| 41 | |
| 42 | @amp.autocast(enabled=False) |
| 43 | def rope_apply(x, grid_sizes, freqs): |
| 44 | n, c = x.size(2), x.size(3) // 2 |
| 45 | |
| 46 | # split freqs |
| 47 | freqs = freqs.split([c - 2 * (c // 3), c // 3, c // 3], dim=1) |
| 48 | |
| 49 | # loop over samples |
| 50 | output = [] |
| 51 | for i, (f, h, w) in enumerate(grid_sizes.tolist()): |
| 52 | seq_len = f * h * w |
| 53 | |
| 54 | # precompute multipliers |
| 55 | x_i = torch.view_as_complex(x[i, :seq_len].to(torch.float64).reshape( |
| 56 | seq_len, n, -1, 2)) |
| 57 | freqs_i = torch.cat([ |
| 58 | freqs[0][:f].view(f, 1, 1, -1).expand(f, h, w, -1), |
| 59 | freqs[1][:h].view(1, h, 1, -1).expand(f, h, w, -1), |
| 60 | freqs[2][:w].view(1, 1, w, -1).expand(f, h, w, -1) |
| 61 | ], |
| 62 | dim=-1).reshape(seq_len, 1, -1) |
| 63 | |
| 64 | # apply rotary embedding |
| 65 | x_i = torch.view_as_real(x_i * freqs_i).flatten(2) |
| 66 | x_i = torch.cat([x_i, x[i, seq_len:]]) |
| 67 | |
| 68 | # append to collection |
| 69 | output.append(x_i) |
| 70 | return torch.stack(output).float() |
| 71 | |
| 72 | |
| 73 | class WanRMSNorm(nn.Module): |