vllm.model_executor.models.kimi_k25_vit ¶

class KimiK25MultiModalProjector(nn.Module):
    """Multi-modal projector with patch merging for Kimi-K2.5."""

    def __init__(
        self,
        config: KimiK25VisionConfig,
        use_data_parallel: bool = False,
        prefix: str = "",
    ):
        super().__init__()
        self.use_data_parallel = use_data_parallel

        # Hidden size after patch merging
        merge_h, merge_w = config.merge_kernel_size
        self.hidden_size = config.hidden_size * merge_h * merge_w

        self.pre_norm = torch.nn.LayerNorm(config.hidden_size, eps=1e-5)
        self.linear_1 = ReplicatedLinear(
            self.hidden_size,
            self.hidden_size,
            bias=True,
            prefix=f"{prefix}.linear_1",
        )
        self.linear_2 = ReplicatedLinear(
            self.hidden_size,
            config.mm_hidden_size,
            bias=True,
            prefix=f"{prefix}.linear_2",
        )
        self.act = GELUActivation()

    def forward(self, image_features: torch.Tensor) -> torch.Tensor:
        hidden_states = self.pre_norm(image_features).view(-1, self.hidden_size)
        hidden_states, _ = self.linear_1(hidden_states)
        hidden_states = self.act(hidden_states)
        hidden_states, _ = self.linear_2(hidden_states)
        return hidden_states

class Learnable2DInterpPosEmbDivided_fixed(nn.Module):
    """2D learnable position embedding with temporal extension."""

    def __init__(
        self,
        height: int,
        width: int,
        num_frames: int,
        dim: int,
        interpolation_mode: str = "bicubic",
    ) -> None:
        super().__init__()
        self.height = height
        self.width = width
        self.num_frames = num_frames
        self.dim = dim
        self.interpolation_mode = interpolation_mode
        self.weight = nn.Parameter(torch.empty(height, width, dim))
        self.register_buffer(
            "time_weight",
            torch.from_numpy(get_1d_sincos_pos_embed(self.dim, self.num_frames))
            .float()
            .unsqueeze(1),
            persistent=False,
        )

        self.reset_parameters()

    def reset_parameters(self):
        nn.init.normal_(self.weight)

    def forward(self, x: torch.Tensor, grid_thws: torch.Tensor) -> torch.Tensor:
        pos_embs = []
        for t, h, w in grid_thws.tolist():
            assert t <= self.num_frames, f"t:{t} > self.num_frames:{self.num_frames}"
            if (h, w) == self.weight.shape[:-1]:
                pos_emb_2d = self.weight.flatten(end_dim=1)
            else:
                pos_emb_2d = get_rope_shape(
                    self.weight,
                    interpolation_mode=self.interpolation_mode,
                    shape=(h, w),
                )

            if t == 1:
                pos_emb_3d = pos_emb_2d
            else:
                pos_emb_3d = (
                    pos_emb_2d.unsqueeze(0).repeat(t, 1, 1) + self.time_weight[0:t]
                )

            pos_embs.append(pos_emb_3d.reshape(-1, pos_emb_3d.shape[-1]))

        out = x + torch.cat(pos_embs)
        return out

class MLP2(nn.Module):
    """Two-layer MLP with tensor parallel support."""

    def __init__(
        self,
        dims: list[int],
        activation,
        bias: bool = True,
        prefix: str = "",
        use_data_parallel: bool = False,
    ):
        super().__init__()
        assert len(dims) == 3
        self.use_data_parallel = use_data_parallel
        self.fc0 = ColumnParallelLinear(
            dims[0],
            dims[1],
            bias=bias,
            prefix=maybe_prefix(prefix, "fc0"),
            disable_tp=self.use_data_parallel,
        )
        self.fc1 = RowParallelLinear(
            dims[1],
            dims[2],
            bias=bias,
            prefix=maybe_prefix(prefix, "fc1"),
            disable_tp=self.use_data_parallel,
        )
        self.activation = activation

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x, _ = self.fc0(x)
        x = self.activation(x)
        x, _ = self.fc1(x)
        return x

class MoonViT3dEncoder(nn.Module):
    """Full encoder stack for MoonViT 3D."""

    def __init__(
        self,
        hidden_dim: int,
        num_layers: int,
        block_cfg: dict,
        video_attn_type: str = "spatial_temporal",
        prefix: str = "",
    ) -> None:
        super().__init__()

        assert video_attn_type == "spatial_temporal", (
            f'video_attn_type must be "spatial_temporal", got {video_attn_type}'
        )
        self.video_attn_type = video_attn_type
        self.rope_2d = Rope2DPosEmbRepeated(
            block_cfg["hidden_dim"] // block_cfg["num_heads"], 512, 512
        )
        self.blocks = nn.ModuleList(
            [
                MoonViTEncoderLayer(
                    **block_cfg,
                    prefix=f"{prefix}.blocks.{layer_idx}",
                )
                for layer_idx in range(num_layers)
            ]
        )
        self.final_layernorm = nn.LayerNorm(hidden_dim)

    def forward(
        self,
        hidden_states: torch.Tensor,
        grid_thws: torch.Tensor,
    ) -> torch.Tensor:
        rope_freqs_cis = self.rope_2d.get_freqs_cis(
            grid_thws=grid_thws, device=hidden_states.device
        )

        lengths = torch.cat(
            (
                torch.zeros(1, dtype=grid_thws.dtype, device=grid_thws.device),
                grid_thws[:, 0] * grid_thws[:, 1] * grid_thws[:, 2],
            )
        )

        cu_seqlens = lengths.to(hidden_states.device).cumsum(dim=0, dtype=torch.int32)

        for block in self.blocks:
            hidden_states = block(
                hidden_states, cu_seqlens, rope_freqs_cis=rope_freqs_cis
            )

        hidden_states = self.final_layernorm(hidden_states)

        return hidden_states

class MoonViT3dPretrainedModel(nn.Module):
    """Main vision tower model.

    Uses KimiK25VisionConfig directly from transformers_utils/configs/kimi_k25.py.
    """

    def __init__(
        self,
        config: KimiK25VisionConfig,
        prefix: str = "",
    ):
        super().__init__()
        config = deepcopy(config)
        self.config = config  # Required for run_dp_sharded_mrope_vision_model
        self.merge_kernel_size = config.merge_kernel_size
        self.patch_size = config.patch_size
        self.merge_type = config.merge_type

        self.patch_embed = MoonVision3dPatchEmbed(
            out_dim=config.hidden_size,
            patch_size=config.patch_size,
            pos_emb_height=config.init_pos_emb_height,
            pos_emb_width=config.init_pos_emb_width,
            pos_emb_time=config.init_pos_emb_time,
            pos_emb_type=config.pos_emb_type,
        )

        self.encoder = MoonViT3dEncoder(
            hidden_dim=config.hidden_size,
            num_layers=config.num_hidden_layers,
            block_cfg={
                "num_heads": config.num_attention_heads,
                "hidden_dim": config.hidden_size,
                "mlp_dim": config.intermediate_size,
                "activation": get_act_fn("gelu_pytorch_tanh"),
                "attn_bias": True,
            },
            video_attn_type=config.video_attn_type,
            prefix=maybe_prefix(prefix, "encoder"),
        )

    def forward(
        self, pixel_values: torch.Tensor, grid_thws: torch.Tensor
    ) -> torch.Tensor:
        """
        Args:
            pixel_values (torch.Tensor): The input pixel values.
            grid_thws (torch.Tensor): Temporal, height and width.

        Returns:
            torch.Tensor: The output tokens.
        """
        hidden_states = self.patch_embed(pixel_values, grid_thws)
        hidden_states = self.encoder(hidden_states, grid_thws)
        if (
            self.merge_type == "sd2_tpool"
        ):  # spatial downsampling 2x with temporal pooling all
            hidden_states = tpool_patch_merger(
                hidden_states, grid_thws, merge_kernel_size=self.merge_kernel_size
            )
        else:
            raise NotImplementedError(f"Not support {self.merge_type}")

        return hidden_states

class MoonViTEncoderLayer(nn.Module):
    """Single encoder layer for MoonViT with TP/DP support."""

    def __init__(
        self,
        num_heads: int,
        hidden_dim: int,
        mlp_dim: int,
        prefix: str = "",
        *,
        activation=F.gelu,
        attn_bias: bool = False,
    ):
        super().__init__()
        self.use_data_parallel = is_vit_use_data_parallel()

        self.num_heads = num_heads
        self.hidden_dim = hidden_dim
        self.hidden_size_per_attention_head = self.hidden_dim // self.num_heads
        self.tp_size = (
            1 if self.use_data_parallel else get_tensor_model_parallel_world_size()
        )
        self.num_attention_heads_per_partition = divide(num_heads, self.tp_size)

        self.norm0 = nn.LayerNorm(hidden_dim)
        self.norm1 = nn.LayerNorm(hidden_dim)
        self.mlp = MLP2(
            [hidden_dim, mlp_dim, hidden_dim],
            activation,
            prefix=f"{prefix}.mlp",
            use_data_parallel=self.use_data_parallel,
        )
        self.wqkv = QKVParallelLinear(
            hidden_size=hidden_dim,
            head_size=self.hidden_size_per_attention_head,
            total_num_heads=num_heads,
            total_num_kv_heads=num_heads,
            bias=attn_bias,
            prefix=f"{prefix}.wqkv",
            disable_tp=self.use_data_parallel,
        )
        self.wo = RowParallelLinear(
            hidden_dim,
            hidden_dim,
            bias=attn_bias,
            prefix=f"{prefix}.wo",
            disable_tp=self.use_data_parallel,
        )
        self.attn = MMEncoderAttention(
            num_heads=self.num_attention_heads_per_partition,
            head_size=self.hidden_size_per_attention_head,
            scale=self.hidden_size_per_attention_head**-0.5,
            prefix=f"{prefix}.attn",
        )

    def attention_qkvpacked(
        self,
        x: torch.Tensor,
        cu_seqlens: torch.Tensor,
        rope_freqs_cis: torch.Tensor | None = None,
    ):
        """Compute self-attention with packed QKV.

        Args:
            x (torch.Tensor): (seqlen, hidden_dim)
            cu_seqlens (torch.Tensor): cumulative sequence lengths
        """
        seq_length = x.size(0)
        xqkv, _ = self.wqkv(x)

        qkv_shape = xqkv.size()[:-1] + (
            3,
            self.num_attention_heads_per_partition,
            self.hidden_size_per_attention_head,
        )
        # xqkv: (seqlen, 3, nheads, headdim)
        xqkv = xqkv.view(*qkv_shape)
        xq, xk, xv = torch.unbind(xqkv, dim=-3)

        xq, xk = apply_rope(xq, xk, rope_freqs_cis)

        max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
        attn_out = self.attn(
            xq.unsqueeze(0),
            xk.unsqueeze(0),
            xv.unsqueeze(0),
            cu_seqlens=cu_seqlens,
            max_seqlen=max_seqlen,
        )
        attn_out = attn_out.reshape(
            seq_length,
            self.num_attention_heads_per_partition
            * self.hidden_size_per_attention_head,
        )
        attn_out, _ = self.wo(attn_out)
        return attn_out

    def forward(
        self,
        hidden_states: torch.Tensor,
        cu_seqlens: torch.Tensor,
        rope_freqs_cis: torch.Tensor | None = None,
    ):
        residual = hidden_states
        hidden_states = self.norm0(hidden_states)

        hidden_states = self.attention_qkvpacked(
            hidden_states, cu_seqlens, rope_freqs_cis
        )
        hidden_states = residual + hidden_states

        residual = hidden_states
        hidden_states = self.norm1(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = residual + hidden_states

        return hidden_states