mirror of
https://github.com/Soul-AILab/SoulX-Podcast.git
synced 2026-05-07 05:59:16 +08:00
1ad755aa4e
Major updates: - Add vLLM engine support with automatic fallback to HuggingFace - Complete REST API implementation with sync/async modes - Add comprehensive API documentation - Organize scripts into dedicated directory API Features: - Support both HuggingFace and vLLM inference engines - Sync and async generation endpoints - Task queue management with concurrency control - Health check with engine information - Automatic file cleanup Configuration: - Environment variable based configuration - Engine validation and auto-fallback - Configurable concurrency limits Documentation: - README_API.md: Complete API usage guide - CHANGELOG_API.md: API version history - VLLM_UPGRADE_SUMMARY.md: Detailed upgrade guide - scripts/README.md: Scripts documentation Scripts Organization: - Move all test and utility scripts to scripts/ - Add configuration test script - Add singleton pattern test Performance: - vLLM engine provides 2-3x speedup - Better GPU memory utilization - Support for prefix caching 🤖 Generated with Claude Code Co-Authored-By: Claude <noreply@anthropic.com>
95 lines
2.6 KiB
Python
95 lines
2.6 KiB
Python
"""
|
||
GPU Memory Monitor for SoulX-Podcast API
|
||
用于监控GPU内存使用情况，帮助诊断内存泄漏问题
|
||
"""
|
||
import time
|
||
import torch
|
||
import psutil
|
||
import logging
|
||
from datetime import datetime
|
||
|
||
logging.basicConfig(
|
||
level=logging.INFO,
|
||
format='%(asctime)s - %(message)s'
|
||
)
|
||
logger = logging.getLogger(__name__)
|
||
|
||
|
||
def get_gpu_memory_info():
|
||
"""获取GPU内存信息"""
|
||
if not torch.cuda.is_available():
|
||
return None
|
||
|
||
info = {
|
||
'allocated_gb': torch.cuda.memory_allocated() / 1024**3,
|
||
'reserved_gb': torch.cuda.memory_reserved() / 1024**3,
|
||
'free_gb': (torch.cuda.memory_reserved() - torch.cuda.memory_allocated()) / 1024**3,
|
||
'max_allocated_gb': torch.cuda.max_memory_allocated() / 1024**3,
|
||
}
|
||
return info
|
||
|
||
|
||
def get_system_memory_info():
|
||
"""获取系统内存信息"""
|
||
memory = psutil.virtual_memory()
|
||
return {
|
||
'total_gb': memory.total / 1024**3,
|
||
'available_gb': memory.available / 1024**3,
|
||
'used_gb': memory.used / 1024**3,
|
||
'percent': memory.percent
|
||
}
|
||
|
||
|
||
def monitor_memory(interval=5):
|
||
"""持续监控内存使用"""
|
||
logger.info("Starting memory monitor...")
|
||
logger.info("Press Ctrl+C to stop")
|
||
|
||
try:
|
||
while True:
|
||
# GPU内存
|
||
gpu_info = get_gpu_memory_info()
|
||
if gpu_info:
|
||
logger.info(
|
||
f"GPU Memory - Allocated: {gpu_info['allocated_gb']:.2f}GB, "
|
||
f"Reserved: {gpu_info['reserved_gb']:.2f}GB, "
|
||
f"Free: {gpu_info['free_gb']:.2f}GB, "
|
||
f"Max: {gpu_info['max_allocated_gb']:.2f}GB"
|
||
)
|
||
|
||
# 系统内存
|
||
sys_info = get_system_memory_info()
|
||
logger.info(
|
||
f"System Memory - Used: {sys_info['used_gb']:.2f}GB / "
|
||
f"{sys_info['total_gb']:.2f}GB ({sys_info['percent']:.1f}%)"
|
||
)
|
||
|
||
time.sleep(interval)
|
||
|
||
except KeyboardInterrupt:
|
||
logger.info("Memory monitor stopped")
|
||
|
||
|
||
def clear_gpu_cache():
|
||
"""清理GPU缓存"""
|
||
if torch.cuda.is_available():
|
||
torch.cuda.empty_cache()
|
||
torch.cuda.reset_peak_memory_stats()
|
||
logger.info("GPU cache cleared")
|
||
|
||
# 显示清理后的状态
|
||
gpu_info = get_gpu_memory_info()
|
||
if gpu_info:
|
||
logger.info(
|
||
f"After clearing - Allocated: {gpu_info['allocated_gb']:.2f}GB, "
|
||
f"Reserved: {gpu_info['reserved_gb']:.2f}GB"
|
||
)
|
||
|
||
|
||
if __name__ == "__main__":
|
||
import sys
|
||
|
||
if len(sys.argv) > 1 and sys.argv[1] == "clear":
|
||
clear_gpu_cache()
|
||
else:
|
||
monitor_memory() |