mps-watch 0.1.2

Apple Silicon MPS (Metal Performance Shaders) memory monitoring tool for PyTorch.

mps-watch

The Apple Silicon equivalent of nvidia-smi or gpustat for PyTorch. Monitor your MPS (Metal Performance Shaders) memory usage in real-time.

Features

• Python API: Context manager and decorator for tracking VRAM usage of specific code blocks.
• CLI Tool: Terminal-based dashboard for system unified memory monitoring.
• Safe Fallbacks: Graceful handling on non-Mac systems or without MPS.

Installation

pip install mps-watch

Usage

Python API

Monitor memory usage of specific blocks or functions:

from mps_watch import monitor
import torch

# As context manager
with monitor("Training Block"):
    # Your MPS code here
    model.to("mps")
    output = model(input)

# As decorator
@monitor("Inference")
def predict(x):
    return model(x)

# Get current usage directly
from mps_watch import get_current_memory_usage, get_system_memory

allocated, reserved = get_current_memory_usage()
print(f"Current Allocated: {allocated} bytes")

# Get system-wide unified memory
sys_mem = get_system_memory()
print(f"System Total: {sys_mem['total']} bytes")
print(f"System Used: {sys_mem['used']} bytes")

Output: It will print a table showing Initial, Final, and Delta memory usage (Allocated vs Reserved).

CLI Tool

Run the dashboard in your terminal:

mps-watch

This will show a live-updating dashboard of your System Unified Memory and MPS availability status.

Requirements

• macOS 12.3+ (for MPS support)
• Python 3.8+
• PyTorch
• rich
• psutil

Understanding the Stats

CLI vs Python API

• The CLI Tool (mps-watch): Shows System-wide memory usage. Since Apple Silicon uses Unified Memory, this reflects the total RAM used by the OS and all applications combined. It gives you an idea of the overall system pressure.
• The Python API (monitor / get_current_memory_usage): Shows Process-specific MPS memory usage. This is the amount of memory specifically allocated by PyTorch on the Metal device for your script.

Why are they different? The CLI shows the whole pie (System RAM). The Python API shows just your slice (PyTorch MPS Usage).

License

MIT