(qkv, scale, eps)
| 120 | |
| 121 | @flags.compile_wrap |
| 122 | def scale_for_cosine_sim_qkv(qkv, scale, eps): |
| 123 | q, k, v = qkv.unbind(2) |
| 124 | q, k = scale_for_cosine_sim(q, k, scale[:, None], eps) |
| 125 | return torch.stack((q, k, v), dim=2) |
| 126 | |
| 127 | |
| 128 | class ConvGEGLU(nn.Module): |
no test coverage detected