llama4

type ImageProcessor struct {
	// contains filtered or unexported fields
}

type Model struct {
	model.Base
	model.BytePairEncoding
	ImageProcessor

	*VisionModel `gguf:"v"`
	*Projector   `gguf:"mm"`
	*TextModel
}

type PatchEmbedding struct {
	*nn.Linear
}

type Projector struct {
	Linear1 *nn.Linear `gguf:"linear_1"`
}

type TextAttention struct {
	Query       *nn.Linear `gguf:"attn_q"`
	Key         *nn.Linear `gguf:"attn_k"`
	Value       *nn.Linear `gguf:"attn_v"`
	Output      *nn.Linear `gguf:"attn_output"`
	RopeFactors ml.Tensor  `gguf:"rope_factors"`
}

type TextExperts struct {
	Gate *nn.LinearBatch `gguf:"ffn_gate_exps"`
	Up   *nn.LinearBatch `gguf:"ffn_up_exps"`
	Down *nn.LinearBatch `gguf:"ffn_down_exps"`
}

type TextFeedForward interface {
	Forward(ctx ml.Context, hiddenStates ml.Tensor, opts *TextOptions) ml.Tensor
}

type TextLayer struct {
	AttentionNorm *nn.LayerNorm `gguf:"attn_norm"`
	Attention     *TextAttention

	FFNNorm     *nn.LayerNorm `gguf:"ffn_norm"`
	FeedForward TextFeedForward
}

type TextMLP struct {
	Gate *nn.Linear `gguf:"ffn_gate"`
	Up   *nn.Linear `gguf:"ffn_up"`
	Down *nn.Linear `gguf:"ffn_down"`
}

type TextMOE struct {
	Router       *nn.Linear `gguf:"ffn_gate_inp"`
	Experts      *TextExperts
	SharedExpert *TextMLP `gguf:",suf:_shexp"`
}

type TextModel struct {
	Layers []TextLayer `gguf:"blk"`

	TokenEmbedding *nn.Embedding `gguf:"token_embd"`
	OutputNorm     *nn.LayerNorm `gguf:"output_norm"`
	Output         *nn.Linear    `gguf:"output,alt:token_embd"`

	*TextOptions
}

type TextOptions struct {
	// contains filtered or unexported fields
}

type VisionAdapter struct {
	FC1 *nn.Linear `gguf:"mlp.fc1"`
	FC2 *nn.Linear `gguf:"mlp.fc2"`
}

type VisionAttention struct {
	Query  *nn.Linear `gguf:"attn_q"`
	Key    *nn.Linear `gguf:"attn_k"`
	Value  *nn.Linear `gguf:"attn_v"`
	Output *nn.Linear `gguf:"attn_output"`
}

type VisionLayer struct {
	InputLayerNorm *nn.LayerNorm `gguf:"attn_norm"`
	*VisionAttention

	PostAttentionNorm *nn.LayerNorm `gguf:"ffn_norm"`
	*VisionMLP        `gguf:"mlp"`
}

type VisionMLP struct {
	FC1 *nn.Linear `gguf:"fc1"`
	FC2 *nn.Linear `gguf:"fc2"`
}

type VisionModel struct {
	Layers []VisionLayer `gguf:"blk"`

	*PatchEmbedding     `gguf:"patch_embedding"`
	ClassEmbedding      ml.Tensor `gguf:"class_embedding"`
	PositionalEmbedding ml.Tensor `gguf:"positional_embedding_vlm"`

	LayerNormPre  *nn.LayerNorm `gguf:"layernorm_pre"`
	LayerNormPost *nn.LayerNorm `gguf:"layernorm_post"`

	*VisionAdapter `gguf:"vision_adapter"`

	*VisionOptions
}

type VisionOptions struct {
	// contains filtered or unexported fields
}