Trying to use triton on torch inductor

compile torch._inductor
1

Trying to use the triton kernel for this convolution, however I get an error. Am I doing something wrong?

import torch

import torch._dynamo
from torch._inductor import config

@config.patch({"triton.convolution" : "triton"})
def run_conv():
    class M(torch.nn.Module):
        def __init__(
            self,
            **kwargs,
        ):
            super().__init__()
            self.conv1 = torch.nn.Conv2d(
                             4,
                             320,
                             (3,3),
                             (1,1),
                             (0,0),
                             (1,1),
                         )

        def forward(self, x):
            x1 = self.conv1(x)
            return x1

    mod = M().cuda()

    opt_mod = torch._dynamo.optimize("inductor")(mod)
    inputs = (
        torch.randn([2, 4, 66, 66], dtype=torch.float32, device="cuda"),
    )
    opt_y = opt_mod(*inputs)
    print(opt_y)

run_conv()

…

File "/home/alexander/triton/python/triton/runtime/autotuner.py", line 95, in prune_configs
    pruned_configs = self.early_config_prune(self.configs, self.nargs)
  File "/home/alexander/pytorch/torch/_inductor/triton_ops/conv_perf_model.py", line 100, in early_config_prune
    cc = _triton.runtime.cc(backend, device)
AttributeError: module 'triton._C.libtriton.triton.runtime' has no attribute 'cc'

in “torch/_inductor/triton_ops/conv.py”,
I tried to remove the 'return’s in the @triton.jit functions
and to not use the autotuner, and I had to remove the ‘tl.store’ in _kernel_delta_x() or it would get stuck

tl.store(y_ptrs, acc, mask=mask_y)

 return

but I’m not sure what is wrong with the tl.store.

2

Hi, there is a related issue here. From the link it seems that @jansel is working on this.
This is a version mismatch between triton and pytorch and believe it will be resolved in later version.