File size: 5,626 Bytes
8ae78b0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import time
import functools
import logging
import os
import psutil
import gc
try:
    from app.utils.logging_utils import setup_logger
except ImportError:
    # Try relative imports for running from project root
    from behavior_backend.app.utils.logging_utils import setup_logger

# Configure logging
logger = setup_logger(__name__)

def get_system_memory_info():
    """
    Get system memory information.
    
    Returns:
        dict: Memory information
    """
    memory = psutil.virtual_memory()
    return {
        "total": memory.total / (1024 ** 3),  # GB
        "available": memory.available / (1024 ** 3),  # GB
        "percent_used": memory.percent,
        "process_usage": psutil.Process(os.getpid()).memory_info().rss / (1024 ** 3)  # GB
    }

def log_memory_usage(message=""):
    """
    Log current memory usage.
    
    Args:
        message: Optional message to include in the log
    """
    mem_info = get_system_memory_info()
    logger.info(f"Memory usage {message}: "
                f"Total: {mem_info['total']:.2f}GB, "
                f"Available: {mem_info['available']:.2f}GB, "
                f"Used: {mem_info['percent_used']}%, "
                f"Process: {mem_info['process_usage']:.2f}GB")

def get_available_device():
    """
    Determine the best available device with proper error handling.
    
    Returns:
        str: 'cuda', 'mps', or 'cpu' depending on availability
    """
    logger.info("=== GPU DETECTION ===")
    
    # Check available memory first
    mem_info = get_system_memory_info()
    if mem_info['available'] < 2.0:  # Less than 2GB available
        logger.warning(f"Low system memory: {mem_info['available']:.2f}GB available. Forcing CPU usage.")
        return "cpu"
    
    # First try CUDA (NVIDIA GPUs)
    if torch.cuda.is_available():
        try:
            # Simplified CUDA test with better error handling
            logger.info("CUDA detected - attempting verification")
            # Use a smaller and simpler operation
            test_tensor = torch.tensor([1.0], device="cuda")
            test_tensor = test_tensor + 1.0  # Simple operation
            result = test_tensor.item()  # Get the value back to validate operation
            
            # If we get here, the CUDA operation worked
            test_tensor = test_tensor.cpu()  # Move back to CPU to free CUDA memory
            torch.cuda.empty_cache()  # Clear CUDA cache
            logger.info(f" NVIDIA GPU (CUDA) detected and verified working (test result: {result})")
            return "cuda"
        except Exception as e:
            logger.warning(f"CUDA detected but test failed: {e}")
            torch.cuda.empty_cache()  # Clear CUDA cache
    
    # Then try MPS (Apple Silicon)
    if hasattr(torch.backends, 'mps') and torch.backends.mps.is_available():
        try:
            # Test MPS with a small operation
            test_tensor = torch.zeros(1).to('mps')
            test_tensor = test_tensor + 1
            test_tensor.cpu()  # Move back to CPU to free MPS memory
            logger.info(" Apple Silicon GPU (MPS) detected and verified working")
            return "mps"
        except Exception as e:
            logger.warning(f" MPS detected but test failed: {e}")
    
    # Fall back to CPU
    logger.info(" No GPU detected or all GPU tests failed, using CPU")
    return "cpu"

def run_on_device(func):
    """
    Decorator to run a function on the best available device.
    
    Args:
        func: The function to decorate
        
    Returns:
        A wrapped function that runs on the best available device
    """
    @functools.wraps(func)
    def wrapper(*args, **kwargs):
        # Log memory before operation
        log_memory_usage(f"before {func.__name__}")
        
        # Force garbage collection before operation
        gc.collect()
        
        # Get device if not already specified
        device = get_available_device()
        
        # Add device to kwargs if not already present
        if 'device' not in kwargs:
            kwargs['device'] = device
        
        try:
            start_time = time.time()
            result = func(*args, **kwargs)
            end_time = time.time()
            
            logger.debug(f"Function {func.__name__} ran on {device} in {end_time - start_time:.4f} seconds")
            return result
        except Exception as e:
            # Check if this is the SparseMPS error
            if "SparseMPS" in str(e) and device == "mps":
                logger.warning(f"MPS error detected: {e}")
                logger.warning("Falling back to CPU for this operation")
                
                # Update device to CPU and retry
                kwargs['device'] = 'cpu'
                
                # Force garbage collection before retry
                gc.collect()
                
                start_time = time.time()
                result = func(*args, **kwargs)
                end_time = time.time()
                
                logger.debug(f"Function {func.__name__} ran on CPU (fallback) in {end_time - start_time:.4f} seconds")
                return result
            else:
                # Re-raise other exceptions
                raise
        finally:
            # Force garbage collection after operation
            gc.collect()
            if device == 'cuda':
                torch.cuda.empty_cache()
            
            # Log memory after operation
            log_memory_usage(f"after {func.__name__}")
    
    return wrapper

# Initialize device once at module level
device = get_available_device()