Method forward

gpu/model.py:283–296

(
        self,
        token_values: torch.Tensor,
        token_lengths: torch.Tensor,
        start_pos: torch.Tensor,
        cache: list[LayerCache],
        kv_padding: int,
    )

        return logits.float()

    def forward(
        self,
        token_values: torch.Tensor,
        token_lengths: torch.Tensor,
        start_pos: torch.Tensor,
        cache: list[LayerCache],
        kv_padding: int,
    ) -> torch.Tensor:
        attn_bias = AttnBias.from_seqlens(
            q_seqlen=token_lengths.tolist(),
            kv_seqlen=(start_pos + token_lengths).tolist(),
            kv_padding=kv_padding,
        )
        return self.forward_with_attn_bias(token_values, attn_bias, cache)


def make_cache(