Spaces:
Running
Running
File size: 12,496 Bytes
ab4e093 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 |
"""
Advanced CPU Optimizer for training on CPU-only systems
Optimized for maximum performance on limited hardware
"""
import os
import logging
import threading
from typing import Dict, Any, Optional, List
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader
import numpy as np
from .memory_manager import AdvancedMemoryManager
logger = logging.getLogger(__name__)
class CPUOptimizer:
"""
Advanced CPU optimization for training and inference
"""
def __init__(self, memory_manager: AdvancedMemoryManager):
"""
Initialize CPU optimizer
Args:
memory_manager: Memory manager instance
"""
self.memory_manager = memory_manager
self.cpu_count = os.cpu_count()
self.optimizations_applied = []
# Apply initial optimizations
self._apply_global_optimizations()
logger.info(f"CPU Optimizer initialized for {self.cpu_count} cores")
def _apply_global_optimizations(self):
"""Apply global CPU optimizations"""
# Set optimal thread count for PyTorch
optimal_threads = min(self.cpu_count, 8) # Cap at 8 for stability
torch.set_num_threads(optimal_threads)
self.optimizations_applied.append(f"PyTorch threads: {optimal_threads}")
# Set thread count for inter-op parallelism
torch.set_num_interop_threads(min(self.cpu_count // 2, 4))
self.optimizations_applied.append("Inter-op parallelism configured")
# Enable Intel MKL optimizations if available
try:
import intel_extension_for_pytorch as ipex
self.optimizations_applied.append("Intel Extension for PyTorch enabled")
except ImportError:
logger.warning("Intel Extension for PyTorch not available")
# Set environment variables for CPU optimization
os.environ['OMP_NUM_THREADS'] = str(optimal_threads)
os.environ['MKL_NUM_THREADS'] = str(optimal_threads)
os.environ['NUMEXPR_NUM_THREADS'] = str(optimal_threads)
os.environ['OPENBLAS_NUM_THREADS'] = str(optimal_threads)
self.optimizations_applied.append("Environment variables optimized")
# Enable CPU-specific optimizations
torch.backends.mkl.enabled = True
torch.backends.mkldnn.enabled = True
self.optimizations_applied.append("MKL and MKLDNN enabled")
logger.info(f"Applied optimizations: {', '.join(self.optimizations_applied)}")
def optimize_model(self, model: nn.Module,
use_jit: bool = True,
use_channels_last: bool = True) -> nn.Module:
"""
Optimize model for CPU inference/training
Args:
model: PyTorch model to optimize
use_jit: Whether to use TorchScript JIT compilation
use_channels_last: Whether to use channels-last memory format
Returns:
Optimized model
"""
with self.memory_manager.memory_context("optimize_model"):
logger.info("Optimizing model for CPU")
# Set model to CPU
model = model.cpu()
# Set to evaluation mode for optimization
was_training = model.training
model.eval()
try:
# Apply Intel Extension optimizations if available
try:
import intel_extension_for_pytorch as ipex
model = ipex.optimize(model, dtype=torch.float32)
logger.info("Applied Intel Extension optimizations")
except ImportError:
pass
# Apply channels-last memory format for conv models
if use_channels_last and self._has_conv_layers(model):
model = model.to(memory_format=torch.channels_last)
logger.info("Applied channels-last memory format")
# Apply TorchScript JIT compilation
if use_jit:
try:
# Create dummy input for tracing
dummy_input = self._create_dummy_input(model)
if dummy_input is not None:
model = torch.jit.trace(model, dummy_input)
logger.info("Applied TorchScript JIT compilation")
except Exception as e:
logger.warning(f"JIT compilation failed: {e}")
# Restore training mode if needed
if was_training:
model.train()
return model
except Exception as e:
logger.error(f"Model optimization failed: {e}")
return model
def _has_conv_layers(self, model: nn.Module) -> bool:
"""Check if model has convolutional layers"""
for module in model.modules():
if isinstance(module, (nn.Conv1d, nn.Conv2d, nn.Conv3d)):
return True
return False
def _create_dummy_input(self, model: nn.Module) -> Optional[torch.Tensor]:
"""Create dummy input for model tracing"""
try:
# Try to infer input shape from model
for name, param in model.named_parameters():
if 'embedding' in name.lower() and param.dim() == 2:
# Text model - create token input
vocab_size = param.shape[0]
return torch.randint(0, min(vocab_size, 1000), (1, 32))
elif 'conv' in name.lower() and param.dim() == 4:
# Vision model - create image input
channels = param.shape[1]
return torch.randn(1, channels, 224, 224)
# Default fallback
return torch.randn(1, 512)
except Exception:
return None
def optimize_dataloader(self, dataloader: DataLoader) -> DataLoader:
"""
Optimize DataLoader for CPU training
Args:
dataloader: Original DataLoader
Returns:
Optimized DataLoader
"""
# Calculate optimal number of workers
optimal_workers = min(self.cpu_count // 2, 4)
# Create new DataLoader with optimized settings
optimized_loader = DataLoader(
dataloader.dataset,
batch_size=dataloader.batch_size,
shuffle=dataloader.drop_last if hasattr(dataloader, 'drop_last') else False,
num_workers=optimal_workers,
pin_memory=False, # Not needed for CPU
persistent_workers=True if optimal_workers > 0 else False,
prefetch_factor=2 if optimal_workers > 0 else 2,
)
logger.info(f"Optimized DataLoader with {optimal_workers} workers")
return optimized_loader
def optimize_optimizer(self, optimizer: optim.Optimizer,
model: nn.Module) -> optim.Optimizer:
"""
Optimize optimizer settings for CPU training
Args:
optimizer: PyTorch optimizer
model: Model being optimized
Returns:
Optimized optimizer
"""
# Apply gradient clipping
for param_group in optimizer.param_groups:
if 'weight_decay' not in param_group:
param_group['weight_decay'] = 0.01
logger.info("Applied optimizer optimizations")
return optimizer
def enable_mixed_precision(self) -> bool:
"""
Enable mixed precision training for CPU (if supported)
Returns:
Whether mixed precision was enabled
"""
try:
# Check if CPU supports mixed precision
if torch.cpu.amp.autocast is not None:
logger.info("CPU mixed precision available")
return True
except AttributeError:
pass
logger.warning("CPU mixed precision not available")
return False
def optimize_batch_size(self, base_batch_size: int,
model_size_mb: float) -> int:
"""
Calculate optimal batch size based on available memory
Args:
base_batch_size: Base batch size to start from
model_size_mb: Model size in MB
Returns:
Optimized batch size
"""
memory_info = self.memory_manager.get_memory_info()
available_memory_mb = memory_info['system_memory_available_gb'] * 1024
# Reserve memory for model and overhead
usable_memory_mb = available_memory_mb - model_size_mb - 2000 # 2GB overhead
# Estimate memory per sample (rough approximation)
memory_per_sample_mb = model_size_mb * 0.1 # 10% of model size per sample
if memory_per_sample_mb > 0:
max_batch_size = int(usable_memory_mb / memory_per_sample_mb)
optimal_batch_size = min(base_batch_size, max_batch_size, 32) # Cap at 32
else:
optimal_batch_size = min(base_batch_size, 8) # Conservative fallback
optimal_batch_size = max(1, optimal_batch_size) # At least 1
logger.info(f"Optimized batch size: {optimal_batch_size} (was {base_batch_size})")
return optimal_batch_size
def get_performance_recommendations(self, model: nn.Module) -> List[str]:
"""
Get performance recommendations for the current setup
Args:
model: Model to analyze
Returns:
List of recommendations
"""
recommendations = []
# Check model size
param_count = sum(p.numel() for p in model.parameters())
model_size_mb = param_count * 4 / (1024**2) # Assume float32
if model_size_mb > 2000: # > 2GB
recommendations.append("Consider using model sharding for large models")
recommendations.append("Use gradient checkpointing to reduce memory usage")
# Check CPU utilization
if self.cpu_count > 8:
recommendations.append("Consider using distributed training across CPU cores")
# Check memory
memory_info = self.memory_manager.get_memory_info()
if memory_info['system_memory_percent'] > 80:
recommendations.append("Reduce batch size to lower memory usage")
recommendations.append("Enable gradient accumulation instead of large batches")
# Check for optimization opportunities
if not any('Intel Extension' in opt for opt in self.optimizations_applied):
recommendations.append("Install Intel Extension for PyTorch for better CPU performance")
return recommendations
def benchmark_performance(self, model: nn.Module,
input_shape: tuple,
num_iterations: int = 100) -> Dict[str, float]:
"""
Benchmark model performance
Args:
model: Model to benchmark
input_shape: Input tensor shape
num_iterations: Number of iterations to run
Returns:
Performance metrics
"""
model.eval()
dummy_input = torch.randn(*input_shape)
# Warmup
with torch.no_grad():
for _ in range(10):
_ = model(dummy_input)
# Benchmark
import time
start_time = time.time()
with torch.no_grad():
for _ in range(num_iterations):
_ = model(dummy_input)
end_time = time.time()
total_time = end_time - start_time
avg_time_per_inference = total_time / num_iterations
throughput = 1.0 / avg_time_per_inference
return {
'total_time_seconds': total_time,
'avg_time_per_inference_ms': avg_time_per_inference * 1000,
'throughput_inferences_per_second': throughput,
'iterations': num_iterations
}
|