Spaces:
Sleeping
Sleeping
File size: 11,647 Bytes
fcf0a07 |
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 |
"""
Memory Manager for Mamba Swarm
Handles memory optimization, caching, and distributed memory management
"""
import torch
import torch.nn as nn
import gc
import psutil
import threading
from typing import Dict, Any, Optional, List, Tuple
from dataclasses import dataclass
from collections import OrderedDict
import numpy as np
import logging
@dataclass
class MemoryStats:
total_memory: float
used_memory: float
free_memory: float
gpu_memory: float
gpu_free: float
cache_size: float
class LRUCache:
"""Least Recently Used cache for model states and activations"""
def __init__(self, max_size: int = 1000):
self.max_size = max_size
self.cache = OrderedDict()
self.lock = threading.Lock()
def get(self, key: str) -> Optional[torch.Tensor]:
with self.lock:
if key in self.cache:
# Move to end (most recently used)
value = self.cache.pop(key)
self.cache[key] = value
return value
return None
def put(self, key: str, value: torch.Tensor):
with self.lock:
if key in self.cache:
self.cache.pop(key)
elif len(self.cache) >= self.max_size:
# Remove least recently used
oldest_key = next(iter(self.cache))
old_value = self.cache.pop(oldest_key)
del old_value
self.cache[key] = value.clone() if isinstance(value, torch.Tensor) else value
def clear(self):
with self.lock:
self.cache.clear()
gc.collect()
class GradientAccumulator:
"""Manages gradient accumulation across multiple steps"""
def __init__(self, accumulation_steps: int = 8):
self.accumulation_steps = accumulation_steps
self.current_step = 0
self.accumulated_gradients = {}
def accumulate(self, model: nn.Module):
"""Accumulate gradients from current backward pass"""
for name, param in model.named_parameters():
if param.grad is not None:
if name not in self.accumulated_gradients:
self.accumulated_gradients[name] = param.grad.clone()
else:
self.accumulated_gradients[name] += param.grad
self.current_step += 1
def should_update(self) -> bool:
"""Check if we should perform optimizer step"""
return self.current_step % self.accumulation_steps == 0
def get_averaged_gradients(self) -> Dict[str, torch.Tensor]:
"""Get accumulated gradients averaged over accumulation steps"""
averaged = {}
for name, grad in self.accumulated_gradients.items():
averaged[name] = grad / self.accumulation_steps
return averaged
def reset(self):
"""Reset accumulator"""
self.accumulated_gradients.clear()
self.current_step = 0
class MemoryManager:
"""Comprehensive memory management for Mamba Swarm"""
def __init__(self,
max_cache_size: int = 2000,
gradient_accumulation_steps: int = 8,
auto_cleanup: bool = True,
memory_threshold: float = 0.85):
self.logger = logging.getLogger(__name__)
self.max_cache_size = max_cache_size
self.gradient_accumulation_steps = gradient_accumulation_steps
self.auto_cleanup = auto_cleanup
self.memory_threshold = memory_threshold
# Initialize components
self.activation_cache = LRUCache(max_cache_size)
self.state_cache = LRUCache(max_cache_size // 2)
self.gradient_accumulator = GradientAccumulator(gradient_accumulation_steps)
# Memory tracking
self.peak_memory_usage = 0.0
self.memory_history = []
self.cleanup_threshold = memory_threshold
# Device management
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
self.setup_memory_optimization()
def setup_memory_optimization(self):
"""Setup memory optimization settings"""
if torch.cuda.is_available():
# Enable memory mapping for large tensors
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
# Set memory fraction
if hasattr(torch.cuda, 'set_per_process_memory_fraction'):
torch.cuda.set_per_process_memory_fraction(0.9)
def get_memory_stats(self) -> MemoryStats:
"""Get current memory statistics"""
# System memory
memory = psutil.virtual_memory()
total_memory = memory.total / (1024**3) # GB
used_memory = memory.used / (1024**3)
free_memory = memory.available / (1024**3)
# GPU memory
gpu_memory = 0.0
gpu_free = 0.0
if torch.cuda.is_available():
gpu_memory = torch.cuda.memory_allocated() / (1024**3)
gpu_free = (torch.cuda.memory_reserved() - torch.cuda.memory_allocated()) / (1024**3)
# Cache size estimation
cache_size = (len(self.activation_cache.cache) + len(self.state_cache.cache)) * 0.001 # Rough estimate
stats = MemoryStats(
total_memory=total_memory,
used_memory=used_memory,
free_memory=free_memory,
gpu_memory=gpu_memory,
gpu_free=gpu_free,
cache_size=cache_size
)
# Update peak usage
current_usage = used_memory + gpu_memory
if current_usage > self.peak_memory_usage:
self.peak_memory_usage = current_usage
return stats
def check_memory_pressure(self) -> bool:
"""Check if system is under memory pressure"""
stats = self.get_memory_stats()
memory_usage_ratio = stats.used_memory / stats.total_memory
if torch.cuda.is_available():
gpu_usage_ratio = stats.gpu_memory / (stats.gpu_memory + stats.gpu_free + 1e-6)
return memory_usage_ratio > self.cleanup_threshold or gpu_usage_ratio > self.cleanup_threshold
return memory_usage_ratio > self.cleanup_threshold
def cleanup_memory(self, aggressive: bool = False):
"""Perform memory cleanup"""
if aggressive:
self.activation_cache.clear()
self.state_cache.clear()
self.gradient_accumulator.reset()
# Python garbage collection
gc.collect()
# GPU memory cleanup
if torch.cuda.is_available():
torch.cuda.empty_cache()
torch.cuda.synchronize()
self.logger.info(f"Memory cleanup completed. Aggressive: {aggressive}")
def cache_activation(self, key: str, activation: torch.Tensor):
"""Cache activation with memory pressure check"""
if self.auto_cleanup and self.check_memory_pressure():
self.cleanup_memory()
self.activation_cache.put(key, activation)
def get_cached_activation(self, key: str) -> Optional[torch.Tensor]:
"""Retrieve cached activation"""
return self.activation_cache.get(key)
def cache_hidden_state(self, key: str, state: torch.Tensor):
"""Cache hidden state"""
self.state_cache.put(key, state)
def get_cached_state(self, key: str) -> Optional[torch.Tensor]:
"""Retrieve cached hidden state"""
return self.state_cache.get(key)
def manage_gradient_accumulation(self, model: nn.Module) -> bool:
"""Manage gradient accumulation and return if optimizer step should be taken"""
self.gradient_accumulator.accumulate(model)
if self.gradient_accumulator.should_update():
# Apply accumulated gradients
averaged_grads = self.gradient_accumulator.get_averaged_gradients()
for name, param in model.named_parameters():
if name in averaged_grads:
param.grad = averaged_grads[name]
self.gradient_accumulator.reset()
return True
return False
def optimize_model_memory(self, model: nn.Module):
"""Optimize model memory usage"""
# Enable gradient checkpointing for large models
for module in model.modules():
if hasattr(module, 'gradient_checkpointing'):
module.gradient_checkpointing = True
# Convert to half precision if possible
if torch.cuda.is_available() and torch.cuda.get_device_capability()[0] >= 7:
model = model.half()
return model
def create_memory_efficient_dataloader(self, dataset, batch_size: int, **kwargs):
"""Create memory-efficient dataloader"""
# Adjust batch size based on available memory
stats = self.get_memory_stats()
if stats.free_memory < 2.0: # Less than 2GB free
batch_size = max(1, batch_size // 2)
self.logger.warning(f"Reduced batch size to {batch_size} due to low memory")
return torch.utils.data.DataLoader(
dataset,
batch_size=batch_size,
num_workers=min(4, psutil.cpu_count()),
pin_memory=torch.cuda.is_available(),
prefetch_factor=2,
**kwargs
)
def monitor_memory_usage(self):
"""Monitor and log memory usage"""
stats = self.get_memory_stats()
self.memory_history.append({
'timestamp': torch.cuda.Event(enable_timing=True) if torch.cuda.is_available() else None,
'stats': stats
})
# Keep only recent history
if len(self.memory_history) > 100:
self.memory_history = self.memory_history[-50:]
self.logger.debug(f"Memory - System: {stats.used_memory:.2f}GB/{stats.total_memory:.2f}GB, "
f"GPU: {stats.gpu_memory:.2f}GB, Cache: {stats.cache_size:.2f}GB")
def get_memory_report(self) -> Dict[str, Any]:
"""Generate comprehensive memory report"""
stats = self.get_memory_stats()
return {
'current_stats': stats.__dict__,
'peak_usage': self.peak_memory_usage,
'cache_stats': {
'activation_cache_size': len(self.activation_cache.cache),
'state_cache_size': len(self.state_cache.cache),
'max_cache_size': self.max_cache_size
},
'gradient_accumulation': {
'current_step': self.gradient_accumulator.current_step,
'accumulation_steps': self.gradient_accumulation_steps,
'accumulated_params': len(self.gradient_accumulator.accumulated_gradients)
},
'memory_pressure': self.check_memory_pressure(),
'device': str(self.device)
}
def __enter__(self):
"""Context manager entry"""
return self
def __exit__(self, exc_type, exc_val, exc_tb):
"""Context manager exit with cleanup"""
self.cleanup_memory(aggressive=True) |