model-merging

模型合并：组合预训练模型

何时使用本技能

当你需要以下情况时，使用模型合并：

• 组合能力：将多个微调模型的能力结合起来，而无需重新训练
• 创建专用模型：通过混合特定领域的专业知识（数学 + 编程 + 对话）来创建专用模型
• 提升性能：超越单一模型（在基准测试中通常提升 5-10%）
• 降低训练成本：无需 GPU，合并在 CPU 上运行
• 快速实验：在几分钟内创建新的模型变体，而不是几天
• 保留多项技能：合并而不会导致灾难性遗忘

成功案例：Marcoro14-7B-slerp（在 2024 年 2 月的 Open LLM Leaderboard 上排名第一），许多 HuggingFace 顶级模型都使用合并

工具：mergekit（Arcee AI）、LazyMergekit、Model Soup

安装

# Install mergekit
git clone https://github.com/arcee-ai/mergekit.git
cd mergekit
pip install -e .

# Or via pip
pip install mergekit

# Optional: Transformer library
pip install transformers torch

快速开始

简单线性合并

# config.yml - Merge two models with equal weights
merge_method: linear
models:
  - model: mistralai/Mistral-7B-v0.1
    parameters:
      weight: 0.5
  - model: teknium/OpenHermes-2.5-Mistral-7B
    parameters:
      weight: 0.5
dtype: bfloat16

# Run merge
mergekit-yaml config.yml ./merged-model --cuda

# Use merged model
python -m transformers.models.auto --model_name_or_path ./merged-model

SLERP 合并（最适合 2 个模型）

# config.yml - Spherical interpolation
merge_method: slerp
slices:
  - sources:
      - model: mistralai/Mistral-7B-v0.1
        layer_range: [0, 32]
      - model: teknium/OpenHermes-2.5-Mistral-7B
        layer_range: [0, 32]
parameters:
  t: 0.5  # Interpolation factor (0=model1, 1=model2)
dtype: bfloat16

核心概念

1. 合并方法

线性（Model Soup）

• 参数的简单加权平均
• 快速，对相似模型效果良好
• 可以合并 2 个及以上模型

merged_weights = w1 * model1_weights + w2 * model2_weights + w3 * model3_weights
# where w1 + w2 + w3 = 1

SLERP（球面线性插值）

• 在权重空间中沿球面进行插值
• 保留权重向量的模长
• 最适合合并 2 个模型
• 比线性更平滑

# SLERP formula
merged = (sin((1-t)*θ) / sin(θ)) * model1 + (sin(t*θ) / sin(θ)) * model2
# where θ = arccos(dot(model1, model2))
# t ∈ [0, 1]

任务算术

• 提取"任务向量"（微调模型 - 基础模型）
• 组合任务向量，添加到基础模型
• 适合合并多个专用模型

# Task vector
task_vector = finetuned_model - base_model

# Merge multiple task vectors
merged = base_model + α₁*task_vector₁ + α₂*task_vector₂

TIES-Merging

• 任务算术 + 稀疏化
• 解决参数中的符号冲突
• 最适合合并多个任务特定模型

DARE（Drop And REscale）

• 随机丢弃微调参数
• 重新缩放剩余参数
• 减少冗余，保持性能

2. 配置结构

# Basic structure
merge_method: <method>  # linear, slerp, ties, dare_ties, task_arithmetic
base_model: <path>      # Optional: base model for task arithmetic

models:
  - model: <path/to/model1>
    parameters:
      weight: <float>   # Merge weight
      density: <float>  # For TIES/DARE

  - model: <path/to/model2>
    parameters:
      weight: <float>

parameters:
  # Method-specific parameters

dtype: <dtype>  # bfloat16, float16, float32

# Optional
slices:  # Layer-wise merging
tokenizer:  # Tokenizer configuration

合并方法指南

线性合并

最适合：简单的模型组合，等权重

merge_method: linear
models:
  - model: WizardLM/WizardMath-7B-V1.1
    parameters:
      weight: 0.4
  - model: teknium/OpenHermes-2.5-Mistral-7B
    parameters:
      weight: 0.3
  - model: NousResearch/Nous-Hermes-2-Mistral-7B-DPO
    parameters:
      weight: 0.3
dtype: bfloat16

SLERP 合并

最适合：两个模型，平滑插值

merge_method: slerp
slices:
  - sources:
      - model: mistralai/Mistral-7B-v0.1
        layer_range: [0, 32]
      - model: teknium/OpenHermes-2.5-Mistral-7B
        layer_range: [0, 32]
parameters:
  t: 0.5  # 0.0 = first model, 1.0 = second model
dtype: bfloat16

按层 SLERP：

merge_method: slerp
slices:
  - sources:
      - model: model_a
        layer_range: [0, 32]
      - model: model_b
        layer_range: [0, 32]
parameters:
  t:
    - filter: self_attn    # Attention layers
      value: 0.3
    - filter: mlp          # MLP layers
      value: 0.7
    - value: 0.5           # Default for other layers
dtype: bfloat16

任务算术

最适合：组合专业技能

merge_method: task_arithmetic
base_model: mistralai/Mistral-7B-v0.1
models:
  - model: WizardLM/WizardMath-7B-V1.1  # Math
    parameters:
      weight: 0.5
  - model: teknium/OpenHermes-2.5-Mistral-7B  # Chat
    parameters:
      weight: 0.3
  - model: ajibawa-2023/Code-Mistral-7B  # Code
    parameters:
      weight: 0.2
dtype: bfloat16

TIES-Merging

最适合：多个模型，解决冲突

merge_method: ties
base_model: mistralai/Mistral-7B-v0.1
models:
  - model: WizardLM/WizardMath-7B-V1.1
    parameters:
      density: 0.5  # Keep top 50% of parameters
      weight: 1.0
  - model: teknium/OpenHermes-2.5-Mistral-7B
    parameters:
      density: 0.5
      weight: 1.0
  - model: NousResearch/Nous-Hermes-2-Mistral-7B-DPO
    parameters:
      density: 0.5
      weight: 1.0
parameters:
  normalize: true
dtype: bfloat16

DARE 合并

最适合：减少冗余

merge_method: dare_ties
base_model: mistralai/Mistral-7B-v0.1
models:
  - model: WizardLM/WizardMath-7B-V1.1
    parameters:
      density: 0.5    # Drop 50% of deltas
      weight: 0.6
  - model: teknium/OpenHermes-2.5-Mistral-7B
    parameters:
      density: 0.5
      weight: 0.4
parameters:
  int8_mask: true  # Use int8 for masks (saves memory)
dtype: bfloat16

高级模式

逐层合并

# Different models for different layers
merge_method: passthrough
slices:
  - sources:
      - model: mistralai/Mistral-7B-v0.1
        layer_range: [0, 16]   # First half
  - sources:
      - model: teknium/OpenHermes-2.5-Mistral-7B
        layer_range: [16, 32]  # Second half
dtype: bfloat16

从合并模型创建 MoE

# Create Mixture of Experts
merge_method: moe
base_model: mistralai/Mistral-7B-v0.1
experts:
  - source_model: WizardLM/WizardMath-7B-V1.1
    positive_prompts:
      - "math"
      - "calculate"
  - source_model: teknium/OpenHermes-2.5-Mistral-7B
    positive_prompts:
      - "chat"
      - "conversation"
  - source_model: ajibawa-2023/Code-Mistral-7B
    positive_prompts:
      - "code"
      - "python"
dtype: bfloat16

分词器合并

merge_method: linear
models:
  - model: mistralai/Mistral-7B-v0.1
  - model: custom/specialized-model

tokenizer:
  source: "union"  # Combine vocabularies from both models
  tokens:
    <|special_token|>:
      source: "custom/specialized-model"

最佳实践

1. 模型兼容性

# ✅ Good: Same architecture
models = [
    "mistralai/Mistral-7B-v0.1",
    "teknium/OpenHermes-2.5-Mistral-7B",  # Both Mistral 7B
]

# ❌ Bad: Different architectures
models = [
    "meta-llama/Llama-2-7b-hf",  # Llama
    "mistralai/Mistral-7B-v0.1",  # Mistral (incompatible!)
]

2. 权重选择

# ✅ Good: Weights sum to 1.0
models:
  - model: model_a
    parameters:
      weight: 0.6
  - model: model_b
    parameters:
      weight: 0.4  # 0.6 + 0.4 = 1.0

# ⚠️  Acceptable: Weights don't sum to 1 (for task arithmetic)
models:
  - model: model_a
    parameters:
      weight: 0.8
  - model: model_b
    parameters:
      weight: 0.8  # May boost performance

3. 方法选择

# Choose merge method based on use case:

# 2 models, smooth blend → SLERP
merge_method = "slerp"

# 3+ models, simple average → Linear
merge_method = "linear"

# Multiple task-specific models → Task Arithmetic or TIES
merge_method = "ties"

# Want to reduce redundancy → DARE
merge_method = "dare_ties"

4. 密度调优（TIES/DARE）

# Start conservative (keep more parameters)
parameters:
  density: 0.8  # Keep 80%

# If performance good, increase sparsity
parameters:
  density: 0.5  # Keep 50%

# If performance degrades, reduce sparsity
parameters:
  density: 0.9  # Keep 90%

5. 按层合并

# Preserve base model's beginning and end
merge_method: passthrough
slices:
  - sources:
      - model: base_model
        layer_range: [0, 2]     # Keep first layers
  - sources:
      - model: merged_middle    # Merge middle layers
        layer_range: [2, 30]
  - sources:
      - model: base_model
        layer_range: [30, 32]   # Keep last layers

评估与测试

基准测试合并模型

from transformers import AutoModelForCausalLM, AutoTokenizer

# Load merged model
model = AutoModelForCausalLM.from_pretrained("./merged-model")
tokenizer = AutoTokenizer.from_pretrained("./merged-model")

# Test on various tasks
test_prompts = {
    "math": "Calculate: 25 * 17 =",
    "code": "Write a Python function to reverse a string:",
    "chat": "What is the capital of France?",
}

for task, prompt in test_prompts.items():
    inputs = tokenizer(prompt, return_tensors="pt")
    outputs = model.generate(**inputs, max_length=100)
    print(f"{task}: {tokenizer.decode(outputs[0])}")

常见基准测试

• Open LLM Leaderboard：通用能力
• MT-Bench：多轮对话
• MMLU：多任务准确率
• HumanEval：代码生成
• GSM8K：数学推理

生产部署

保存并上传

from transformers import AutoModelForCausalLM, AutoTokenizer

# Load merged model
model = AutoModelForCausalLM.from_pretrained("./merged-model")
tokenizer = AutoTokenizer.from_pretrained("./merged-model")

# Upload to HuggingFace Hub
model.push_to_hub("username/my-merged-model")
tokenizer.push_to_hub("username/my-merged-model")

量化合并模型

# Quantize with GGUF
python convert.py ./merged-model --outtype f16 --outfile merged-model.gguf

# Quantize with GPTQ
python quantize_gptq.py ./merged-model --bits 4 --group_size 128

常见陷阱

❌ 陷阱 1：合并不兼容的模型

# Wrong: Different architectures
models:
  - model: meta-llama/Llama-2-7b  # Llama architecture
  - model: mistralai/Mistral-7B   # Mistral architecture

修复：只合并具有相同架构的模型

❌ 陷阱 2：过度加权某个模型

# Suboptimal: One model dominates
models:
  - model: model_a
    parameters:
      weight: 0.95  # Too high
  - model: model_b
    parameters:
      weight: 0.05  # Too low

修复：使用更均衡的权重（0.3-0.7 范围）

❌ 陷阱 3：不进行评估

# Wrong: Merge and deploy without testing
mergekit-yaml config.yml ./merged-model
# Deploy immediately (risky!)

修复：部署前务必进行基准测试

资源

• mergekit GitHub：https://github.com/arcee-ai/mergekit
• HuggingFace 教程：https://huggingface.co/blog/mlabonne/merge-models
• LazyMergekit：自动化合并 notebook
• TIES 论文：https://arxiv.org/abs/2306.01708
• DARE 论文：https://arxiv.org/abs/2311.03099

另请参阅

• references/methods.md - 深入探讨合并算法
• references/examples.md - 真实世界的合并配置
• references/evaluation.md - 基准测试和测试策略