File size: 8,734 Bytes
27871e7 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 |
"""
DataLoader utilities for SLM training.
Provides efficient batching and data loading for training.
"""
import os
from typing import Dict, Optional, List
import torch
from torch.utils.data import DataLoader, Dataset, DistributedSampler
from .dataset import ConversationalDataset, StreamingTextDataset, PackedDataset
from .tokenizer import SLMTokenizer
def create_dataloader(
dataset: Dataset,
batch_size: int,
shuffle: bool = True,
num_workers: int = 4,
pin_memory: bool = None, # Auto-detect based on device
drop_last: bool = True,
distributed: bool = False,
world_size: int = 1,
rank: int = 0,
) -> DataLoader:
"""Create a DataLoader with optimal settings.
Args:
dataset: The dataset to load from
batch_size: Batch size per device
shuffle: Whether to shuffle data
num_workers: Number of data loading workers
pin_memory: Pin memory for faster GPU transfer
drop_last: Drop last incomplete batch
distributed: Whether using distributed training
world_size: Number of distributed processes
rank: Current process rank
Returns:
Configured DataLoader
"""
sampler = None
if distributed:
sampler = DistributedSampler(
dataset,
num_replicas=world_size,
rank=rank,
shuffle=shuffle,
)
shuffle = False # Sampler handles shuffling
# Auto-detect pin_memory: disable for MPS (not supported)
if pin_memory is None:
import torch
pin_memory = torch.cuda.is_available() # Only True for CUDA
return DataLoader(
dataset,
batch_size=batch_size,
shuffle=shuffle if sampler is None else False,
sampler=sampler,
num_workers=num_workers,
pin_memory=pin_memory,
drop_last=drop_last,
collate_fn=default_collate_fn,
)
def default_collate_fn(batch: List[Dict[str, torch.Tensor]]) -> Dict[str, torch.Tensor]:
"""Collate function for batching samples.
Args:
batch: List of sample dictionaries
Returns:
Batched dictionary with stacked tensors
"""
return {
"input_ids": torch.stack([s["input_ids"] for s in batch]),
"attention_mask": torch.stack([s["attention_mask"] for s in batch]),
"labels": torch.stack([s["labels"] for s in batch]),
}
class DataModule:
"""Data module for managing train/val dataloaders.
Provides a unified interface for data loading during training.
"""
def __init__(
self,
data_dir: str,
tokenizer_path: str,
max_length: int = 1024,
batch_size: int = 32,
num_workers: int = 4,
val_batch_size: Optional[int] = None,
):
"""Initialize data module.
Args:
data_dir: Directory containing processed data
tokenizer_path: Path to tokenizer.json
max_length: Maximum sequence length
batch_size: Training batch size
num_workers: Number of data loading workers
val_batch_size: Validation batch size (defaults to batch_size)
"""
self.data_dir = data_dir
self.max_length = max_length
self.batch_size = batch_size
self.val_batch_size = val_batch_size or batch_size
self.num_workers = num_workers
# Load tokenizer
self.tokenizer = SLMTokenizer.from_file(tokenizer_path)
# Datasets (created on first access)
self._train_dataset = None
self._val_dataset = None
@property
def train_dataset(self) -> Dataset:
"""Get or create training dataset."""
if self._train_dataset is None:
self._train_dataset = ConversationalDataset(
data_path=self.data_dir,
tokenizer=self.tokenizer,
max_length=self.max_length,
split="train",
)
return self._train_dataset
@property
def val_dataset(self) -> Dataset:
"""Get or create validation dataset."""
if self._val_dataset is None:
self._val_dataset = ConversationalDataset(
data_path=self.data_dir,
tokenizer=self.tokenizer,
max_length=self.max_length,
split="val",
)
return self._val_dataset
def train_dataloader(
self,
distributed: bool = False,
world_size: int = 1,
rank: int = 0,
) -> DataLoader:
"""Get training dataloader."""
return create_dataloader(
self.train_dataset,
batch_size=self.batch_size,
shuffle=True,
num_workers=self.num_workers,
drop_last=True,
distributed=distributed,
world_size=world_size,
rank=rank,
)
def val_dataloader(self) -> DataLoader:
"""Get validation dataloader."""
return create_dataloader(
self.val_dataset,
batch_size=self.val_batch_size,
shuffle=False,
num_workers=self.num_workers,
drop_last=False,
)
class StreamingDataModule:
"""Data module for streaming large datasets.
Memory-efficient loading for large text corpora.
"""
def __init__(
self,
data_files: List[str],
tokenizer_path: str,
max_length: int = 1024,
batch_size: int = 32,
num_workers: int = 4,
):
"""Initialize streaming data module.
Args:
data_files: List of text file paths
tokenizer_path: Path to tokenizer.json
max_length: Maximum sequence length
batch_size: Batch size
num_workers: Number of data loading workers
"""
self.data_files = data_files
self.max_length = max_length
self.batch_size = batch_size
self.num_workers = num_workers
# Load tokenizer
self.tokenizer = SLMTokenizer.from_file(tokenizer_path)
def train_dataloader(self) -> DataLoader:
"""Get training dataloader for streaming data."""
dataset = StreamingTextDataset(
data_files=self.data_files,
tokenizer=self.tokenizer,
max_length=self.max_length,
shuffle=True,
)
return DataLoader(
dataset,
batch_size=self.batch_size,
num_workers=self.num_workers,
pin_memory=True,
collate_fn=default_collate_fn,
)
def estimate_dataset_tokens(data_dir: str, tokenizer_path: str) -> Dict[str, int]:
"""Estimate total tokens in a dataset.
Args:
data_dir: Directory containing data files
tokenizer_path: Path to tokenizer
Returns:
Dictionary with token counts
"""
import json
from pathlib import Path
tokenizer = SLMTokenizer.from_file(tokenizer_path)
total_tokens = 0
total_samples = 0
for file_path in Path(data_dir).glob("*.json*"):
with open(file_path, "r") as f:
if file_path.suffix == ".jsonl":
samples = [json.loads(line) for line in f if line.strip()]
else:
samples = json.load(f)
if not isinstance(samples, list):
samples = [samples]
for sample in samples:
if "user" in sample and "assistant" in sample:
tokens = tokenizer.encode_conversation(
sample["user"], sample["assistant"]
)
elif "text" in sample:
tokens = tokenizer.encode(sample["text"])
else:
continue
total_tokens += len(tokens)
total_samples += 1
return {
"total_tokens": total_tokens,
"total_samples": total_samples,
"avg_tokens_per_sample": total_tokens / max(total_samples, 1),
}
def get_dataloader_stats(dataloader: DataLoader) -> Dict[str, float]:
"""Get statistics from a dataloader.
Args:
dataloader: The dataloader to analyze
Returns:
Dictionary with statistics
"""
total_batches = 0
total_tokens = 0
total_non_pad_tokens = 0
for batch in dataloader:
total_batches += 1
total_tokens += batch["input_ids"].numel()
total_non_pad_tokens += batch["attention_mask"].sum().item()
# Only sample first 100 batches
if total_batches >= 100:
break
return {
"batches_sampled": total_batches,
"tokens_per_batch": total_tokens / max(total_batches, 1),
"non_pad_ratio": total_non_pad_tokens / max(total_tokens, 1),
}
|