jbilcke-hf's picture
jbilcke-hf HF Staff
add gpu tracking
0d34ea8
"""
GPU monitoring service for Video Model Studio.
Tracks NVIDIA GPU resources like utilization, memory, and temperature.
"""
import os
import time
import logging
from typing import Dict, List, Any, Optional, Tuple
from collections import deque
from datetime import datetime
# Force the use of the Agg backend which is thread-safe
import matplotlib
matplotlib.use('Agg') # Must be before importing pyplot
import matplotlib.pyplot as plt
import numpy as np
logger = logging.getLogger(__name__)
logger.setLevel(logging.INFO)
# Optional import of pynvml
try:
import pynvml
PYNVML_AVAILABLE = True
except ImportError:
PYNVML_AVAILABLE = False
logger.info("pynvml not available, GPU monitoring will be limited")
class GPUMonitoringService:
"""Service for monitoring NVIDIA GPU resources"""
def __init__(self, history_minutes: int = 10, sample_interval: int = 5):
"""Initialize the GPU monitoring service
Args:
history_minutes: How many minutes of history to keep
sample_interval: How many seconds between samples
"""
self.history_minutes = history_minutes
self.sample_interval = sample_interval
self.max_samples = (history_minutes * 60) // sample_interval
# Track if the monitoring thread is running
self.is_running = False
self.thread = None
# Check if NVIDIA GPUs are available
self.has_nvidia_gpus = False
self.gpu_count = 0
self.device_info = []
self.history = {}
# Try to initialize NVML
self._initialize_nvml()
# Initialize history data structures if GPUs are available
if self.has_nvidia_gpus:
self._initialize_history()
def _initialize_nvml(self):
"""Initialize NVIDIA Management Library"""
if not PYNVML_AVAILABLE:
logger.info("pynvml module not installed, GPU monitoring disabled")
return
try:
pynvml.nvmlInit()
self.gpu_count = pynvml.nvmlDeviceGetCount()
self.has_nvidia_gpus = self.gpu_count > 0
if self.has_nvidia_gpus:
logger.info(f"Successfully initialized NVML, found {self.gpu_count} GPU(s)")
# Get static information about each GPU
for i in range(self.gpu_count):
self.device_info.append(self._get_device_info(i))
else:
logger.info("No NVIDIA GPUs found")
except Exception as e:
logger.warning(f"Failed to initialize NVML: {str(e)}")
self.has_nvidia_gpus = False
def _initialize_history(self):
"""Initialize data structures for storing metric history"""
for i in range(self.gpu_count):
self.history[i] = {
'timestamps': deque(maxlen=self.max_samples),
'utilization': deque(maxlen=self.max_samples),
'memory_used': deque(maxlen=self.max_samples),
'memory_total': deque(maxlen=self.max_samples),
'memory_percent': deque(maxlen=self.max_samples),
'temperature': deque(maxlen=self.max_samples),
'power_usage': deque(maxlen=self.max_samples),
'power_limit': deque(maxlen=self.max_samples),
}
def _get_device_info(self, device_index: int) -> Dict[str, Any]:
"""Get static information about a GPU device
Args:
device_index: Index of the GPU device
Returns:
Dictionary with device information
"""
if not PYNVML_AVAILABLE or not self.has_nvidia_gpus:
return {"error": "NVIDIA GPUs not available"}
try:
handle = pynvml.nvmlDeviceGetHandleByIndex(device_index)
# Get device name (decode if it's bytes)
name = pynvml.nvmlDeviceGetName(handle)
if isinstance(name, bytes):
name = name.decode('utf-8')
# Get device UUID
uuid = pynvml.nvmlDeviceGetUUID(handle)
if isinstance(uuid, bytes):
uuid = uuid.decode('utf-8')
# Get memory info, compute capability
memory_info = pynvml.nvmlDeviceGetMemoryInfo(handle)
compute_capability = pynvml.nvmlDeviceGetCudaComputeCapability(handle)
# Get power limits if available
try:
power_limit = pynvml.nvmlDeviceGetPowerManagementLimit(handle) / 1000.0 # Convert to watts
except pynvml.NVMLError:
power_limit = None
return {
'index': device_index,
'name': name,
'uuid': uuid,
'memory_total': memory_info.total,
'memory_total_gb': memory_info.total / (1024**3), # Convert to GB
'compute_capability': f"{compute_capability[0]}.{compute_capability[1]}",
'power_limit': power_limit
}
except Exception as e:
logger.error(f"Error getting device info for GPU {device_index}: {str(e)}")
return {"error": str(e), "index": device_index}
def collect_gpu_metrics(self) -> List[Dict[str, Any]]:
"""Collect current GPU metrics for all available GPUs
Returns:
List of dictionaries with current metrics for each GPU
"""
if not PYNVML_AVAILABLE or not self.has_nvidia_gpus:
return []
metrics = []
timestamp = datetime.now()
for i in range(self.gpu_count):
try:
handle = pynvml.nvmlDeviceGetHandleByIndex(i)
# Get utilization rates
utilization = pynvml.nvmlDeviceGetUtilizationRates(handle)
# Get memory information
memory_info = pynvml.nvmlDeviceGetMemoryInfo(handle)
# Get temperature
temperature = pynvml.nvmlDeviceGetTemperature(handle, pynvml.NVML_TEMPERATURE_GPU)
# Get power usage if available
try:
power_usage = pynvml.nvmlDeviceGetPowerUsage(handle) / 1000.0 # Convert to watts
except pynvml.NVMLError:
power_usage = None
# Get process information
processes = []
try:
for proc in pynvml.nvmlDeviceGetComputeRunningProcesses(handle):
try:
process_name = pynvml.nvmlSystemGetProcessName(proc.pid)
if isinstance(process_name, bytes):
process_name = process_name.decode('utf-8')
except pynvml.NVMLError:
process_name = f"Unknown (PID: {proc.pid})"
processes.append({
'pid': proc.pid,
'name': process_name,
'memory_used': proc.usedGpuMemory,
'memory_used_mb': proc.usedGpuMemory / (1024**2) # Convert to MB
})
except pynvml.NVMLError:
# Unable to get process information, continue with empty list
pass
gpu_metrics = {
'index': i,
'timestamp': timestamp,
'utilization_gpu': utilization.gpu,
'utilization_memory': utilization.memory,
'memory_total': memory_info.total,
'memory_used': memory_info.used,
'memory_free': memory_info.free,
'memory_percent': (memory_info.used / memory_info.total) * 100,
'temperature': temperature,
'power_usage': power_usage,
'processes': processes
}
metrics.append(gpu_metrics)
except Exception as e:
logger.error(f"Error collecting metrics for GPU {i}: {str(e)}")
metrics.append({
'index': i,
'error': str(e)
})
return metrics
def update_history(self):
"""Update GPU metrics history"""
if not self.has_nvidia_gpus:
return
current_metrics = self.collect_gpu_metrics()
timestamp = datetime.now()
for gpu_metrics in current_metrics:
if 'error' in gpu_metrics:
continue
idx = gpu_metrics['index']
self.history[idx]['timestamps'].append(timestamp)
self.history[idx]['utilization'].append(gpu_metrics['utilization_gpu'])
self.history[idx]['memory_used'].append(gpu_metrics['memory_used'])
self.history[idx]['memory_total'].append(gpu_metrics['memory_total'])
self.history[idx]['memory_percent'].append(gpu_metrics['memory_percent'])
self.history[idx]['temperature'].append(gpu_metrics['temperature'])
if gpu_metrics['power_usage'] is not None:
self.history[idx]['power_usage'].append(gpu_metrics['power_usage'])
else:
self.history[idx]['power_usage'].append(0)
# Store power limit in history (static but kept for consistency)
info = self.device_info[idx]
if 'power_limit' in info and info['power_limit'] is not None:
self.history[idx]['power_limit'].append(info['power_limit'])
else:
self.history[idx]['power_limit'].append(0)
def start_monitoring(self):
"""Start background thread for collecting GPU metrics"""
if self.is_running:
logger.warning("GPU monitoring thread already running")
return
if not self.has_nvidia_gpus:
logger.info("No NVIDIA GPUs found, not starting monitoring thread")
return
import threading
self.is_running = True
def _monitor_loop():
while self.is_running:
try:
self.update_history()
time.sleep(self.sample_interval)
except Exception as e:
logger.error(f"Error in GPU monitoring thread: {str(e)}", exc_info=True)
time.sleep(self.sample_interval)
self.thread = threading.Thread(target=_monitor_loop, daemon=True)
self.thread.start()
logger.info("GPU monitoring thread started")
def stop_monitoring(self):
"""Stop the GPU monitoring thread"""
if not self.is_running:
return
self.is_running = False
if self.thread:
self.thread.join(timeout=1.0)
logger.info("GPU monitoring thread stopped")
def get_gpu_info(self) -> List[Dict[str, Any]]:
"""Get information about all available GPUs
Returns:
List of dictionaries with GPU information
"""
return self.device_info
def get_current_metrics(self) -> List[Dict[str, Any]]:
"""Get current metrics for all GPUs
Returns:
List of dictionaries with current GPU metrics
"""
return self.collect_gpu_metrics()
def generate_utilization_plot(self, gpu_index: int) -> plt.Figure:
"""Generate a plot of GPU utilization over time
Args:
gpu_index: Index of the GPU to plot
Returns:
Matplotlib figure with utilization plot
"""
plt.close('all') # Close all existing figures
fig, ax = plt.subplots(figsize=(10, 5))
if not self.has_nvidia_gpus or gpu_index not in self.history:
ax.set_title(f"No data available for GPU {gpu_index}")
return fig
history = self.history[gpu_index]
if not history['timestamps']:
ax.set_title(f"No history data for GPU {gpu_index}")
return fig
# Convert timestamps to strings
x = [t.strftime('%H:%M:%S') for t in history['timestamps']]
# If we have many points, show fewer labels for readability
if len(x) > 10:
step = len(x) // 10
ax.set_xticks(range(0, len(x), step))
ax.set_xticklabels([x[i] for i in range(0, len(x), step)], rotation=45)
# Plot utilization
ax.plot(x, list(history['utilization']), 'b-', label='GPU Utilization %')
ax.set_ylim(0, 100)
# Add temperature on secondary y-axis
ax2 = ax.twinx()
ax2.plot(x, list(history['temperature']), 'r-', label='Temperature °C')
ax2.set_ylabel('Temperature (°C)', color='r')
ax2.tick_params(axis='y', colors='r')
# Set labels and title
ax.set_title(f'GPU {gpu_index} Utilization Over Time')
ax.set_xlabel('Time')
ax.set_ylabel('Utilization %')
ax.grid(True, alpha=0.3)
# Add legend
lines, labels = ax.get_legend_handles_labels()
lines2, labels2 = ax2.get_legend_handles_labels()
ax.legend(lines + lines2, labels + labels2, loc='upper left')
plt.tight_layout()
return fig
def generate_memory_plot(self, gpu_index: int) -> plt.Figure:
"""Generate a plot of GPU memory usage over time
Args:
gpu_index: Index of the GPU to plot
Returns:
Matplotlib figure with memory usage plot
"""
plt.close('all') # Close all existing figures
fig, ax = plt.subplots(figsize=(10, 5))
if not self.has_nvidia_gpus or gpu_index not in self.history:
ax.set_title(f"No data available for GPU {gpu_index}")
return fig
history = self.history[gpu_index]
if not history['timestamps']:
ax.set_title(f"No history data for GPU {gpu_index}")
return fig
# Convert timestamps to strings
x = [t.strftime('%H:%M:%S') for t in history['timestamps']]
# If we have many points, show fewer labels for readability
if len(x) > 10:
step = len(x) // 10
ax.set_xticks(range(0, len(x), step))
ax.set_xticklabels([x[i] for i in range(0, len(x), step)], rotation=45)
# Plot memory percentage
ax.plot(x, list(history['memory_percent']), 'g-', label='Memory Usage %')
ax.set_ylim(0, 100)
# Add absolute memory values on secondary y-axis (convert to GB)
ax2 = ax.twinx()
memory_used_gb = [m / (1024**3) for m in history['memory_used']]
memory_total_gb = [m / (1024**3) for m in history['memory_total']]
ax2.plot(x, memory_used_gb, 'm--', label='Used (GB)')
ax2.set_ylabel('Memory (GB)')
# Set labels and title
ax.set_title(f'GPU {gpu_index} Memory Usage Over Time')
ax.set_xlabel('Time')
ax.set_ylabel('Usage %')
ax.grid(True, alpha=0.3)
# Add legend
lines, labels = ax.get_legend_handles_labels()
lines2, labels2 = ax2.get_legend_handles_labels()
ax.legend(lines + lines2, labels + labels2, loc='upper left')
plt.tight_layout()
return fig
def generate_power_plot(self, gpu_index: int) -> plt.Figure:
"""Generate a plot of GPU power usage over time
Args:
gpu_index: Index of the GPU to plot
Returns:
Matplotlib figure with power usage plot
"""
plt.close('all') # Close all existing figures
fig, ax = plt.subplots(figsize=(10, 5))
if not self.has_nvidia_gpus or gpu_index not in self.history:
ax.set_title(f"No data available for GPU {gpu_index}")
return fig
history = self.history[gpu_index]
if not history['timestamps'] or not any(history['power_usage']):
ax.set_title(f"No power data for GPU {gpu_index}")
return fig
# Convert timestamps to strings
x = [t.strftime('%H:%M:%S') for t in history['timestamps']]
# If we have many points, show fewer labels for readability
if len(x) > 10:
step = len(x) // 10
ax.set_xticks(range(0, len(x), step))
ax.set_xticklabels([x[i] for i in range(0, len(x), step)], rotation=45)
# Plot power usage
power_usage = list(history['power_usage'])
if any(power_usage): # Only plot if we have actual power data
ax.plot(x, power_usage, 'b-', label='Power Usage (W)')
# Get power limit if available
power_limit = list(history['power_limit'])
if any(power_limit): # Only plot if we have power limit data
# Show power limit as horizontal line
limit = max(power_limit) # Should be constant, but take max just in case
if limit > 0:
ax.axhline(y=limit, color='r', linestyle='--', label=f'Power Limit ({limit}W)')
# Set labels and title
ax.set_title(f'GPU {gpu_index} Power Usage Over Time')
ax.set_xlabel('Time')
ax.set_ylabel('Power (Watts)')
ax.grid(True, alpha=0.3)
ax.legend(loc='upper left')
else:
ax.set_title(f"Power data not available for GPU {gpu_index}")
plt.tight_layout()
return fig
def shutdown(self):
"""Clean up resources when shutting down"""
self.stop_monitoring()
# Shutdown NVML if it was initialized
if PYNVML_AVAILABLE and self.has_nvidia_gpus:
try:
pynvml.nvmlShutdown()
logger.info("NVML shutdown complete")
except Exception as e:
logger.error(f"Error during NVML shutdown: {str(e)}")