Torch.compile [WARNING] not enough SMs to use max_autotune_gemm mode

torch.compile
1

I have a small model with 900K in wights

compiled_model = torch.compile(net_model,backend="inductor",mode="max-autotune")

When I compile the model using max_autotune model, he is giving me this warning :
[WARNING] not enough SMs to use max_autotune_gemm mode and i don't know how to exploit thi feature.

PS: I’m working on complex- valued models model with a ComplexConvolution class that uses two nn.Conv2D (for real and imaginary part )

2

The warning seems to be raised here, which assumes your GPU needs at least 80 SMs for this mode, which seems to be a hard-coded limitation in torch.compile.

4 Likes
3

OK, i get it. I have only 48 streaming, thank you for your response.

4

Did you somehow resolve this and were able to use the max-autotune mode?

5

@Prakhar I think after providing enough GPU resource it should automatically fix