makarna/pkg/model/tensors.go

package model

// Standard tensor names for various architectures
// These constants helps in normalizing and identifying tensor roles across different models (Llama, Qwen, etc.)

const (
	// Embeddings & Head
	TensorTokenEmbeddings = "embed_tokens"
	TensorOutputHead      = "lm_head"
	TensorNorm            = "norm"
	
	// Attention
	TensorAttnQ    = "self_attn.q_proj"
	TensorAttnK    = "self_attn.k_proj"
	TensorAttnV    = "self_attn.v_proj"
	TensorAttnO    = "self_attn.o_proj"
	TensorAttnNorm = "input_layernorm" // pre-attention norm
	
	// MLP / FFN
	TensorMlpGate = "mlp.gate_proj"
	TensorMlpUp   = "mlp.up_proj"
	TensorMlpDown = "mlp.down_proj"
	TensorMlpNorm = "post_attention_layernorm" // pre-mlp norm
	
	// Biases (Common suffixes)
	SuffixWeight = ".weight"
	SuffixBias   = ".bias"
)

// ImportantPattern defines a mapping between a conceptual role and a regex pattern
type ImportantPattern struct {
	Role    string
	Pattern string
}

// DefaultCriticalPatterns returns a list of patterns typically considered critical (sensitive)
func DefaultCriticalPatterns() []string {
	return []string{
		"embed_tokens",
		"lm_head",
	}
}