Fix: causal_conv1d with channel last layout requires strides (x.stride(0) and x.stride(2)) to be multiples of 8

[Dexterp37]

This addresses the following problem:

RuntimeError: causal_conv1d with channel last layout requires strides (x.stride(0) and x.stride(2)) to be multiples of 8

I'm not sure if this is the correct fix, but I'm quite sure I can't run without these changes.

[catalpaaa]

you should add contiguous() at line 880 882 and 883 too for training to run

dxBC_given = rearrange(dxBC_given, "b s d -> b d s").contiguous()
dxBC_given, dweight, dbias, *_ = causal_conv1d_cuda.causal_conv1d_bwd(
    rearrange(xBC, "b s d -> b d s").contiguous(), conv1d_weight, conv1d_bias,
    rearrange(dxBC, "b s d -> b d s").contiguous(), seq_idx, None, None, dxBC_given, False, ctx.activation in ["silu", "swish"]
)

(sorry if i got the line numbers wrong)

[Dexterp37]

you should add contiguous() at line 880 882 and 883 too for training to run

Whoops, you're indeed right! Thanks @catalpaaa , I updated the PR .

[Dexterp37]

@tridao , is this the expected way to address the mentioned problem? I believe (but may be wrong!) that calling contiguous might have negative performance implications.

[catalpaaa]

@tridao , is this the expected way to address the mentioned problem? I believe (but may be wrong!) that calling contiguous might have negative performance implications.

#351 (comment) try this

[Dexterp37]

@tridao , is this the expected way to address the mentioned problem? I believe (but may be wrong!) that calling contiguous might have negative performance implications.

#351 (comment) try this

You're indeed right!