Method concat_encodings

library/strategy_sd3.py:249–255

(
        self, lg_out: torch.Tensor, t5_out: Optional[torch.Tensor], lg_pooled: torch.Tensor
    )

        return [lg_out, t5_out, lg_pooled, l_attn_mask, g_attn_mask, t5_attn_mask]

    def concat_encodings(
        self, lg_out: torch.Tensor, t5_out: Optional[torch.Tensor], lg_pooled: torch.Tensor
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        lg_out = torch.nn.functional.pad(lg_out, (0, 4096 - lg_out.shape[-1]))
        if t5_out is None:
            t5_out = torch.zeros((lg_out.shape[0], 77, 4096), device=lg_out.device, dtype=lg_out.dtype)
        return torch.cat([lg_out, t5_out], dim=-2), lg_pooled


class Sd3TextEncoderOutputsCachingStrategy(TextEncoderOutputsCachingStrategy):