profiling采集
1.进程级别采集
比较原始的profiling采集方法，使用cann相关的工具包，对于训练和推理都可以进行相关的性能采集任务
注意
该采集方式是通过服务启动用户工作目录当前的sock文件进行通信的，比如是root用户启动推理任务的话，/root/目录会有一个sock文件，我们启动msprof采集任务的话也需要以root用户启动，如果推理任务的是HwHiAiUser启动的话，msprof也需要使用HwHiAiUser启动，防止msprof找不到对应的sock文件
2.常规采集
见高亮的20行和49行，如果只需要traceview.json的话只需要49行可以去除20行，加上20行的话torch npu会采集更详细的信息到指定文件夹，包含各个算子的耗时统计一系列csv文件
import torch
import torch_npu
import time
def test_linear(num_runs=200):
forward_times = []
backward_times = []
input = torch.randn(
[32768, 1, 1024], dtype=torch.bfloat16, device="npu", requires_grad=True
)
weight = torch.randn(
[8192, 1024], dtype=torch.bfloat16, device="npu", requires_grad=True
)
llm_profile = torch_npu.profiler.profile(
activities=[torch_npu.profiler.ProfilerActivity.CPU, torch_npu.profiler.ProfilerActivity.NPU],
with_stack=True,
with_modules=True,
on_trace_ready=torch_npu.profiler.tensorboard_trace_handler("./910b_linear")
)
llm_profile.start()
for _ in range(num_runs):
# Measure forward time
torch.npu.synchronize()
fwd_start_time = time.time()
# 模型层计算linear，使用torch npu总计有4种方式可实现该功能
# 无论下面4种方式最终调用的是aclnnmm还是aclnnmatmul，最终在设备计算流上都是调用aclnnMm_MatMulCommon_MatMulV2，耗时也都相同
# output_parallel = torch_npu.npu_linear(torch.squeeze(input), weight).unsqueeze(1)
# output_parallel = torch.mm(torch.squeeze(input), weight.t()).unsqueeze(1)
# output_parallel = torch.matmul(input, weight.t())
output_parallel = torch.nn.functional.linear(input, weight)
torch.npu.synchronize()
fwd_end_time = time.time()
forward_times.append(fwd_end_time - fwd_start_time)
# Measure backward time
if torch.is_tensor(output_parallel):
label = torch.ones_like(output_parallel)
torch.npu.synchronize()
bwd_start_time = time.time()
output_parallel.backward(label)
torch.npu.synchronize()
bwd_end_time = time.time()
backward_times.append(bwd_end_time - bwd_start_time)
else:
raise RuntimeError("the result of forward is not a tensor")
llm_profile.stop()
llm_profile.export_chrome_trace(f"./910b_linear.json")
avg_forward_time = (sum(forward_times) / num_runs) * 1000 # Convert to milliseconds
avg_backward_time = (
sum(backward_times) / num_runs
) * 1000 # Convert to milliseconds
print("第一组测试")
print(f"Average forward time over {num_runs} runs: {avg_forward_time:.3f} ms")
print(f"Average backward time over {num_runs} runs: {avg_backward_time:.3f} ms")
print(
f"Average fwd+bwd time over {num_runs} runs: {avg_forward_time + avg_backward_time:.3f} ms"
)
test_linear()