefix, args.multimodal, max_split_size_bytes, args.num_splits)
    
    print("Dataset splitting completed!")


if __name__ == '__main__':
    main()