How to disable triton when calling torch.compile in max-autotune mode

1

I train my model in DDP and introduce torch.compile:
model = torch.compile(model, mode=“max-autotune”)
it worked until triton was installed. So I wonder how to disable triton before torch.compile in max-autotune mode

2

You could try to disable Triton via selecting the desired backends in torch._inductor.config.max_autotune_gemm_backends.

1 Like