How to make tensor cores work?

I’m trying to make my network utilize my A100 tensor cores. To debug I’ve created simple model with single convolutional layer.

import torch.nn as nn
import torch
import nvidia_dlprof_pytorch_nvtx
from triton.testing import do_bench
import contextlib

class TestNet(nn.Module):
    def __init__(self):
        self.conv1 = nn.Conv2d(64, 512, kernel_size=(3,3), padding=1)

    def forward(self, x):
        return self.conv1(x)

def run_step(opt, model):
    # with torch.cuda.amp.autocast():
    x = torch.randn(64,64,128,128, dtype=torch.float16, requires_grad=True).to(device)
    model = model.half()
    out = model(x)
    loss = out.sum()

if __name__ == '__main__':
    device = 'cuda:0'
    do_dlprof = True
    do_benchmark = False
    do_profiler = False
    profiler_path = ...
    torch.backends.cudnn.benchmark = True
    torch.backends.cuda.matmul.allow_tf32 = True
    torch.backends.cudnn.allow_tf32 = True
    if do_dlprof:

    model = TestNet()
    opt = torch.optim.Adam(model.parameters())
    model =

    dlprof_ctx = torch.autograd.profiler.emit_nvtx(enabled=do_dlprof)
    with dlprof_ctx:
        profiler_ctx = torch.profiler.profile(activities=[torch.profiler.ProfilerActivity.CPU,
                    with_stack=True) if do_profiler else contextlib.nullcontext()
        with profiler_ctx as p:
            for i in range(15):
                if do_benchmark:
                    do_bench(lambda: run_step(opt, model), warmup=10, rep=10)
                    run_step(opt, model)
                if do_profiler:

I’ve already tried:

  • making dimensions of input tensor and in\out channels divisible by 8
  • mixed precision\half precision
  • switch to tf32
    But no matter what I do, dlprof profiler tells that tensor cores are not used for convolution operation.

    At the same time pytorch profiler tells that tensor cores were used.

    If it helps, the name of used convolution kernel is ‘cutlass_tensorop_f16_s16816fprop_optimized_f16_128x128_32x3_nhwc’
    Why the results of both profilers are inconsistent do the tensor cores work indeed?

The problem was connected with environment setup.
I’ve set up all libraries in my working conda environment but dlprof didn’t log any info about kernel usage. After I’had switched to nvidia docker image as it was recommended in dlprof installation guide, kernel info started logging, dlprof started correctly displaying tc usage as torch profiler and the problem has gone.

1 Like

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.