PyTorch专用的性能采集

PyTorch推荐使用接口进行性能采集，采集完之后会自动解析性能数据，详细请参考《CANN 性能调优工具用户指南》中的“msprof采集通用命令”。

完整示例如下：

import torch
import torch_npu
device = torch.device("npu")
class DemoModel(torch.nn.Module):
    def forward(self, in0, in1):
        mul0 = in0 * in1
        sub0 = mul0 ** 2 - in0
        sub1 = mul0 ** 2 - in1
        slice0 = sub0[:, :2]
        slice1 = sub1[:, 2:]
        cat0 = torch.cat([slice0, slice1], dim=1)
        return cat0
def main():
    torch.manual_seed(2025)
    in0 = torch.randn(3,5).to(device)
    in1 = torch.randn(3,5).to(device)
    model = DemoModel().to(device)
    output = model(in0, in1)
    print(output.shape)
    print(output.device)
    print(output)
    # perf code
    experimental_config = torch_npu.profiler._ExperimentalConfig(
        export_type=[
            torch_npu.profiler.ExportType.Text,
            torch_npu.profiler.ExportType.Db
            ],
        profiler_level=torch_npu.profiler.ProfilerLevel.Level1,
        msprof_tx=False,
        aic_metrics=torch_npu.profiler.AiCMetrics.AiCoreNone,
        l2_cache=False,
        op_attr=False,
        data_simplification=False,
        record_op_args=False,
        gc_detect_threshold=None
    )
    steps = 10
    with torch_npu.profiler.profile(
        activities=[
            torch_npu.profiler.ProfilerActivity.CPU,
            torch_npu.profiler.ProfilerActivity.NPU
            ],
        schedule=torch_npu.profiler.schedule(wait=3, warmup=0, active=1, repeat=1, skip_first=1),
        on_trace_ready=torch_npu.profiler.tensorboard_trace_handler("./result"),
        record_shapes=False,
        profile_memory=False,
        with_stack=False,
        with_modules=False,
        with_flops=False,
        experimental_config=experimental_config) as prof:
        print("profiling start...")
        for step in range(steps):
            output = model(in0, in1)
            prof.step()
if __name__ == "__main__":
    main()