Error in stride shape

[Mxdy21]

Hi,

I noticed an error/warning because of the strides misshape in DDP when you train the architecture ddpmpp, a simple fix for this is to fix the operation in the attention modules contiguously. Here's the error :

params[166] in this process with sizes [256, 256, 1, 1] appears not to match strides of the same param in process 0.

To fix it you need to modify 5 lines of code.
For the AttentionOP modules you modify the backward and forward from :

class AttentionOp(torch.autograd.Function):
    @staticmethod
    def forward(ctx, q, k):
        w = torch.einsum('ncq,nck->nqk', q.to(torch.float32), (k / np.sqrt(k.shape[1])).to(torch.float32)).softmax(dim=2).to(q.dtype)
        ctx.save_for_backward(q, k, w)
        return w

    @staticmethod
    def backward(ctx, dw):
        q, k, w = ctx.saved_tensors
        db = torch._softmax_backward_data(grad_output=dw.to(torch.float32), output=w.to(torch.float32), dim=2, input_dtype=torch.float32)
        dq = torch.einsum('nck,nqk->ncq', k.to(torch.float32), db).to(q.dtype) / np.sqrt(k.shape[1])
        dk = torch.einsum('ncq,nqk->nck', q.to(torch.float32), db).to(k.dtype) / np.sqrt(k.shape[1])
        return dq, dk

to :

class AttentionOp(torch.autograd.Function):
    @staticmethod
    def forward(ctx, q, k):
        w = torch.einsum('ncq,nck->nqk', q.to(torch.float32), (k / np.sqrt(k.shape[1])).to(torch.float32)).softmax(dim=2).to(q.dtype).contiguous()
        ctx.save_for_backward(q, k, w)
        return w

    @staticmethod
    def backward(ctx, dw):
        q, k, w = ctx.saved_tensors
        db = torch._softmax_backward_data(grad_output=dw.to(torch.float32), output=w.to(torch.float32), dim=2, input_dtype=torch.float32).contiguous()
        dq = torch.einsum('nck,nqk->ncq', k.to(torch.float32), db).to(q.dtype).contiguous() / np.sqrt(k.shape[1])
        dk = torch.einsum('ncq,nqk->nck', q.to(torch.float32), db).to(k.dtype).contiguous() / np.sqrt(k.shape[1])
        return dq, dk

and in the UNetBlock module, from :

def forward(self, x, emb):
        orig = x
        x = self.conv0(silu(self.norm0(x)))

        params = self.affine(emb).unsqueeze(2).unsqueeze(3).to(x.dtype)
        if self.adaptive_scale:
            scale, shift = params.chunk(chunks=2, dim=1)
            x = silu(torch.addcmul(shift, self.norm1(x), scale + 1))
        else:
            x = silu(self.norm1(x.add_(params)))

        x = self.conv1(torch.nn.functional.dropout(x, p=self.dropout, training=self.training))
        x = x.add_(self.skip(orig) if self.skip is not None else orig)
        x = x * self.skip_scale

        if self.num_heads:
            q, k, v = self.qkv(self.norm2(x)).reshape(x.shape[0] * self.num_heads, x.shape[1] // self.num_heads, 3, -1).unbind(2)
            w = AttentionOp.apply(q, k)
            a = torch.einsum('nqk,nck->ncq', w, v)
            x = self.proj(a.reshape(*x.shape)).add_(x)
            x = x * self.skip_scale
        return x

to :

    def forward(self, x, emb):
        orig = x
        x = self.conv0(silu(self.norm0(x)))

        params = self.affine(emb).unsqueeze(2).unsqueeze(3).to(x.dtype)
        if self.adaptive_scale:
            scale, shift = params.chunk(chunks=2, dim=1)
            x = silu(torch.addcmul(shift, self.norm1(x), scale + 1))
        else:
            x = silu(self.norm1(x.add_(params)))

        x = self.conv1(torch.nn.functional.dropout(x, p=self.dropout, training=self.training))
        x = x.add_(self.skip(orig) if self.skip is not None else orig)
        x = x * self.skip_scale

        if self.num_heads:
            q, k, v = self.qkv(self.norm2(x)).reshape(x.shape[0] * self.num_heads, x.shape[1] // self.num_heads, 3, -1).unbind(2)
            w = AttentionOp.apply(q, k)
            a = torch.einsum('nqk,nck->ncq', w, v).contiguous()
            x = self.proj(a.reshape(*x.shape)).add_(x)
            x = x * self.skip_scale
        return x

I know that the author find this warning/error and decided to mute it, to solve it you just need those few lines of code

[Newbeeer]

Thanks for the fix! Will this change slow the forward/backward pass?