Why not have test_conv tests make use of fast/cuda backends for accessing matmul?

[paulbricman]

Hi,

I wanted to try using the matmul op in the conv1d implementation (i.e. task 4.1), but the tests are being run on the simple backend, the only one of the three which lacks a matmul implementation. I find this a bit weird especially because (1) conv1d is supposed to be implemented in fast_conv.py which made me expect the tests to use the fast backend, and (2) the guide on convolution actually makes use of the matmul op itself.

Should students essentially ignore both previous matmul implementations when working on conv1d and write it more from scratch with multiple nested fors?

Edit: It wasn't that hard to do it from scratch after all, just a bit confusing given the above