pmetal-merge

Model merging toolkit for Apple Silicon.

Overview

This crate provides utilities for merging multiple fine-tuned models into a single model. It supports various merging strategies and is optimized for memory-efficient processing of large models.

Merge Methods

Features

• Lazy Loading: Stream weights without loading full models
• Memory Efficient: Process layer-by-layer for large models
• Multiple Formats: SafeTensors, PyTorch, GGUF support
• GPU-Accelerated Merging: Metal-based merge operations for large models
• FP8-Aware Merging: Merge with FP8 quantization for memory efficiency
• Async Merge Pipeline: Double-buffered streaming merge for large models
• LoRA Merge: Fuse LoRA adapters into base weights (standard and accurate modes)
• Configurable: Fine-grained control over merge parameters

Usage

Linear Merge

use pmetal_merge::{MergeConfig, LinearMerge, run_merge};

let config = MergeConfig {
    method: MergeMethod::Linear,
    models: vec![
        ModelWeight { path: "model_a", weight: 0.7 },
        ModelWeight { path: "model_b", weight: 0.3 },
    ],
    output: "merged_model",
};

run_merge(&config)?;

SLERP Merge

use pmetal_merge::{MergeConfig, MergeMethod};

let config = MergeConfig {
    method: MergeMethod::Slerp { t: 0.5 },
    models: vec![
        ModelWeight { path: "model_a", weight: 1.0 },
        ModelWeight { path: "model_b", weight: 1.0 },
    ],
    output: "merged_model",
};

TIES Merge

use pmetal_merge::{MergeConfig, MergeMethod};

let config = MergeConfig {
    method: MergeMethod::Ties {
        density: 0.5,      // Keep top 50% of weights
        majority_sign: true,
    },
    models: vec![
        ModelWeight { path: "task_a", weight: 1.0 },
        ModelWeight { path: "task_b", weight: 1.0 },
        ModelWeight { path: "task_c", weight: 1.0 },
    ],
    base_model: Some("base_model"),
    output: "merged_model",
};

Merge Methods Explained

Linear

Simple weighted average: merged = w1*m1 + w2*m2 + ...

SLERP

Spherical linear interpolation for smooth blending between two models. Parameter t controls interpolation (0.0 = model A, 1.0 = model B).

TIES

Task Arithmetic with Interference Elimination:

1. Compute task vectors (fine-tuned - base)
2. Trim low-magnitude weights
3. Resolve sign conflicts by majority vote
4. Merge remaining weights

DARE

Drop And REscale:

1. Randomly drop weights with probability p
2. Rescale remaining weights by 1/(1-p)
3. Reduces interference between models. Available in TIES and Linear variants.

Task Arithmetic

Direct task vector addition: merged = base + w1*(m1-base) + w2*(m2-base) + ...

DELLA

Adaptive magnitude-based pruning. Prunes weights based on their magnitude relative to the base model, preserving important changes.

Model Stock

Geometric interpolation using task vector similarity. Computes merge weights based on geometric properties of the fine-tuning directions.

Passthrough

Layer passthrough composition — select layers from different models to build a "Frankenstein" model.

Modules

License

MIT OR Apache-2.0