| 368 | |
| 369 | class SelfAttentionBlock(nn.Module): |
| 370 | def __init__(self, d_model, d_head, cond_features, dropout=0.0): |
| 371 | super().__init__() |
| 372 | self.d_head = d_head |
| 373 | self.n_heads = d_model // d_head |
| 374 | self.norm = AdaRMSNorm(d_model, cond_features) |
| 375 | self.qkv_proj = apply_wd(Linear(d_model, d_model * 3, bias=False)) |
| 376 | self.scale = nn.Parameter(torch.full([self.n_heads], 10.0)) |
| 377 | self.pos_emb = AxialRoPE(d_head // 2, self.n_heads) |
| 378 | self.dropout = nn.Dropout(dropout) |
| 379 | self.out_proj = apply_wd(zero_init(Linear(d_model, d_model, bias=False))) |
| 380 | |
| 381 | def extra_repr(self): |
| 382 | return f"d_head={self.d_head}," |