File size: 13,344 Bytes
3cb0dc4 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 |
"""
Model Optimization Module
Advanced optimizations for ChatterboxTTS models including mixed precision, caching, and memory management
"""
import torch
import logging
import time
import hashlib
from typing import Dict, Any, Optional, Tuple
from pathlib import Path
import gc
class MixedPrecisionOptimizer:
"""Mixed precision optimization for faster inference"""
def __init__(self, enabled=True):
self.enabled = enabled and torch.cuda.is_available()
self.autocast_context = None
self.logger = logging.getLogger(__name__)
if self.enabled:
self.logger.info("π Mixed precision optimization enabled")
else:
self.logger.info("π Mixed precision optimization disabled")
def __enter__(self):
if self.enabled:
self.autocast_context = torch.autocast(device_type='cuda', dtype=torch.float16)
return self.autocast_context.__enter__()
return None
def __exit__(self, exc_type, exc_val, exc_tb):
if self.enabled and self.autocast_context:
return self.autocast_context.__exit__(exc_type, exc_val, exc_tb)
return False
class TextTokenizationCache:
"""Smart caching system for text tokenization"""
def __init__(self, max_size=1000):
self.cache = {}
self.max_size = max_size
self.hit_count = 0
self.miss_count = 0
self.logger = logging.getLogger(__name__)
def _get_cache_key(self, text: str, params: Dict = None) -> str:
"""Generate cache key from text and parameters"""
key_str = text
if params:
# Include relevant parameters that affect tokenization
param_str = str(sorted(params.items()))
key_str += f"|{param_str}"
return hashlib.md5(key_str.encode()).hexdigest()
def get(self, text: str, params: Dict = None) -> Optional[Any]:
"""Get cached tokenization result"""
key = self._get_cache_key(text, params)
if key in self.cache:
self.hit_count += 1
self.logger.debug(f"π Tokenization cache HIT for: '{text[:30]}...'")
return self.cache[key]
self.miss_count += 1
self.logger.debug(f"π Tokenization cache MISS for: '{text[:30]}...'")
return None
def put(self, text: str, result: Any, params: Dict = None):
"""Store tokenization result in cache"""
if len(self.cache) >= self.max_size:
# Simple LRU eviction - remove oldest entry
oldest_key = next(iter(self.cache))
del self.cache[oldest_key]
self.logger.debug("ποΈ Evicted oldest cache entry")
key = self._get_cache_key(text, params)
self.cache[key] = result
self.logger.debug(f"πΎ Cached tokenization for: '{text[:30]}...'")
def get_stats(self) -> Dict[str, Any]:
"""Get cache performance statistics"""
total_requests = self.hit_count + self.miss_count
hit_rate = (self.hit_count / total_requests * 100) if total_requests > 0 else 0
return {
'hit_count': self.hit_count,
'miss_count': self.miss_count,
'hit_rate': hit_rate,
'cache_size': len(self.cache),
'max_size': self.max_size
}
def clear(self):
"""Clear the cache"""
self.cache.clear()
self.hit_count = 0
self.miss_count = 0
self.logger.info("ποΈ Tokenization cache cleared")
class MemoryOptimizer:
"""Advanced memory optimization strategies"""
def __init__(self):
self.logger = logging.getLogger(__name__)
self.memory_threshold_gb = 6.0 # Memory threshold for aggressive cleanup
def optimize_memory_usage(self, aggressive=False):
"""Optimize memory usage with various strategies"""
initial_memory = self.get_gpu_memory_usage()
# Standard cleanup
gc.collect()
if torch.cuda.is_available():
torch.cuda.empty_cache()
torch.cuda.synchronize()
if aggressive:
# More aggressive cleanup
if torch.cuda.is_available():
torch.cuda.ipc_collect()
# Clear unused memory
torch.cuda.reset_peak_memory_stats()
final_memory = self.get_gpu_memory_usage()
freed_memory = initial_memory - final_memory
if freed_memory > 0.1: # More than 100MB freed
self.logger.info(f"ποΈ Memory cleanup freed {freed_memory:.2f}GB")
return freed_memory
def get_gpu_memory_usage(self) -> float:
"""Get current GPU memory usage in GB"""
if torch.cuda.is_available():
return torch.cuda.memory_allocated() / 1024**3
return 0.0
def should_optimize_memory(self) -> bool:
"""Check if memory optimization is needed"""
current_memory = self.get_gpu_memory_usage()
return current_memory > self.memory_threshold_gb
def monitor_memory_usage(self, operation_name: str = ""):
"""Context manager for monitoring memory usage during operations"""
return MemoryMonitor(self, operation_name)
class MemoryMonitor:
"""Context manager for monitoring memory usage"""
def __init__(self, optimizer: MemoryOptimizer, operation_name: str):
self.optimizer = optimizer
self.operation_name = operation_name
self.start_memory = 0.0
def __enter__(self):
self.start_memory = self.optimizer.get_gpu_memory_usage()
return self
def __exit__(self, exc_type, exc_val, exc_tb):
end_memory = self.optimizer.get_gpu_memory_usage()
memory_diff = end_memory - self.start_memory
if memory_diff > 0.1: # Significant memory increase
self.optimizer.logger.info(
f"π {self.operation_name}: +{memory_diff:.2f}GB memory usage"
)
elif memory_diff < -0.1: # Significant memory decrease
self.optimizer.logger.info(
f"π {self.operation_name}: {abs(memory_diff):.2f}GB memory freed"
)
class InferenceOptimizer:
"""Optimizations specifically for inference performance"""
def __init__(self):
self.logger = logging.getLogger(__name__)
self.setup_inference_optimizations()
def setup_inference_optimizations(self):
"""Setup general inference optimizations"""
# Disable gradient computation globally for inference
torch.set_grad_enabled(False)
# Enable cuDNN benchmarking for consistent input sizes
if torch.backends.cudnn.is_available():
torch.backends.cudnn.benchmark = True
torch.backends.cudnn.deterministic = False
self.logger.info("π cuDNN benchmark mode enabled")
# Enable optimized attention if available
if hasattr(torch.backends.cuda, 'enable_flash_sdp'):
torch.backends.cuda.enable_flash_sdp(True)
self.logger.info("β‘ Flash attention enabled")
def optimize_model_for_inference(self, model):
"""Apply inference-specific optimizations to a model"""
optimizations_applied = []
# Set model to eval mode if possible
if hasattr(model, 'eval'):
model.eval()
optimizations_applied.append("eval_mode")
# Apply eval mode to submodules
if hasattr(model, 't3') and hasattr(model.t3, 'eval'):
model.t3.eval()
optimizations_applied.append("t3_eval")
if hasattr(model, 's3gen') and hasattr(model.s3gen, 'eval'):
model.s3gen.eval()
optimizations_applied.append("s3gen_eval")
# Enable inference mode context
if hasattr(torch, 'inference_mode'):
# Note: This would need to be applied as a context manager
optimizations_applied.append("inference_mode_available")
self.logger.info(f"π§ Applied inference optimizations: {', '.join(optimizations_applied)}")
return model
class OptimizedTTSWrapper:
"""Wrapper that applies all optimizations to TTS operations"""
def __init__(self, model, enable_mixed_precision=True, enable_caching=True):
self.model = model
self.mixed_precision = MixedPrecisionOptimizer(enable_mixed_precision)
self.tokenization_cache = TextTokenizationCache() if enable_caching else None
self.memory_optimizer = MemoryOptimizer()
self.inference_optimizer = InferenceOptimizer()
self.logger = logging.getLogger(__name__)
# Apply inference optimizations to the model
self.inference_optimizer.optimize_model_for_inference(self.model)
# Performance tracking
self.stats = {
'optimized_generations': 0,
'cache_hits': 0,
'memory_optimizations': 0,
'total_time_saved': 0.0
}
def generate(self, text: str, **kwargs) -> torch.Tensor:
"""Optimized generation with all optimizations applied"""
start_time = time.time()
# Check if memory optimization is needed
if self.memory_optimizer.should_optimize_memory():
freed = self.memory_optimizer.optimize_memory_usage(aggressive=True)
if freed > 0:
self.stats['memory_optimizations'] += 1
# Use mixed precision if enabled
with self.mixed_precision:
with self.memory_optimizer.monitor_memory_usage(f"TTS generation"):
# Check cache for identical requests
cache_key = None
if self.tokenization_cache:
cache_key = f"{text}|{str(sorted(kwargs.items()))}"
cached_result = self.tokenization_cache.get(text, kwargs)
if cached_result is not None:
self.stats['cache_hits'] += 1
self.logger.debug("π Using cached generation result")
return cached_result
# Generate audio with optimizations
with torch.inference_mode() if hasattr(torch, 'inference_mode') else torch.no_grad():
audio = self.model.generate(text, **kwargs)
# Cache the result if caching is enabled
if self.tokenization_cache and cache_key:
self.tokenization_cache.put(text, audio, kwargs)
generation_time = time.time() - start_time
self.stats['optimized_generations'] += 1
# Estimate time saved (rough heuristic)
estimated_baseline_time = generation_time * 1.15 # Assume 15% overhead without optimizations
time_saved = estimated_baseline_time - generation_time
self.stats['total_time_saved'] += time_saved
return audio
def prepare_conditionals(self, *args, **kwargs):
"""Optimized voice conditioning preparation"""
with self.mixed_precision:
with self.memory_optimizer.monitor_memory_usage("Voice conditioning"):
return self.model.prepare_conditionals(*args, **kwargs)
def get_optimization_stats(self) -> Dict[str, Any]:
"""Get comprehensive optimization statistics"""
stats = self.stats.copy()
if self.tokenization_cache:
cache_stats = self.tokenization_cache.get_stats()
stats.update({
'cache_hit_rate': cache_stats['hit_rate'],
'cache_size': cache_stats['cache_size']
})
stats['current_memory_gb'] = self.memory_optimizer.get_gpu_memory_usage()
return stats
def log_performance_summary(self):
"""Log comprehensive performance summary"""
stats = self.get_optimization_stats()
self.logger.info("π OPTIMIZATION PERFORMANCE SUMMARY")
self.logger.info("=" * 50)
self.logger.info(f"Optimized generations: {stats['optimized_generations']}")
self.logger.info(f"Memory optimizations: {stats['memory_optimizations']}")
self.logger.info(f"Total time saved: {stats['total_time_saved']:.2f}s")
if 'cache_hit_rate' in stats:
self.logger.info(f"Cache hit rate: {stats['cache_hit_rate']:.1f}%")
self.logger.info(f"Cache size: {stats['cache_size']} entries")
self.logger.info(f"Current memory usage: {stats['current_memory_gb']:.2f}GB")
# Calculate estimated performance improvement
if stats['optimized_generations'] > 0:
avg_time_saved = stats['total_time_saved'] / stats['optimized_generations']
self.logger.info(f"Average time saved per generation: {avg_time_saved:.2f}s")
def create_optimized_model(model, enable_mixed_precision=True, enable_caching=True):
"""Create an optimized wrapper around a ChatterboxTTS model"""
return OptimizedTTSWrapper(
model,
enable_mixed_precision=enable_mixed_precision,
enable_caching=enable_caching
) |