convert

type ByteOrder interface {
	binary.ByteOrder
	binary.AppendByteOrder
}

type GemmaModel struct {
	ModelData
}

type LlamaModel struct {
	ModelData
}

type MistralModel struct {
	ModelData
}

type MixtralModel struct {
	ModelData
}

type ModelArch interface {
	GetTensors() error
	LoadVocab() error
	WriteGGUF(io.WriteSeeker) error
}

type ModelData struct {
	Path    string
	Name    string
	Params  *Params
	Vocab   *Vocab
	Tensors []llm.Tensor
	Format  ModelFormat
}

type ModelFormat interface {
	GetLayerName(string) (string, error)
	GetTensors(string, *Params) ([]llm.Tensor, error)
	GetParams(string) (*Params, error)
	GetModelArch(string, string, *Params) (ModelArch, error)
}

type Params struct {
	Architectures     []string `json:"architectures"`
	VocabSize         int      `json:"vocab_size"`
	HiddenSize        int      `json:"hidden_size"`       // n_embd
	HiddenLayers      int      `json:"num_hidden_layers"` // n_layer
	ContextSize       int      `json:"max_position_embeddings"`
	IntermediateSize  int      `json:"intermediate_size"`
	AttentionHeads    int      `json:"num_attention_heads"` // n_head
	KeyValHeads       int      `json:"num_key_value_heads"`
	NormEPS           float64  `json:"rms_norm_eps"`
	BoSTokenID        int      `json:"bos_token_id"`
	EoSTokenID        int      `json:"eos_token_id"`
	HeadDimension     int      `json:"head_dim"`
	PaddingTokenID    int      `json:"pad_token_id"`
	RopeFrequencyBase float64  `json:"rope_theta"`

	Experts     int `json:"num_local_experts"`
	ExpertsUsed int `json:"num_experts_per_tok"`

	ByteOrder
}

type SafetensorFormat struct{}

type TorchFormat struct{}

type Vocab struct {
	Tokens []string
	Scores []float32
	Types  []int32
}