profiler.md

October 31, 2023 · View on GitHub

1 PyTorch 2.1 里的 Profiler: torch.profiler

可以分析 CPU、GPU 执行时间,也可以分析显存占用,有 trace 功能,有 stack trace

from torch.profiler import profile, record_function, ProfilerActivity

with profile(activities=[ProfilerActivity.CPU, ProfilerActivity.CUDA], record_shapes=True) as prof: # 加这两行就行
    with record_funciton("model_inference"):
        model(inputs)

结果都是 pytorch 里算子级别的,比如 aten::conv2d 这种

2 autograd.profiler

with torch.autograd.profilerr.profiler() as prof