vllm.model_executor.models.siglip2navit

class Siglip2Attention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(
        self,
        config: Siglip2VisionConfig,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
        use_data_parallel: bool = False,
    ):
        super().__init__()
        self.config = config
        self.embed_dim = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.head_dim = self.embed_dim // self.num_heads
        if self.head_dim * self.num_heads != self.embed_dim:
            raise ValueError(
                f"embed_dim must be divisible by num_heads "
                f"(got `embed_dim`: {self.embed_dim} and `num_heads`:"
                f" {self.num_heads}).")
        self.scale = self.head_dim**-0.5
        self.dropout = config.attention_dropout
        self.is_causal = False

        # TODO(Isotr0py): Enable data parallel after we support
        # disabling TP on parallel linear layer
        self.qkv_proj = QKVParallelLinear(
            hidden_size=self.embed_dim,
            head_size=self.head_dim,
            total_num_heads=self.num_heads,
            quant_config=quant_config,
            prefix=f"{prefix}.qkv_proj",
        )
        self.out_proj = RowParallelLinear(
            input_size=self.embed_dim,
            output_size=self.embed_dim,
            quant_config=quant_config,
            prefix=f"{prefix}.out_proj",
        )

        self.tp_size = (1 if use_data_parallel else
                        get_tensor_model_parallel_world_size())
        self.num_heads_per_partition = divide(self.num_heads, self.tp_size)
        self.use_rope = config.use_rope

        # Detect attention implementation.
        self.attn_backend: _Backend = get_vit_attn_backend(support_fa=True)
        if self.attn_backend not in {
                _Backend.FLASH_ATTN, _Backend.TORCH_SDPA,
                _Backend.ROCM_AITER_FA
        }:
            self.attn_backend = _Backend.TORCH_SDPA
        self.is_flash_attn_backend = self.attn_backend in {
            _Backend.FLASH_ATTN, _Backend.ROCM_AITER_FA
        }

    def forward(
        self,
        hidden_states: torch.Tensor,
        cu_seqlens: torch.Tensor,
        position_embeddings: Optional[tuple[torch.Tensor,
                                            torch.Tensor]] = None,
    ) -> tuple[torch.Tensor, Optional[torch.Tensor]]:
        """Input shape: Batch x Time x Channel"""

        seq_length, embed_dim = hidden_states.shape

        qkv_states, _ = self.qkv_proj(hidden_states)
        queries, keys, values = qkv_states.chunk(3, dim=-1)

        queries = queries.view(seq_length, self.num_heads_per_partition,
                               self.head_dim)
        keys = keys.view(seq_length, self.num_heads_per_partition,
                         self.head_dim)
        values = values.view(seq_length, self.num_heads_per_partition,
                             self.head_dim)

        if self.use_rope:
            cos, sin = position_embeddings
            queries, keys = apply_rotary_pos_emb(queries.unsqueeze(0),
                                                 keys.unsqueeze(0), cos, sin,
                                                 self.is_flash_attn_backend)
            queries = queries.squeeze(0)
            keys = keys.squeeze(0)

        max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max().item()
        if self.is_flash_attn_backend:
            if self.attn_backend == _Backend.ROCM_AITER_FA:
                from aiter import flash_attn_varlen_func
            else:
                from flash_attn import flash_attn_varlen_func
            attn_output = flash_attn_varlen_func(
                queries, keys, values, cu_seqlens, cu_seqlens, max_seqlen,
                max_seqlen).reshape(seq_length, -1)
        elif self.attn_backend == _Backend.TORCH_SDPA:
            # Execute attention entry by entry for speed & less VRAM.
            batch_size = cu_seqlens.shape[0] - 1
            outputs = []
            cu = cu_seqlens.tolist()
            for i in range(batch_size):
                start_idx = cu[i]
                end_idx = cu[i + 1]

                # Each sequence is processed independently.
                q_i = queries[start_idx:end_idx].unsqueeze(0)
                k_i = keys[start_idx:end_idx].unsqueeze(0)
                v_i = values[start_idx:end_idx].unsqueeze(0)

                # (1, seq_len, num_heads, head_dim) ->
                # (1, num_heads, seq_len, head_dim)
                q_i, k_i, v_i = [x.transpose(1, 2) for x in (q_i, k_i, v_i)]

                output_i = F.scaled_dot_product_attention(q_i,
                                                          k_i,
                                                          v_i,
                                                          dropout_p=0.0)
                # (1, num_heads, seq_len, head_dim) -> (seq_len, embed_dim)
                output_i = output_i.transpose(1, 2).reshape(
                    end_idx - start_idx, -1)
                outputs.append(output_i)

            attn_output = torch.cat(outputs, dim=0)
        attn_output, _ = self.out_proj(attn_output)
        return attn_output

class Siglip2Encoder(nn.Module):
    """
    Transformer encoder consisting of `config.num_hidden_layers` 
    self attention layers. Each layer is a [`Siglip2EncoderLayer`].

    Args:
        config: PretrainedConfig
    """

    def __init__(
        self,
        config: Siglip2VisionConfig,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
        use_data_parallel: bool = False,
    ):
        super().__init__()
        self.config = config
        self.layers = nn.ModuleList([
            Siglip2EncoderLayer(config,
                                quant_config=quant_config,
                                prefix=f"{prefix}.layers.{idx}",
                                use_data_parallel=use_data_parallel)
            for idx in range(config.num_hidden_layers)
        ])

        self.rotary_pos_emb = VisionRotaryEmbedding(
            config.hidden_size // config.num_attention_heads // 2)
        self.patch_size = config.patch_size
        self.hidden_stride = config.hidden_stride
        self.window_size = config.window_size
        self.spatial_merge_unit = config.hidden_stride * config.hidden_stride
        if config.fullatt_block_indexes is None:
            self.fullatt_block_indexes = None
        else:
            self.fullatt_block_indexes = [
                int(i) for i in config.fullatt_block_indexes.split('|')
            ]

    # copied from qwen2.5_vl
    def rot_pos_emb(self, grid_thw):
        pos_ids = []
        for t, h, w in grid_thw:
            hpos_ids = torch.arange(h).unsqueeze(1).expand(-1, w)
            hpos_ids = hpos_ids.reshape(
                h // self.hidden_stride,
                self.hidden_stride,
                w // self.hidden_stride,
                self.hidden_stride,
            )
            hpos_ids = hpos_ids.permute(0, 2, 1, 3)
            hpos_ids = hpos_ids.flatten()

            wpos_ids = torch.arange(w).unsqueeze(0).expand(h, -1)
            wpos_ids = wpos_ids.reshape(
                h // self.hidden_stride,
                self.hidden_stride,
                w // self.hidden_stride,
                self.hidden_stride,
            )
            wpos_ids = wpos_ids.permute(0, 2, 1, 3)
            wpos_ids = wpos_ids.flatten()
            pos_ids.append(
                torch.stack([hpos_ids, wpos_ids], dim=-1).repeat(t, 1))
        pos_ids = torch.cat(pos_ids, dim=0)
        max_grid_size = grid_thw[:, 1:].max()
        rotary_pos_emb_full = self.rotary_pos_emb(max_grid_size)
        rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
        return rotary_pos_emb

    def get_window_index(self, grid_thw):
        window_index: list = []
        cu_window_seqlens: list = [0]
        window_index_id = 0
        # patch (after merge) number in each window
        vit_merger_window_size = (self.window_size // self.hidden_stride //
                                  self.patch_size)

        for grid_t, grid_h, grid_w in grid_thw:
            llm_grid_h, llm_grid_w = (
                grid_h // self.hidden_stride,  # number of patch after merge
                grid_w // self.hidden_stride,
            )
            index = torch.arange(grid_t * llm_grid_h * llm_grid_w).reshape(
                grid_t, llm_grid_h, llm_grid_w)
            pad_h = vit_merger_window_size - llm_grid_h % vit_merger_window_size
            pad_w = vit_merger_window_size - llm_grid_w % vit_merger_window_size
            num_windows_h = (llm_grid_h + pad_h) // vit_merger_window_size
            num_windows_w = (llm_grid_w + pad_w) // vit_merger_window_size
            index_padded = F.pad(index, (0, pad_w, 0, pad_h), "constant", -100)
            index_padded = index_padded.reshape(
                grid_t,
                num_windows_h,
                vit_merger_window_size,
                num_windows_w,
                vit_merger_window_size,
            )
            index_padded = index_padded.permute(0, 1, 3, 2, 4).reshape(
                grid_t,
                num_windows_h * num_windows_w,
                vit_merger_window_size,
                vit_merger_window_size,
            )
            seqlens = (index_padded != -100).sum([2, 3]).reshape(-1)
            index_padded = index_padded.reshape(-1)
            index_new = index_padded[index_padded != -100]
            window_index.append(index_new + window_index_id)
            cu_seqlens_tmp = seqlens.cumsum(
                0) * self.spatial_merge_unit + cu_window_seqlens[-1]
            cu_window_seqlens.extend(cu_seqlens_tmp.tolist())
            window_index_id += (grid_t * llm_grid_h * llm_grid_w).item()
        window_index = torch.cat(window_index, dim=0)

        return window_index, cu_window_seqlens

    def forward(
        self,
        inputs_embeds: torch.Tensor,
        grid_thws: torch.Tensor,
    ) -> torch.Tensor:
        r"""
        Args:
            inputs_embeds (`torch.FloatTensor` of shape
                `(batch_size, sequence_length, hidden_size)`):
                Optionally, instead of passing `input_ids` you can choose to
                directly pass an embedded representation. This is useful if
                you want more control over how to convert `input_ids` indices
                into associated vectors than the model's internal embedding
                lookup matrix.
            grid_thws (`torch.LongTensor`):
                grid shape (num_patches, 3)
            output_hidden_states (`bool`, *optional*):
                Whether or not to return the hidden states of all layers. See
                `hidden_states` under returned tensors for more detail.
            return_dict (`bool`, *optional*):
                Whether or not to return a [`~utils.ModelOutput`] instead of
                a plain tuple.
        """
        rotary_pos_emb = self.rot_pos_emb(grid_thws)
        window_index, cu_window_seqlens = self.get_window_index(grid_thws)
        cu_window_seqlens = torch.tensor(
            cu_window_seqlens,
            device=inputs_embeds.device,
            dtype=grid_thws.dtype if torch.jit.is_tracing() else torch.int32,
        )
        cu_window_seqlens = torch.unique_consecutive(cu_window_seqlens)

        seq_len, _ = inputs_embeds.size()
        inputs_embeds = inputs_embeds.reshape(
            seq_len // self.spatial_merge_unit, self.spatial_merge_unit, -1)
        inputs_embeds = inputs_embeds[window_index, :, :]
        inputs_embeds = inputs_embeds.reshape(seq_len, -1)
        rotary_pos_emb = rotary_pos_emb.reshape(
            seq_len // self.spatial_merge_unit, self.spatial_merge_unit, -1)
        rotary_pos_emb = rotary_pos_emb[window_index, :, :]
        rotary_pos_emb = rotary_pos_emb.reshape(seq_len, -1)
        emb = torch.cat((rotary_pos_emb, rotary_pos_emb), dim=-1)
        position_embeddings = (emb.cos(), emb.sin())

        cu_seqlens = torch.repeat_interleave(
            grid_thws[:, 1] * grid_thws[:, 2], grid_thws[:, 0]
        ).cumsum(
            dim=0,
            # Select dtype based on the following factors:
            #  - FA2 requires that cu_seqlens_q must have dtype int32
            #  - torch.onnx.export requires that cu_seqlens_q must have
            #    same dtype as grid_thw
            # See https://github.com/huggingface/transformers/pull/34852
            # for more information
            dtype=grid_thws.dtype if torch.jit.is_tracing() else torch.int32,
        )
        cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0)

        reverse_indices = torch.argsort(window_index)

        hidden_states = inputs_embeds
        for index, block in enumerate(self.layers):
            if (not self.fullatt_block_indexes
                    or index in self.fullatt_block_indexes):
                cu_seqlens_tmp = cu_seqlens
            else:
                cu_seqlens_tmp = cu_window_seqlens
            hidden_states = block(hidden_states, cu_seqlens_tmp,
                                  position_embeddings)

        hidden_states = hidden_states.reshape(
            seq_len // self.spatial_merge_unit, self.spatial_merge_unit, -1)
        hidden_states = hidden_states[reverse_indices, :].reshape(seq_len, -1)

        return hidden_states

class Siglip2EncoderLayer(nn.Module):

    def __init__(
        self,
        config: Siglip2VisionConfig,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
        use_data_parallel: bool = False,
    ):
        super().__init__()
        self.embed_dim = config.hidden_size
        self.layer_norm1 = nn.LayerNorm(self.embed_dim,
                                        eps=config.layer_norm_eps)
        self.self_attn = Siglip2Attention(config,
                                          quant_config=quant_config,
                                          prefix=f"{prefix}.self_attn",
                                          use_data_parallel=use_data_parallel)
        self.layer_norm2 = nn.LayerNorm(self.embed_dim,
                                        eps=config.layer_norm_eps)
        self.mlp = Siglip2MLP(config,
                              quant_config=quant_config,
                              prefix=f"{prefix}.mlp",
                              use_data_parallel=use_data_parallel)

    def forward(self, hidden_states: torch.Tensor, cu_seqlens: torch.Tensor,
                position_embeddings: torch.Tensor) -> tuple[torch.FloatTensor]:
        """
        Args:
            hidden_states (`torch.FloatTensor`):
                Input to the layer of shape `(batch, seq_len, embed_dim)`.
            output_attentions (`bool`, *optional*, defaults to `False`):
                Whether or not to return the attentions tensors of all 
                attention layers. See `attentions` under
                returned tensors for more detail.
        """
        residual = hidden_states

        hidden_states = self.layer_norm1(hidden_states)
        hidden_states = self.self_attn(hidden_states=hidden_states,
                                       cu_seqlens=cu_seqlens,
                                       position_embeddings=position_embeddings)
        hidden_states = residual + hidden_states

        residual = hidden_states
        hidden_states = self.layer_norm2(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = residual + hidden_states
        return hidden_states