Spaces:
Paused
Paused
File size: 10,916 Bytes
fb867c3 |
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 |
"""
GPU resource management for ZeroGPU integration.
This module provides GPU resource management specifically designed
for deployment on Hugging Face Spaces with ZeroGPU support.
Key Features:
- Automatic GPU detection and allocation
- Resource limits for fair sharing
- Memory management and cleanup
- Fallback to CPU when GPU unavailable
"""
import os
import time
import threading
from typing import Optional, Dict, Any, List
from contextlib import contextmanager
from dataclasses import dataclass
import logging
logger = logging.getLogger(__name__)
@dataclass
class GPUContext:
"""Context for GPU resource allocation."""
device_id: Optional[int]
memory_allocated: int
priority: str
start_time: float
is_gpu: bool
@property
def elapsed_time(self) -> float:
"""Get elapsed time since allocation."""
return time.time() - self.start_time
class GPUResourceManager:
"""
Manages GPU resources for ZeroGPU deployment.
Provides automatic resource management with fallback to CPU
when GPU is unavailable or quota exceeded.
"""
# ZeroGPU typical limits
DEFAULT_TIME_LIMIT = 120 # 2 minutes per request
DEFAULT_MEMORY_LIMIT = 4096 # 4GB VRAM
# Priority levels
PRIORITY_LEVELS = {
"high": {"time_limit": 180, "memory_limit": 6144},
"normal": {"time_limit": 120, "memory_limit": 4096},
"low": {"time_limit": 60, "memory_limit": 2048}
}
def __init__(self,
enable_gpu: bool = True,
auto_cleanup: bool = True,
max_concurrent: int = 3):
"""
Initialize GPU resource manager.
Args:
enable_gpu: Whether to attempt GPU usage
auto_cleanup: Enable automatic resource cleanup
max_concurrent: Maximum concurrent GPU operations
"""
self.enable_gpu = enable_gpu and self._check_gpu_available()
self.auto_cleanup = auto_cleanup
self.max_concurrent = max_concurrent
# Resource tracking
self._lock = threading.Lock()
self._active_contexts: List[GPUContext] = []
self._total_memory_used = 0
self._request_count = 0
# Check for ZeroGPU environment
self.is_zerogpu = self._detect_zerogpu_environment()
if self.is_zerogpu:
logger.info("ZeroGPU environment detected")
self._setup_zerogpu()
elif self.enable_gpu:
logger.info("GPU available but not ZeroGPU environment")
else:
logger.info("Running in CPU-only mode")
def _check_gpu_available(self) -> bool:
"""Check if GPU is available."""
try:
import torch
return torch.cuda.is_available()
except ImportError:
return False
def _detect_zerogpu_environment(self) -> bool:
"""Detect if running in ZeroGPU environment."""
# Check for ZeroGPU-specific environment variables
zerogpu_indicators = [
"SPACES_ZERO_GPU",
"ZEROGPU_AVAILABLE",
"HF_SPACES"
]
for indicator in zerogpu_indicators:
if os.environ.get(indicator):
return True
return False
def _setup_zerogpu(self):
"""Setup ZeroGPU-specific configurations."""
try:
# Import spaces library if available
import spaces
# Register with ZeroGPU
@spaces.GPU
def dummy_gpu_function():
pass
# Test GPU allocation
dummy_gpu_function()
logger.info("ZeroGPU setup successful")
except ImportError:
logger.warning("spaces library not available, ZeroGPU features limited")
except Exception as e:
logger.warning(f"ZeroGPU setup failed: {e}")
@contextmanager
def acquire_resources(self, priority: str = "normal",
memory_required: int = 2048):
"""
Context manager for acquiring GPU resources.
Args:
priority: Priority level (high/normal/low)
memory_required: Required memory in MB
Yields:
GPUContext object with resource allocation details
"""
context = self._allocate_resources(priority, memory_required)
try:
yield context
finally:
self._release_resources(context)
def _allocate_resources(self, priority: str, memory_required: int) -> GPUContext:
"""Allocate GPU resources."""
with self._lock:
# Check if GPU is available and within limits
if not self.enable_gpu or len(self._active_contexts) >= self.max_concurrent:
# Fallback to CPU
context = GPUContext(
device_id=None,
memory_allocated=0,
priority=priority,
start_time=time.time(),
is_gpu=False
)
logger.info(f"Allocated CPU resources (GPU unavailable or limit reached)")
return context
# Check memory limits
limits = self.PRIORITY_LEVELS.get(priority, self.PRIORITY_LEVELS["normal"])
if memory_required > limits["memory_limit"]:
memory_required = limits["memory_limit"]
logger.warning(f"Memory request reduced to {memory_required}MB (priority limit)")
if self._total_memory_used + memory_required > self.DEFAULT_MEMORY_LIMIT * 2:
# Total system limit exceeded, use CPU
context = GPUContext(
device_id=None,
memory_allocated=0,
priority=priority,
start_time=time.time(),
is_gpu=False
)
logger.warning("System memory limit exceeded, falling back to CPU")
return context
# Allocate GPU
device_id = self._get_available_device()
context = GPUContext(
device_id=device_id,
memory_allocated=memory_required,
priority=priority,
start_time=time.time(),
is_gpu=True
)
self._active_contexts.append(context)
self._total_memory_used += memory_required
self._request_count += 1
logger.info(f"Allocated GPU {device_id} with {memory_required}MB for {priority} priority")
# Start cleanup timer if needed
if self.auto_cleanup:
self._schedule_cleanup(context, limits["time_limit"])
return context
def _release_resources(self, context: GPUContext):
"""Release GPU resources."""
with self._lock:
if context in self._active_contexts:
self._active_contexts.remove(context)
self._total_memory_used -= context.memory_allocated
if context.is_gpu:
logger.info(f"Released GPU {context.device_id} ({context.memory_allocated}MB)")
# Clear GPU cache if available
try:
import torch
if torch.cuda.is_available():
torch.cuda.empty_cache()
except:
pass
def _get_available_device(self) -> int:
"""Get available GPU device ID."""
# For ZeroGPU, typically use device 0
# In multi-GPU setups, could implement round-robin
return 0
def _schedule_cleanup(self, context: GPUContext, timeout: int):
"""Schedule automatic cleanup after timeout."""
def cleanup():
time.sleep(timeout)
with self._lock:
if context in self._active_contexts:
logger.warning(f"Auto-cleaning GPU context after {timeout}s timeout")
self._release_resources(context)
thread = threading.Thread(target=cleanup, daemon=True)
thread.start()
def get_device_string(self, context: Optional[GPUContext] = None) -> str:
"""
Get device string for PyTorch/TensorFlow.
Args:
context: GPU context (uses CPU if None or not GPU)
Returns:
Device string like "cuda:0" or "cpu"
"""
if context and context.is_gpu and context.device_id is not None:
return f"cuda:{context.device_id}"
return "cpu"
def get_statistics(self) -> Dict[str, Any]:
"""Get resource usage statistics."""
with self._lock:
stats = {
"gpu_enabled": self.enable_gpu,
"is_zerogpu": self.is_zerogpu,
"active_contexts": len(self._active_contexts),
"total_memory_used_mb": self._total_memory_used,
"total_requests": self._request_count,
"max_concurrent": self.max_concurrent
}
if self._active_contexts:
stats["active_priorities"] = [c.priority for c in self._active_contexts]
stats["oldest_context_age_s"] = max(c.elapsed_time for c in self._active_contexts)
return stats
def cleanup_all(self):
"""Force cleanup of all GPU resources."""
with self._lock:
contexts = list(self._active_contexts)
for context in contexts:
self._release_resources(context)
logger.info("Cleaned up all GPU resources")
def zerogpu_decorator(memory_mb: int = 2048, priority: str = "normal"):
"""
Decorator for functions that need GPU resources.
Usage:
@zerogpu_decorator(memory_mb=4096, priority="high")
def my_gpu_function(data):
# Function automatically gets GPU resources
...
Args:
memory_mb: Required memory in MB
priority: Priority level
"""
def decorator(func):
def wrapper(*args, **kwargs):
manager = GPUResourceManager()
with manager.acquire_resources(priority=priority,
memory_required=memory_mb) as gpu_ctx:
# Inject GPU context into kwargs
kwargs['gpu_context'] = gpu_ctx
return func(*args, **kwargs)
# Preserve function metadata
wrapper.__name__ = func.__name__
wrapper.__doc__ = func.__doc__
return wrapper
return decorator
# Global manager instance for shared use
_global_gpu_manager: Optional[GPUResourceManager] = None
def get_gpu_manager() -> GPUResourceManager:
"""Get or create global GPU manager instance."""
global _global_gpu_manager
if _global_gpu_manager is None:
_global_gpu_manager = GPUResourceManager()
return _global_gpu_manager |