llama

Documentation

Index

• func New(c fs.Config) (model.Model, error)
• type Layer
  • func (l *Layer) Forward(ctx ml.Context, hiddenState, positions, outputs ml.Tensor, cache kvcache.Cache, ...) ml.Tensor
• type MLP
  • func (mlp *MLP) Forward(ctx ml.Context, hiddenState ml.Tensor, opts *Options) ml.Tensor
• type Model
  • func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error)
  • func (m *Model) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error)
• type Options
• type SelfAttention
  • func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, positions ml.Tensor, cache kvcache.Cache, ...) ml.Tensor

Functions

func New

func New(c fs.Config) (model.Model, error)

Types

type Layer

type Layer struct {
	AttentionNorm *nn.RMSNorm `gguf:"attn_norm"`
	SelfAttention *SelfAttention
	MLPNorm       *nn.RMSNorm `gguf:"ffn_norm"`
	MLP           *MLP
}

func (*Layer) Forward

func (l *Layer) Forward(ctx ml.Context, hiddenState, positions, outputs ml.Tensor, cache kvcache.Cache, opts *Options) ml.Tensor

type MLP

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

func (*MLP) Forward

func (mlp *MLP) Forward(ctx ml.Context, hiddenState ml.Tensor, opts *Options) ml.Tensor

type Model

type Model struct {
	model.Base
	model.BytePairEncoding

	TokenEmbedding *nn.Embedding `gguf:"token_embd"`
	Layers         []Layer       `gguf:"blk"`
	OutputNorm     *nn.RMSNorm   `gguf:"output_norm"`
	Output         *nn.Linear    `gguf:"output,alt:token_embd"`

	*Options
}

func (*Model) Forward

func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error)

func (*Model) Shift

func (m *Model) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error)

type Options

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

type SelfAttention

type SelfAttention 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_freqs.weight"`
}

func (*SelfAttention) Forward

func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, positions ml.Tensor, cache kvcache.Cache, opts *Options) ml.Tensor