Implementation:Ollama Ollama Convert Lfm2

Knowledge Sources	Ollama
Domains	Model Conversion, GGUF Format
Last Updated	2025-02-15 00:00 GMT

Overview

Implements the GGUF model converter for the LFM2 hybrid architecture, handling per-layer KV head count arrays based on mixed attention and short convolution layer types, and squeezing convolution weight dimensions.

Description

The lfm2Model struct implements ModelConverter for LFM2 models that combine full attention layers with short convolution layers. The KV method builds a per-layer KV head count array where attention layers have the configured number of KV heads and short convolution layers have zero heads. It also emits the short convolution L-cache size. The Tensors method squeezes 3D convolution weight tensors from shape [D, 1, K] to [D, K] by removing the singleton dimension. Tensor name replacements map the conv.conv, conv.in_proj, and conv.out_proj paths to shortconv.* GGUF names.

Usage

Invoked automatically when the model's architecture matches LFM2ForCausalLM.

Code Reference

Source Location

• Repository: Ollama
• File: convert/convert_lfm2.go
• Lines: 1-100

Signature

type lfm2Model struct {
    ModelParameters
    HiddenSize      uint32   `json:"hidden_size"`
    NumHiddenLayers uint32   `json:"num_hidden_layers"`
    ConvLCache      uint32   `json:"conv_L_cache"`
    LayerTypes      []string `json:"layer_types"`
    TieEmbedding    bool     `json:"tie_embedding"`
}

func (p *lfm2Model) KV(t *Tokenizer) KV
func (p *lfm2Model) Tensors(ts []Tensor) []*ggml.Tensor
func (p *lfm2Model) Replacements() []string

Import

import "github.com/ollama/ollama/convert"

I/O Contract

Inputs

Name	Type	Required	Description
t	*Tokenizer	Yes	Tokenizer data for GGUF metadata
ts	[]Tensor	Yes	Source tensors including conv weights to squeeze

Outputs

Name	Type	Description
KV	KV	GGUF metadata with lfm2.* keys including per-layer KV head counts and shortconv params
[]*ggml.Tensor	slice	Converted tensors with squeezed convolution weights

Usage Examples

// Converter registered for LFM2 architecture
// Per-layer KV head counts: attention layers get NumKeyValueHeads, conv layers get 0
// Conv weights [D, 1, K] are squeezed to [D, K]

Related Pages

• Principle:Ollama_Ollama_GGUF_Model_Conversion_Lfm2