advanced.md

Advanced Usage

Customizing Reports

Report Styling

You can customize the appearance of the HTML report by modifying the CSS in the template. The report uses Bootstrap 5 for styling.

Custom Metrics

Add custom metrics to your profiling data:

from profiling_py import default_profiler

with default_profiler.profile("my_operation") as step:
    # Your code here
    step.metadata["custom_metric"] = 42

GPU Profiling

Multi-GPU Support

Profile across multiple GPUs by creating multiple profilers:

from profiling_py import Profiler

# Profile on GPU 0
with Profiler(enable_gpu=True, gpu_device=0) as p0:
    # Code for GPU 0
    pass

# Profile on GPU 1
with Profiler(enable_gpu=True, gpu_device=1) as p1:
    # Code for GPU 1
    pass

GPU Metrics

Available GPU metrics include:

• gpu_utilization: GPU core utilization (0-100%)
• memory_used_mb: Used GPU memory in MB
• memory_total_mb: Total GPU memory in MB
• memory_utilization: Memory utilization percentage
• temperature_gpu: GPU temperature in Celsius
• power_usage_w: Current power usage in watts
• power_limit_w: Power limit in watts

Performance Tips

Reducing Overhead

For minimal overhead:

1. Only enable the metrics you need
2. Use larger step sizes
3. Profile only critical sections

Memory Profiling

Memory profiling adds overhead. For accurate results:

• Profile for longer durations
• Use larger step sizes
• Avoid frequent small allocations

Integration

Jupyter Notebooks

Use the profiler in Jupyter notebooks:

from profiling_py import profile

@profile
def train_model():
    # Training code
    pass

# Run and display report
train_model()

Web Applications

Profile web applications by creating a profiler middleware:

from fastapi import Request
from profiling_py import Profiler

def profiling_middleware(app):
    async def middleware(request: Request, call_next):
        with Profiler() as prof:
            response = await call_next(request)
            prof.generate_report(output_dir="./profiling_reports")
        return response
    return middleware