`fastvideo.models.dits.matrixgame2.model.DistributedAttention`

DistributedAttention(num_heads: int, head_size: int, num_kv_heads: int | None = None, softmax_scale: float | None = None, causal: bool = False, supported_attention_backends: tuple[AttentionBackendEnum, ...] | None = None, prefix: str = '', **extra_impl_args)

Distributed attention layer.

`fastvideo.models.dits.matrixgame2.model.DistributedAttention.forward`

forward(q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, original_seq_len: int | None = None, replicated_q: torch.Tensor | None = None, replicated_k: torch.Tensor | None = None, replicated_v: torch.Tensor | None = None, freqs_cis: tuple[torch.Tensor, torch.Tensor] | None = None, attn_metadata: AttentionMetadata | None = None) -> tuple[torch.Tensor, torch.Tensor | None]

Forward pass for distributed attention.

Parameters

q: torch.Tensor: Query tensor [batch_size, seq_len, num_heads, head_dim]
k: torch.Tensor: Key tensor [batch_size, seq_len, num_heads, head_dim]
v: torch.Tensor: Value tensor [batch_size, seq_len, num_heads, head_dim]
original_seq_len: int: Original (unpadded) full sequence length
replicated_q: Optional[torch.Tensor]: Replicated query tensor, typically for text tokens
replicated_k: Optional[torch.Tensor]: Replicated key tensor
replicated_v: Optional[torch.Tensor]: Replicated value tensor

Returns

tuple[torch.Tensor, torch.Tensor | None]: Tuple[torch.Tensor, Optional[torch.Tensor]]: A tuple containing: - o (torch.Tensor): Output tensor after attention for the main sequence - replicated_o (Optional[torch.Tensor]): Output tensor for replicated tokens, if provided

`fastvideo.models.dits.matrixgame2.model.MatrixGame2WanVideoConfig`

MatrixGame2WanVideoConfig(arch_config: MatrixGame2WanVideoArchConfig = MatrixGame2WanVideoArchConfig(), prefix: str = 'Wan', quant_config: QuantizationConfig | None = None, _compile_conditions: list = (lambda: [_is_transformer_block])(), *, _resolved_attention_backend: AttentionBackendEnum | None = None)

`fastvideo.models.dits.matrixgame2.model.get_sp_world_size`

get_sp_world_size() -> int

Return world size for the sequence model parallel group.

`fastvideo.models.dits.matrixgame2.model.FP32LayerNorm`

FP32LayerNorm()

`fastvideo.models.dits.matrixgame2.model.FP32LayerNorm.forward`

forward(inputs: torch.Tensor) -> torch.Tensor

`fastvideo.models.dits.matrixgame2.model.LayerNormScaleShift`

LayerNormScaleShift(hidden_size: int, norm_type: str = 'rms', eps: float = 1e-06, elementwise_affine: bool = False, dtype: torch.dtype = torch.float32, compute_dtype: torch.dtype | None = None, prefix: str = '')

Fused operation that combines LayerNorm with scale and shift operations. This reduces memory bandwidth by combining memory-bound operations.

`fastvideo.models.dits.matrixgame2.model.LayerNormScaleShift.forward`

forward(x: torch.Tensor, shift: torch.Tensor, scale: torch.Tensor, convert_modulation_dtype: bool = False) -> torch.Tensor

Apply ln followed by scale and shift in a single fused operation.

`fastvideo.models.dits.matrixgame2.model.RMSNorm`

RMSNorm(hidden_size: int, eps: float = 1e-06, dtype: torch.dtype = torch.float32, var_hidden_size: int | None = None, has_weight: bool = True)

Root mean square normalization.

Computes x -> w * x / sqrt(E[x^2] + eps) where w is the learned weight. Refer to https://arxiv.org/abs/1910.07467

`fastvideo.models.dits.matrixgame2.model.RMSNorm.forward_native`

forward_native(x: torch.Tensor, residual: torch.Tensor | None = None) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]

PyTorch-native implementation equivalent to forward().

`fastvideo.models.dits.matrixgame2.model.RMSNorm.extra_repr`

extra_repr() -> str

`fastvideo.models.dits.matrixgame2.model.ScaleResidual`

ScaleResidual(prefix: str = '')

Applies gated residual connection.

`fastvideo.models.dits.matrixgame2.model.ScaleResidual.forward`

forward(residual: torch.Tensor, x: torch.Tensor, gate: torch.Tensor) -> torch.Tensor

Apply gated residual connection.

`fastvideo.models.dits.matrixgame2.model.ScaleResidualLayerNormScaleShift`

ScaleResidualLayerNormScaleShift(hidden_size: int, norm_type: str = 'rms', eps: float = 1e-06, elementwise_affine: bool = False, dtype: torch.dtype = torch.float32, compute_dtype: torch.dtype | None = None, prefix: str = '', fuse_inference: bool = False)

Fused operation that combines: 1. Gated residual connection 2. LayerNorm 3. Scale and shift operations

This reduces memory bandwidth by combining memory-bound operations.

`fastvideo.models.dits.matrixgame2.model.ScaleResidualLayerNormScaleShift.forward`

forward(residual: torch.Tensor, x: torch.Tensor, gate: torch.Tensor | int, shift: torch.Tensor | None, scale: torch.Tensor | None, convert_modulation_dtype: bool = False) -> tuple[torch.Tensor, torch.Tensor]

Apply gated residual connection, followed by layernorm and scale/shift in a single fused operation.

Pass shift=None, scale=None (both, not just one) to skip the modulation entirely and return the normalized tensor as-is. Callers that need norm-only behavior (e.g. Wan's post-self-attn site) would otherwise pay two full-tensor elementwise passes computing normalized * (1.0 + 0) + 0.

Returns

torch.Tensor

Tuple containing:

torch.Tensor

normalized and modulated output

tuple[torch.Tensor, torch.Tensor]

residual value (value after residual connection but before normalization)

`fastvideo.models.dits.matrixgame2.model.ReplicatedLinear`

ReplicatedLinear(input_size: int, output_size: int, bias: bool = True, skip_bias_add: bool = False, params_dtype: torch.dtype | None = None, quant_config: QuantizationConfig | None = None, prefix: str = '')

Replicated linear layer.

Parameters

input_size: int: input dimension of the linear layer.
output_size: int: output dimension of the linear layer.
bias: bool: If true, add bias.
skip_bias_add: bool: If true, skip adding bias but instead return it.
params_dtype: torch.dtype | None: Data type for the parameters.
quant_config: QuantizationConfig | None: Quantization configure.
prefix: str: The name of the layer in the state dict, including all parents (e.g. model.layers.0.qkv_proj)

`fastvideo.models.dits.matrixgame2.model.ReplicatedLinear.weight_loader`

weight_loader(param: Parameter, loaded_weight: torch.Tensor) -> None

`fastvideo.models.dits.matrixgame2.model.ReplicatedLinear.forward`

forward(x: torch.Tensor) -> tuple[torch.Tensor, Parameter | None]

`fastvideo.models.dits.matrixgame2.model.ReplicatedLinear.extra_repr`

extra_repr() -> str

`fastvideo.models.dits.matrixgame2.model.ReplicatedLinear.get_shape_mapping`

get_shape_mapping() -> dict

Get the mapping from (input_shape, output_shape) to layer types.

`fastvideo.models.dits.matrixgame2.model.ReplicatedLinear.reset_shape_tracking`

reset_shape_tracking() -> None

Clear tracked shapes and layer type mappings.

`fastvideo.models.dits.matrixgame2.model.ReplicatedLinear.print_shape_summary`

print_shape_summary() -> None

Log a summary of all unique shapes and their layer types.

`fastvideo.models.dits.matrixgame2.model.MLP`

MLP(input_dim: int, mlp_hidden_dim: int, output_dim: int | None = None, bias: bool = True, act_type: str = 'gelu_pytorch_tanh', dtype: torch.dtype | None = None, prefix: str = '', quant_config: QuantizationConfig | None = None)

MLP for DiT blocks, NO gated linear units

`fastvideo.models.dits.matrixgame2.model.MLP.forward`

forward(x: torch.Tensor) -> torch.Tensor

`fastvideo.models.dits.matrixgame2.model.get_rotary_pos_embed`

get_rotary_pos_embed(rope_sizes, hidden_size, heads_num, rope_dim_list, rope_theta, theta_rescale_factor = 1.0, interpolation_factor = 1.0, shard_dim: int = 0, do_sp_sharding: bool = False, dtype: torch.dtype = torch.float32, start_frame: int = 0, use_real: bool = True) -> tuple[torch.Tensor, torch.Tensor]

Generate rotary positional embeddings for the given sizes.

Parameters

rope_sizes: Tuple of dimensions (t, h, w)
hidden_size: Hidden dimension size
heads_num: Number of attention heads
rope_dim_list: List of dimensions for each axis, or None
rope_theta: Base for frequency calculations
theta_rescale_factor: Rescale factor for theta. Defaults to 1.0
interpolation_factor: Factor to scale positions. Defaults to 1.0
shard_dim: int: Which dimension to shard for sequence parallelism. Defaults to 0.
do_sp_sharding: bool: Whether to shard the positional embeddings for sequence parallelism. Defaults to False.
use_real: bool: If True, output full head_dim for rotate_half style RoPE. Defaults to True.

Returns

tuple[torch.Tensor, torch.Tensor]: Tuple of (cos, sin) tensors for rotary embeddings. Shape [S, D] if use_real, [S, D/2] otherwise.

`fastvideo.models.dits.matrixgame2.model.PatchEmbed`

PatchEmbed(patch_size = 16, in_chans = 3, embed_dim = 768, norm_layer = None, flatten = True, bias = True, dtype = None, prefix: str = '')

2D Image to Patch Embedding

Image to Patch Embedding using Conv2d

A convolution based approach to patchifying a 2D image w/ embedding projection.

Based on the impl in https://github.com/google-research/vision_transformer

Remove the _assert function in forward function to be compatible with multi-resolution images.

`fastvideo.models.dits.matrixgame2.model.PatchEmbed.forward`

forward(x)

`fastvideo.models.dits.matrixgame2.model.TimestepEmbedder`

TimestepEmbedder(hidden_size, act_layer = 'silu', frequency_embedding_size = 256, max_period = 10000, dtype = None, freq_dtype = torch.float32, prefix: str = '')

Embeds scalar timesteps into vector representations.

`fastvideo.models.dits.matrixgame2.model.TimestepEmbedder.forward`

forward(t: torch.Tensor, timestep_seq_len: int | None = None) -> torch.Tensor

`fastvideo.models.dits.matrixgame2.model.ModulateProjection`

ModulateProjection(hidden_size: int, factor: int = 2, act_layer: str = 'silu', dtype: torch.dtype | None = None, prefix: str = '')

Modulation layer for DiT blocks.

`fastvideo.models.dits.matrixgame2.model.ModulateProjection.forward`

forward(x: torch.Tensor) -> torch.Tensor

`fastvideo.models.dits.matrixgame2.model.init_logger`

init_logger(name: str) -> _FastvideoLogger

The main purpose of this function is to ensure that loggers are retrieved in such a way that we can be sure the root fastvideo logger has already been configured.

`fastvideo.models.dits.matrixgame2.model.BaseDiT`

BaseDiT(config: DiTConfig, hf_config: dict[str, Any], **kwargs)

`fastvideo.models.dits.matrixgame2.model.BaseDiT.forward`

forward(hidden_states: torch.Tensor, encoder_hidden_states: torch.Tensor | list[torch.Tensor], timestep: torch.LongTensor, encoder_hidden_states_image: torch.Tensor | list[torch.Tensor] | None = None, guidance = None, **kwargs) -> torch.Tensor

`fastvideo.models.dits.matrixgame2.model.BaseDiT.preprocess_loaded_weight`

preprocess_loaded_weight(name: str, tensor: torch.Tensor) -> torch.Tensor

Adapt a checkpoint tensor before sharding; default models are unchanged.

`fastvideo.models.dits.matrixgame2.model.BaseDiT.inference_config_overrides`

inference_config_overrides() -> dict[str, Any]

Extra config.json keys a deployable checkpoint of this model needs.

`fastvideo.models.dits.matrixgame2.model.BaseDiT.inference_export_expected_shape`

inference_export_expected_shape(output_key: str, base_shape: tuple[int, ...]) -> tuple[int, ...]

The exported shape of output_key given the base checkpoint's shape.

`fastvideo.models.dits.matrixgame2.model.WanSelfAttention`

WanSelfAttention(dim: int, num_heads: int, window_size = (-1, -1), qk_norm = True, eps = 1e-06, parallel_attention = False, quant_config: QuantizationConfig | None = None, prefix: str = '')

`fastvideo.models.dits.matrixgame2.model.WanSelfAttention.forward`

forward(x: torch.Tensor, context: torch.Tensor, context_lens: int)

Parameters

x: Tensor: Shape [B, L, num_heads, C / num_heads]
seq_lens: Tensor: Shape [B]
grid_sizes: Tensor: Shape [B, 3], the second dimension contains (F, H, W)
freqs: Tensor: Rope freqs, shape [1024, C / num_heads / 2]

`fastvideo.models.dits.matrixgame2.model.WanI2VCrossAttention`

WanI2VCrossAttention(dim: int, num_heads: int, window_size = (-1, -1), qk_norm = True, eps = 1e-06, supported_attention_backends: tuple[AttentionBackendEnum, ...] | None = None, quant_config: QuantizationConfig | None = None, prefix: str = '')

`fastvideo.models.dits.matrixgame2.model.WanI2VCrossAttention.forward`

forward(x, context, context_lens)

Parameters

x: Tensor: Shape [B, L1, C]
context: Tensor: Shape [B, L2, C]
context_lens: Tensor: Shape [B]

`fastvideo.models.dits.matrixgame2.model.WanT2VCrossAttention`

WanT2VCrossAttention(dim: int, num_heads: int, window_size = (-1, -1), qk_norm = True, eps = 1e-06, parallel_attention = False, quant_config: QuantizationConfig | None = None, prefix: str = '')

`fastvideo.models.dits.matrixgame2.model.WanT2VCrossAttention.forward`

forward(x, context, context_lens, crossattn_cache = None)

Parameters

x: Tensor: Shape [B, L1, C]
context: Tensor: Shape [B, L2, C]
context_lens: Tensor: Shape [B]

`fastvideo.models.dits.matrixgame2.model.WanImageEmbedding`

WanImageEmbedding(in_features: int, out_features: int)

`fastvideo.models.dits.matrixgame2.model.WanImageEmbedding.forward`

forward(encoder_hidden_states_image: torch.Tensor) -> torch.Tensor

`fastvideo.models.dits.matrixgame2.model.AttentionBackendEnum`

AttentionBackendEnum()

`fastvideo.models.dits.matrixgame2.model.ActionModule`

ActionModule(mouse_dim_in: int = 2, keyboard_dim_in: int = 6, hidden_size: int = 128, img_hidden_size: int = 1536, keyboard_hidden_dim: int = 1024, mouse_hidden_dim: int = 1024, vae_time_compression_ratio: int = 4, windows_size: int = 3, heads_num: int = 16, patch_size: list | None = None, qk_norm: bool = True, qkv_bias: bool = False, rope_dim_list: list | None = None, rope_theta = 256, mouse_qk_dim_list: list | None = None, enable_mouse = True, enable_keyboard = True, local_attn_size = 6, sink_size = 0, blocks: list | None = None)

action module from https://arxiv.org/pdf/2501.08325

`fastvideo.models.dits.matrixgame2.model.ActionModule.patchify`

patchify(x, patch_size)

x : (N C T H W)

`fastvideo.models.dits.matrixgame2.model.ActionModule.unpatchify`

unpatchify(x, t, h, w, patch_size)

x: (N, T, patch_size**2 * C) imgs: (N, H, W, C)

`fastvideo.models.dits.matrixgame2.model.ActionModule.get_rotary_pos_embed`

get_rotary_pos_embed(video_length, height, width, head_dim, rope_dim_list = None, start_offset = 0)

`fastvideo.models.dits.matrixgame2.model.ActionModule.forward`

forward(x: torch.Tensor, tt: int, th: int, tw: int, mouse_condition: torch.Tensor | None = None, keyboard_condition: torch.Tensor | None = None, block_mask_mouse: BlockMask | None = None, block_mask_keyboard: BlockMask | None = None, is_causal: bool = False, kv_cache_mouse: dict[str, torch.Tensor] | None = None, kv_cache_keyboard: dict[str, torch.Tensor] | None = None, start_frame: int = 0, use_rope_keyboard: bool = True, num_frame_per_block: int = 3)

hidden_states: B, ttthtw, C mouse_condition: B, N_frames, C1 keyboard_condition: B, N_frames, C2

`fastvideo.models.dits.matrixgame2.model.MatrixGame2TimeImageEmbedding`

MatrixGame2TimeImageEmbedding(dim: int, time_freq_dim: int, image_embed_dim: int | None = None)

Source

class MatrixGame2TimeImageEmbedding(nn.Module):

    def __init__(
        self,
        dim: int,
        time_freq_dim: int,
        image_embed_dim: int | None = None,
    ):
        super().__init__()

        self.time_embedder = TimestepEmbedder(dim, frequency_embedding_size=time_freq_dim, act_layer="silu")
        self.time_modulation = ModulateProjection(dim, factor=6, act_layer="silu")

        self.image_embedder = None
        if image_embed_dim is not None:
            self.image_embedder = WanImageEmbedding(image_embed_dim, dim)

    def forward(
        self,
        timestep: torch.Tensor,
        encoder_hidden_states: torch.Tensor,  # Kept for interface compatibility
        encoder_hidden_states_image: torch.Tensor | None = None,
        timestep_seq_len: int | None = None,
    ):
        temb = self.time_embedder(timestep, timestep_seq_len)
        timestep_proj = self.time_modulation(temb)

        # Matrix-Game 2.0 does not use text embeddings, so we ignore encoder_hidden_states

        if encoder_hidden_states_image is not None:
            assert self.image_embedder is not None
            encoder_hidden_states_image = self.image_embedder(encoder_hidden_states_image)

        encoder_hidden_states = torch.zeros(
            (timestep.shape[0], 0, temb.shape[-1]),
            device=temb.device,
            dtype=temb.dtype,
        )

        return (
            temb,
            timestep_proj,
            encoder_hidden_states,
            encoder_hidden_states_image,
        )

`fastvideo.models.dits.matrixgame2.model.MatrixGame2TimeImageEmbedding.forward`

forward(timestep: torch.Tensor, encoder_hidden_states: torch.Tensor, encoder_hidden_states_image: torch.Tensor | None = None, timestep_seq_len: int | None = None)

Source

def forward(
    self,
    timestep: torch.Tensor,
    encoder_hidden_states: torch.Tensor,  # Kept for interface compatibility
    encoder_hidden_states_image: torch.Tensor | None = None,
    timestep_seq_len: int | None = None,
):
    temb = self.time_embedder(timestep, timestep_seq_len)
    timestep_proj = self.time_modulation(temb)

    # Matrix-Game 2.0 does not use text embeddings, so we ignore encoder_hidden_states

    if encoder_hidden_states_image is not None:
        assert self.image_embedder is not None
        encoder_hidden_states_image = self.image_embedder(encoder_hidden_states_image)

    encoder_hidden_states = torch.zeros(
        (timestep.shape[0], 0, temb.shape[-1]),
        device=temb.device,
        dtype=temb.dtype,
    )

    return (
        temb,
        timestep_proj,
        encoder_hidden_states,
        encoder_hidden_states_image,
    )

`fastvideo.models.dits.matrixgame2.model.MatrixGame2CrossAttention`

MatrixGame2CrossAttention(dim: int, num_heads: int, window_size = (-1, -1), qk_norm = True, eps = 1e-06, parallel_attention = False, quant_config: QuantizationConfig | None = None, prefix: str = '')

Source

class MatrixGame2CrossAttention(WanSelfAttention):

    def forward(self, x, context, context_lens=None, crossattn_cache=None):
        r"""
        Args:
            x(Tensor): Shape [B, L1, C]
            context(Tensor): Shape [B, L2, C] - typically 257 image tokens
            context_lens(Tensor): Shape [B]
            crossattn_cache(dict): Optional cache for k/v during inference
        """
        b, n, d = x.size(0), self.num_heads, self.head_dim

        # compute query, key, value
        q = self.norm_q(self.to_q(x)[0]).view(b, -1, n, d)

        if crossattn_cache is not None:
            if not crossattn_cache["is_init"]:
                crossattn_cache["is_init"] = True
                k = self.norm_k(self.to_k(context)[0]).view(b, -1, n, d)
                v = self.to_v(context)[0].view(b, -1, n, d)
                crossattn_cache["k"] = k
                crossattn_cache["v"] = v
            else:
                k = crossattn_cache["k"]
                v = crossattn_cache["v"]
        else:
            k = self.norm_k(self.to_k(context)[0]).view(b, -1, n, d)
            v = self.to_v(context)[0].view(b, -1, n, d)

        # compute attention
        x = self.attn(q, k, v)

        # output
        x = x.flatten(2)
        x, _ = self.to_out(x)
        return x

`fastvideo.models.dits.matrixgame2.model.MatrixGame2CrossAttention.forward`

forward(x, context, context_lens = None, crossattn_cache = None)

Parameters

x: Tensor: Shape [B, L1, C]
context: Tensor: Shape [B, L2, C] - typically 257 image tokens
context_lens: Tensor: Shape [B]
crossattn_cache: dict: Optional cache for k/v during inference

Source

def forward(self, x, context, context_lens=None, crossattn_cache=None):
    r"""
    Args:
        x(Tensor): Shape [B, L1, C]
        context(Tensor): Shape [B, L2, C] - typically 257 image tokens
        context_lens(Tensor): Shape [B]
        crossattn_cache(dict): Optional cache for k/v during inference
    """
    b, n, d = x.size(0), self.num_heads, self.head_dim

    # compute query, key, value
    q = self.norm_q(self.to_q(x)[0]).view(b, -1, n, d)

    if crossattn_cache is not None:
        if not crossattn_cache["is_init"]:
            crossattn_cache["is_init"] = True
            k = self.norm_k(self.to_k(context)[0]).view(b, -1, n, d)
            v = self.to_v(context)[0].view(b, -1, n, d)
            crossattn_cache["k"] = k
            crossattn_cache["v"] = v
        else:
            k = crossattn_cache["k"]
            v = crossattn_cache["v"]
    else:
        k = self.norm_k(self.to_k(context)[0]).view(b, -1, n, d)
        v = self.to_v(context)[0].view(b, -1, n, d)

    # compute attention
    x = self.attn(q, k, v)

    # output
    x = x.flatten(2)
    x, _ = self.to_out(x)
    return x

`fastvideo.models.dits.matrixgame2.model.MatrixGame2TransformerBlock`

MatrixGame2TransformerBlock(dim: int, ffn_dim: int, num_heads: int, qk_norm: str = 'rms_norm_across_heads', cross_attn_norm: bool = False, eps: float = 1e-06, added_kv_proj_dim: int | None = None, supported_attention_backends: tuple[AttentionBackendEnum, ...] | None = None, prefix: str = '', action_config: dict | None = None)

Source

class MatrixGame2TransformerBlock(nn.Module):

    def __init__(
        self,
        dim: int,
        ffn_dim: int,
        num_heads: int,
        qk_norm: str = "rms_norm_across_heads",
        cross_attn_norm: bool = False,
        eps: float = 1e-6,
        added_kv_proj_dim: int | None = None,
        supported_attention_backends: tuple[AttentionBackendEnum, ...]
        | None = None,
        prefix: str = "",
        action_config: dict | None = None,
    ):
        super().__init__()
        action_config = action_config or {}

        # 1. Self-attention
        self.norm1 = FP32LayerNorm(dim, eps, elementwise_affine=False)
        self.to_q = ReplicatedLinear(dim, dim, bias=True)
        self.to_k = ReplicatedLinear(dim, dim, bias=True)
        self.to_v = ReplicatedLinear(dim, dim, bias=True)

        self.to_out = ReplicatedLinear(dim, dim, bias=True)
        self.attn1 = DistributedAttention(
            num_heads=num_heads,
            head_size=dim // num_heads,
            causal=False,
            supported_attention_backends=supported_attention_backends,
            prefix=f"{prefix}.attn1",
        )
        self.hidden_dim = dim
        self.num_attention_heads = num_heads
        dim_head = dim // num_heads
        if qk_norm == "rms_norm":
            self.norm_q = RMSNorm(dim_head, eps=eps)
            self.norm_k = RMSNorm(dim_head, eps=eps)
        elif qk_norm == "rms_norm_across_heads":
            # LTX applies qk norm across all heads
            self.norm_q = RMSNorm(dim, eps=eps)
            self.norm_k = RMSNorm(dim, eps=eps)
        else:
            print("QK Norm type not supported")
            raise Exception
        assert cross_attn_norm is True
        self.self_attn_residual_norm = ScaleResidualLayerNormScaleShift(
            dim,
            norm_type="layer",
            eps=eps,
            elementwise_affine=True,
            dtype=torch.float32,
            compute_dtype=torch.float32,
        )

        # 2. Cross-attention
        if added_kv_proj_dim is not None:
            # I2V
            self.attn2 = WanI2VCrossAttention(dim, num_heads, qk_norm=qk_norm, eps=eps)
        else:
            # T2V
            self.attn2 = WanT2VCrossAttention(dim, num_heads, qk_norm=qk_norm, eps=eps)
        self.cross_attn_residual_norm = ScaleResidualLayerNormScaleShift(
            dim,
            norm_type="layer",
            eps=eps,
            elementwise_affine=False,
            dtype=torch.float32,
            compute_dtype=torch.float32,
        )

        # 2.1. Action Module Integration
        self.use_action_module = len(action_config) > 0
        if self.use_action_module:
            self.action_model = ActionModule(**action_config)
        else:
            self.action_model = None

        # 3. Feed-forward
        self.ffn = MLP(dim, ffn_dim, act_type="gelu_pytorch_tanh")
        self.mlp_residual = ScaleResidual()

        self.scale_shift_table = nn.Parameter(torch.randn(1, 6, dim) / dim**0.5)

    def forward(
        self,
        hidden_states: torch.Tensor,
        encoder_hidden_states: torch.Tensor,
        temb: torch.Tensor,
        freqs_cis: tuple[torch.Tensor, torch.Tensor],
        # Action Module specific args
        grid_sizes: torch.Tensor,
        mouse_cond: torch.Tensor | None = None,
        keyboard_cond: torch.Tensor | None = None,
    ) -> torch.Tensor:
        if hidden_states.dim() == 4:
            hidden_states = hidden_states.squeeze(1)
        orig_dtype = hidden_states.dtype

        if temb.dim() == 4:
            # temb: batch_size, seq_len, 6, inner_dim (wan2.2 ti2v)
            (
                shift_msa,
                scale_msa,
                gate_msa,
                c_shift_msa,
                c_scale_msa,
                c_gate_msa,
            ) = (self.scale_shift_table.unsqueeze(0) + temb.float()).chunk(6, dim=2)
            # batch_size, seq_len, 1, inner_dim
            shift_msa = shift_msa.squeeze(2)
            scale_msa = scale_msa.squeeze(2)
            gate_msa = gate_msa.squeeze(2)
            c_shift_msa = c_shift_msa.squeeze(2)
            c_scale_msa = c_scale_msa.squeeze(2)
            c_gate_msa = c_gate_msa.squeeze(2)
        else:
            # temb: batch_size, 6, inner_dim (wan2.1/wan2.2 14B)
            e = self.scale_shift_table + temb.float()
            (
                shift_msa,
                scale_msa,
                gate_msa,
                c_shift_msa,
                c_scale_msa,
                c_gate_msa,
            ) = e.chunk(6, dim=1)
        assert shift_msa.dtype == torch.float32

        # 1. Self-attention
        norm_hidden_states = (self.norm1(hidden_states.float()) * (1 + scale_msa) + shift_msa).to(orig_dtype)
        query, _ = self.to_q(norm_hidden_states)
        key, _ = self.to_k(norm_hidden_states)
        value, _ = self.to_v(norm_hidden_states)

        if self.norm_q is not None:
            query = self.norm_q(query)
        if self.norm_k is not None:
            key = self.norm_k(key)

        query = query.squeeze(1).unflatten(2, (self.num_attention_heads, -1))
        key = key.squeeze(1).unflatten(2, (self.num_attention_heads, -1))
        value = value.squeeze(1).unflatten(2, (self.num_attention_heads, -1))

        # Apply rotary embeddings
        cos, sin = freqs_cis
        query, key = (
            _apply_rotary_emb(query, cos, sin, is_neox_style=False),
            _apply_rotary_emb(key, cos, sin, is_neox_style=False),
        )

        attn_output, _ = self.attn1(query, key, value)
        attn_output = attn_output.flatten(2)
        attn_output, _ = self.to_out(attn_output)
        attn_output = attn_output.squeeze(1)

        null_shift = null_scale = torch.tensor([0], device=hidden_states.device)
        norm_hidden_states, hidden_states = self.self_attn_residual_norm(hidden_states, attn_output, gate_msa,
                                                                         null_shift, null_scale)
        norm_hidden_states, hidden_states = (
            norm_hidden_states.to(orig_dtype),
            hidden_states.to(orig_dtype),
        )

        # 2. Cross-attention
        attn_output = self.attn2(norm_hidden_states, context=encoder_hidden_states, context_lens=None)
        norm_hidden_states, hidden_states = self.cross_attn_residual_norm(hidden_states, attn_output, 1, c_shift_msa,
                                                                          c_scale_msa)
        norm_hidden_states, hidden_states = (
            norm_hidden_states.to(orig_dtype),
            hidden_states.to(orig_dtype),
        )

        # ================= Action Module =================
        if self.action_model is not None:
            if mouse_cond is not None or keyboard_cond is not None:
                # grid_sizes is expected to be [F, H, W]
                # ActionModule implementation takes hidden_states directly
                hidden_states = self.action_model(
                    hidden_states,
                    int(grid_sizes[0]),
                    int(grid_sizes[1]),
                    int(grid_sizes[2]),
                    mouse_cond,
                    keyboard_cond,
                    num_frame_per_block=int(grid_sizes[0]),
                )
        # =================================================

        # 3. Feed-forward
        ff_output = self.ffn(norm_hidden_states)
        hidden_states = self.mlp_residual(hidden_states, ff_output, c_gate_msa)
        hidden_states = hidden_states.to(orig_dtype)

        return hidden_states

`fastvideo.models.dits.matrixgame2.model.MatrixGame2TransformerBlock.forward`

forward(hidden_states: torch.Tensor, encoder_hidden_states: torch.Tensor, temb: torch.Tensor, freqs_cis: tuple[torch.Tensor, torch.Tensor], grid_sizes: torch.Tensor, mouse_cond: torch.Tensor | None = None, keyboard_cond: torch.Tensor | None = None) -> torch.Tensor

Source

def forward(
    self,
    hidden_states: torch.Tensor,
    encoder_hidden_states: torch.Tensor,
    temb: torch.Tensor,
    freqs_cis: tuple[torch.Tensor, torch.Tensor],
    # Action Module specific args
    grid_sizes: torch.Tensor,
    mouse_cond: torch.Tensor | None = None,
    keyboard_cond: torch.Tensor | None = None,
) -> torch.Tensor:
    if hidden_states.dim() == 4:
        hidden_states = hidden_states.squeeze(1)
    orig_dtype = hidden_states.dtype

    if temb.dim() == 4:
        # temb: batch_size, seq_len, 6, inner_dim (wan2.2 ti2v)
        (
            shift_msa,
            scale_msa,
            gate_msa,
            c_shift_msa,
            c_scale_msa,
            c_gate_msa,
        ) = (self.scale_shift_table.unsqueeze(0) + temb.float()).chunk(6, dim=2)
        # batch_size, seq_len, 1, inner_dim
        shift_msa = shift_msa.squeeze(2)
        scale_msa = scale_msa.squeeze(2)
        gate_msa = gate_msa.squeeze(2)
        c_shift_msa = c_shift_msa.squeeze(2)
        c_scale_msa = c_scale_msa.squeeze(2)
        c_gate_msa = c_gate_msa.squeeze(2)
    else:
        # temb: batch_size, 6, inner_dim (wan2.1/wan2.2 14B)
        e = self.scale_shift_table + temb.float()
        (
            shift_msa,
            scale_msa,
            gate_msa,
            c_shift_msa,
            c_scale_msa,
            c_gate_msa,
        ) = e.chunk(6, dim=1)
    assert shift_msa.dtype == torch.float32

    # 1. Self-attention
    norm_hidden_states = (self.norm1(hidden_states.float()) * (1 + scale_msa) + shift_msa).to(orig_dtype)
    query, _ = self.to_q(norm_hidden_states)
    key, _ = self.to_k(norm_hidden_states)
    value, _ = self.to_v(norm_hidden_states)

    if self.norm_q is not None:
        query = self.norm_q(query)
    if self.norm_k is not None:
        key = self.norm_k(key)

    query = query.squeeze(1).unflatten(2, (self.num_attention_heads, -1))
    key = key.squeeze(1).unflatten(2, (self.num_attention_heads, -1))
    value = value.squeeze(1).unflatten(2, (self.num_attention_heads, -1))

    # Apply rotary embeddings
    cos, sin = freqs_cis
    query, key = (
        _apply_rotary_emb(query, cos, sin, is_neox_style=False),
        _apply_rotary_emb(key, cos, sin, is_neox_style=False),
    )

    attn_output, _ = self.attn1(query, key, value)
    attn_output = attn_output.flatten(2)
    attn_output, _ = self.to_out(attn_output)
    attn_output = attn_output.squeeze(1)

    null_shift = null_scale = torch.tensor([0], device=hidden_states.device)
    norm_hidden_states, hidden_states = self.self_attn_residual_norm(hidden_states, attn_output, gate_msa,
                                                                     null_shift, null_scale)
    norm_hidden_states, hidden_states = (
        norm_hidden_states.to(orig_dtype),
        hidden_states.to(orig_dtype),
    )

    # 2. Cross-attention
    attn_output = self.attn2(norm_hidden_states, context=encoder_hidden_states, context_lens=None)
    norm_hidden_states, hidden_states = self.cross_attn_residual_norm(hidden_states, attn_output, 1, c_shift_msa,
                                                                      c_scale_msa)
    norm_hidden_states, hidden_states = (
        norm_hidden_states.to(orig_dtype),
        hidden_states.to(orig_dtype),
    )

    # ================= Action Module =================
    if self.action_model is not None:
        if mouse_cond is not None or keyboard_cond is not None:
            # grid_sizes is expected to be [F, H, W]
            # ActionModule implementation takes hidden_states directly
            hidden_states = self.action_model(
                hidden_states,
                int(grid_sizes[0]),
                int(grid_sizes[1]),
                int(grid_sizes[2]),
                mouse_cond,
                keyboard_cond,
                num_frame_per_block=int(grid_sizes[0]),
            )
    # =================================================

    # 3. Feed-forward
    ff_output = self.ffn(norm_hidden_states)
    hidden_states = self.mlp_residual(hidden_states, ff_output, c_gate_msa)
    hidden_states = hidden_states.to(orig_dtype)

    return hidden_states

`fastvideo.models.dits.matrixgame2.model.MatrixGame2WanModel`

MatrixGame2WanModel(config: MatrixGame2WanVideoConfig, hf_config: dict[str, Any], **kwargs)

Source

class MatrixGame2WanModel(BaseDiT):
    # Marker for action input support (Matrix-Game)
    supports_action_input = True

    _fsdp_shard_conditions = _DEFAULT_MATRIXGAME2_CONFIG._fsdp_shard_conditions
    _compile_conditions = _DEFAULT_MATRIXGAME2_CONFIG._compile_conditions
    _supported_attention_backends = (_DEFAULT_MATRIXGAME2_CONFIG._supported_attention_backends)
    param_names_mapping = _DEFAULT_MATRIXGAME2_CONFIG.param_names_mapping
    reverse_param_names_mapping = (_DEFAULT_MATRIXGAME2_CONFIG.reverse_param_names_mapping)
    lora_param_names_mapping = (_DEFAULT_MATRIXGAME2_CONFIG.lora_param_names_mapping)

    def __init__(
        self,
        config: MatrixGame2WanVideoConfig,
        hf_config: dict[str, Any],
        **kwargs,
    ) -> None:
        super().__init__(config=config, hf_config=hf_config)

        inner_dim = config.num_attention_heads * config.attention_head_dim
        self.hidden_size = config.hidden_size
        self.num_attention_heads = config.num_attention_heads
        self.num_channels_latents = config.out_channels
        self.patch_size = config.patch_size

        # 1. Patch & position embedding
        self.patch_embedding = PatchEmbed(
            in_chans=config.in_channels,
            embed_dim=inner_dim,
            patch_size=config.patch_size,
            flatten=False,
        )

        # 2. Condition embeddings
        self.condition_embedder = MatrixGame2TimeImageEmbedding(
            dim=inner_dim,
            time_freq_dim=config.freq_dim,
            image_embed_dim=config.image_dim,
        )

        # 2.1. Get action config
        self.action_config = getattr(config, "action_config", {})

        # 3. Transformer blocks
        self.blocks = nn.ModuleList([
            MatrixGame2TransformerBlock(
                inner_dim,
                config.ffn_dim,
                config.num_attention_heads,
                config.qk_norm,
                config.cross_attn_norm,
                config.eps,
                config.added_kv_proj_dim,
                self._supported_attention_backends,
                prefix=f"{getattr(config, 'prefix', 'Wan')}.blocks.{i}",
                action_config=self.action_config,
            ) for i in range(config.num_layers)
        ])

        # 4. Output norm & projection
        self.norm_out = LayerNormScaleShift(
            inner_dim,
            norm_type="layer",
            eps=config.eps,
            elementwise_affine=False,
            dtype=torch.float32,
            compute_dtype=torch.float32,
        )
        self.proj_out = nn.Linear(inner_dim, config.out_channels * math.prod(config.patch_size))
        self.scale_shift_table = nn.Parameter(torch.randn(1, 2, inner_dim) / inner_dim**0.5)

        self.gradient_checkpointing = False

    def forward(
        self,
        hidden_states: torch.Tensor,
        encoder_hidden_states: torch.Tensor | list[torch.Tensor],
        timestep: torch.LongTensor,
        encoder_hidden_states_image: torch.Tensor
        | list[torch.Tensor]
        | None = None,
        # Action inputs
        mouse_cond: torch.Tensor | None = None,
        keyboard_cond: torch.Tensor | None = None,
        **kwargs,
    ) -> torch.Tensor:
        if encoder_hidden_states is not None and not isinstance(encoder_hidden_states, torch.Tensor):
            encoder_hidden_states = encoder_hidden_states[0]
        if (isinstance(encoder_hidden_states_image, list) and len(encoder_hidden_states_image) > 0):
            encoder_hidden_states_image = encoder_hidden_states_image[0]
        elif not isinstance(encoder_hidden_states_image, torch.Tensor):
            encoder_hidden_states_image = None

        batch_size, num_channels, num_frames, height, width = (hidden_states.shape)
        p_t, p_h, p_w = self.patch_size
        post_patch_num_frames = num_frames // p_t
        post_patch_height = height // p_h
        post_patch_width = width // p_w

        # Get rotary embeddings
        d = self.hidden_size // self.num_attention_heads
        rope_dim_list = [d - 4 * (d // 6), 2 * (d // 6), 2 * (d // 6)]
        freqs_cos, freqs_sin = get_rotary_pos_embed(
            (
                post_patch_num_frames * get_sp_world_size(),
                post_patch_height,
                post_patch_width,
            ),
            self.hidden_size,
            self.num_attention_heads,
            rope_dim_list,
            dtype=torch.float32 if current_platform.is_mps() else torch.float64,
            rope_theta=10000,
            do_sp_sharding=True,
        )
        freqs_cos = freqs_cos.to(hidden_states.device)
        freqs_sin = freqs_sin.to(hidden_states.device)
        freqs_cis = ((freqs_cos.float(), freqs_sin.float()) if freqs_cos is not None else None)

        hidden_states = self.patch_embedding(hidden_states)
        hidden_states = hidden_states.flatten(2).transpose(1, 2)

        if timestep.dim() == 2:
            timestep = timestep.flatten()

        (
            temb,
            timestep_proj,
            encoder_hidden_states,
            encoder_hidden_states_image,
        ) = self.condition_embedder(timestep, encoder_hidden_states, encoder_hidden_states_image)
        timestep_proj = timestep_proj.unflatten(1, (6, -1))

        if encoder_hidden_states is not None:
            if isinstance(encoder_hidden_states, list):
                encoder_hidden_states = encoder_hidden_states[0]
            elif encoder_hidden_states.ndim == 2:
                encoder_hidden_states = encoder_hidden_states.unsqueeze(0)
        else:
            # encoder_hidden_states is None (e.g. no text encoder)
            # Matrix-Game 2.0 uses image-action cross-attn.
            pass

        if encoder_hidden_states_image is not None:
            if encoder_hidden_states is not None:
                encoder_hidden_states = torch.concat([encoder_hidden_states_image, encoder_hidden_states], dim=1)
            else:
                encoder_hidden_states = encoder_hidden_states_image

        # This is [F, H, W] for the ActionModule
        grid_sizes = torch.tensor(
            [post_patch_num_frames, post_patch_height, post_patch_width],
            device=hidden_states.device,
        )

        # Blocks
        for block in self.blocks:
            if torch.is_grad_enabled() and self.gradient_checkpointing:
                hidden_states = self._gradient_checkpointing_func(
                    block,
                    hidden_states,
                    encoder_hidden_states,
                    timestep_proj,
                    freqs_cis,
                    grid_sizes=grid_sizes,
                    mouse_cond=mouse_cond,
                    keyboard_cond=keyboard_cond,
                )
            else:
                hidden_states = block(
                    hidden_states,
                    encoder_hidden_states,
                    timestep_proj,
                    freqs_cis,
                    grid_sizes=grid_sizes,
                    mouse_cond=mouse_cond,
                    keyboard_cond=keyboard_cond,
                )

        # Output
        shift, scale = (self.scale_shift_table + temb.unsqueeze(1)).chunk(2, dim=1)
        hidden_states = self.norm_out(hidden_states, shift, scale)
        hidden_states = self.proj_out(hidden_states)

        hidden_states = hidden_states.reshape(
            batch_size,
            post_patch_num_frames,
            post_patch_height,
            post_patch_width,
            p_t,
            p_h,
            p_w,
            -1,
        )
        hidden_states = hidden_states.permute(0, 7, 1, 4, 2, 5, 3, 6)
        output = hidden_states.flatten(6, 7).flatten(4, 5).flatten(2, 3)

        return output

`fastvideo.models.dits.matrixgame2.model.MatrixGame2WanModel.forward`

forward(hidden_states: torch.Tensor, encoder_hidden_states: torch.Tensor | list[torch.Tensor], timestep: torch.LongTensor, encoder_hidden_states_image: torch.Tensor | list[torch.Tensor] | None = None, mouse_cond: torch.Tensor | None = None, keyboard_cond: torch.Tensor | None = None, **kwargs) -> torch.Tensor

Source

def forward(
    self,
    hidden_states: torch.Tensor,
    encoder_hidden_states: torch.Tensor | list[torch.Tensor],
    timestep: torch.LongTensor,
    encoder_hidden_states_image: torch.Tensor
    | list[torch.Tensor]
    | None = None,
    # Action inputs
    mouse_cond: torch.Tensor | None = None,
    keyboard_cond: torch.Tensor | None = None,
    **kwargs,
) -> torch.Tensor:
    if encoder_hidden_states is not None and not isinstance(encoder_hidden_states, torch.Tensor):
        encoder_hidden_states = encoder_hidden_states[0]
    if (isinstance(encoder_hidden_states_image, list) and len(encoder_hidden_states_image) > 0):
        encoder_hidden_states_image = encoder_hidden_states_image[0]
    elif not isinstance(encoder_hidden_states_image, torch.Tensor):
        encoder_hidden_states_image = None

    batch_size, num_channels, num_frames, height, width = (hidden_states.shape)
    p_t, p_h, p_w = self.patch_size
    post_patch_num_frames = num_frames // p_t
    post_patch_height = height // p_h
    post_patch_width = width // p_w

    # Get rotary embeddings
    d = self.hidden_size // self.num_attention_heads
    rope_dim_list = [d - 4 * (d // 6), 2 * (d // 6), 2 * (d // 6)]
    freqs_cos, freqs_sin = get_rotary_pos_embed(
        (
            post_patch_num_frames * get_sp_world_size(),
            post_patch_height,
            post_patch_width,
        ),
        self.hidden_size,
        self.num_attention_heads,
        rope_dim_list,
        dtype=torch.float32 if current_platform.is_mps() else torch.float64,
        rope_theta=10000,
        do_sp_sharding=True,
    )
    freqs_cos = freqs_cos.to(hidden_states.device)
    freqs_sin = freqs_sin.to(hidden_states.device)
    freqs_cis = ((freqs_cos.float(), freqs_sin.float()) if freqs_cos is not None else None)

    hidden_states = self.patch_embedding(hidden_states)
    hidden_states = hidden_states.flatten(2).transpose(1, 2)

    if timestep.dim() == 2:
        timestep = timestep.flatten()

    (
        temb,
        timestep_proj,
        encoder_hidden_states,
        encoder_hidden_states_image,
    ) = self.condition_embedder(timestep, encoder_hidden_states, encoder_hidden_states_image)
    timestep_proj = timestep_proj.unflatten(1, (6, -1))

    if encoder_hidden_states is not None:
        if isinstance(encoder_hidden_states, list):
            encoder_hidden_states = encoder_hidden_states[0]
        elif encoder_hidden_states.ndim == 2:
            encoder_hidden_states = encoder_hidden_states.unsqueeze(0)
    else:
        # encoder_hidden_states is None (e.g. no text encoder)
        # Matrix-Game 2.0 uses image-action cross-attn.
        pass

    if encoder_hidden_states_image is not None:
        if encoder_hidden_states is not None:
            encoder_hidden_states = torch.concat([encoder_hidden_states_image, encoder_hidden_states], dim=1)
        else:
            encoder_hidden_states = encoder_hidden_states_image

    # This is [F, H, W] for the ActionModule
    grid_sizes = torch.tensor(
        [post_patch_num_frames, post_patch_height, post_patch_width],
        device=hidden_states.device,
    )

    # Blocks
    for block in self.blocks:
        if torch.is_grad_enabled() and self.gradient_checkpointing:
            hidden_states = self._gradient_checkpointing_func(
                block,
                hidden_states,
                encoder_hidden_states,
                timestep_proj,
                freqs_cis,
                grid_sizes=grid_sizes,
                mouse_cond=mouse_cond,
                keyboard_cond=keyboard_cond,
            )
        else:
            hidden_states = block(
                hidden_states,
                encoder_hidden_states,
                timestep_proj,
                freqs_cis,
                grid_sizes=grid_sizes,
                mouse_cond=mouse_cond,
                keyboard_cond=keyboard_cond,
            )

    # Output
    shift, scale = (self.scale_shift_table + temb.unsqueeze(1)).chunk(2, dim=1)
    hidden_states = self.norm_out(hidden_states, shift, scale)
    hidden_states = self.proj_out(hidden_states)

    hidden_states = hidden_states.reshape(
        batch_size,
        post_patch_num_frames,
        post_patch_height,
        post_patch_width,
        p_t,
        p_h,
        p_w,
        -1,
    )
    hidden_states = hidden_states.permute(0, 7, 1, 4, 2, 5, 3, 6)
    output = hidden_states.flatten(6, 7).flatten(4, 5).flatten(2, 3)

    return output