Gemma Multimodal Fine-Tuner
Gemma多模态微调工具
Skill by
ara.so — Daily 2026 Skills collection.
Fine-tune Gemma 4 and Gemma 3n models on text, images, and audio data entirely on Apple Silicon (MPS), with support for streaming large datasets from GCS/BigQuery without filling local storage.
由
ara.so提供的Skill — 2026年度每日Skill合集。
你可以完全在Apple Silicon（MPS）设备上基于文本、图像、音频数据微调Gemma 4和Gemma 3n模型，支持从GCS/BigQuery流式加载大规模数据集，不会占用本地存储。
- Text LoRA: instruction-tuning or completion fine-tuning from local CSV
- Image + Text LoRA: captioning and VQA from local CSV
- Audio + Text LoRA: the only Apple-Silicon-native path for this modality
- Cloud streaming: train on terabytes from GCS/BigQuery without local copy
- MPS-native: no NVIDIA GPU required — runs on MacBook Pro/Air/Mac Studio
- Text LoRA：基于本地CSV文件进行指令微调或补全微调
- 图像+文本LoRA：基于本地CSV文件实现图像描述生成和视觉问答（VQA）
- 音频+文本LoRA：目前唯一原生支持Apple Silicon的该模态微调方案
- 云端流式加载：无需本地下载，可直接基于GCS/BigQuery上的TB级数据集训练
- 原生支持MPS：无需NVIDIA GPU，可在MacBook Pro/Air/Mac Studio上运行
- macOS 12.3+ with Apple Silicon (arm64)
- Python 3.10+ (native arm64, not Rosetta)
- Hugging Face account with Gemma access
- 搭载Apple Silicon（arm64架构）的设备，macOS 12.3及以上版本
- Python 3.10及以上版本（原生arm64版本，非Rosetta转译版本）
- 已开通Gemma访问权限的Hugging Face账号
Install Python 3.12 if needed
Install Python 3.12 if needed
python3.12 -m venv .venv
source .venv/bin/activate
python3.12 -m venv .venv
source .venv/bin/activate
Verify arm64 (must show arm64, not x86_64)
Verify arm64 (must show arm64, not x86_64)
python -c "import platform; print(platform.machine())"
python -c "import platform; print(platform.machine())"
Install PyTorch
Install PyTorch
pip install torch torchaudio
pip install torch torchaudio
Clone and install
Clone and install
For Gemma 4 support (separate venv recommended)
For Gemma 4 support (separate venv recommended)
pip install -r requirements/requirements-gemma4.txt
pip install -r requirements/requirements-gemma4.txt
Authenticate with Hugging Face
Hugging Face身份认证
bash
huggingface-cli login
bash
huggingface-cli login
Or set environment variable:
Or set environment variable:
export HF_TOKEN=your_token_here
export HF_TOKEN=your_token_here
Check system is ready
Check system is ready
gemma-macos-tuner system-check
gemma-macos-tuner system-check
Guided setup wizard (recommended for first run)
Guided setup wizard (recommended for first run)
Prepare dataset
Prepare dataset
gemma-macos-tuner prepare <dataset-profile>
gemma-macos-tuner prepare <dataset-profile>
Fine-tune a model
Fine-tune a model
gemma-macos-tuner finetune <profile> --json-logging
gemma-macos-tuner finetune <profile> --json-logging
Evaluate a run
Evaluate a run
gemma-macos-tuner evaluate <profile-or-run>
gemma-macos-tuner evaluate <profile-or-run>
Export merged HF/SafeTensors (merges LoRA when adapter_config.json present)
Export merged HF/SafeTensors (merges LoRA when adapter_config.json present)
gemma-macos-tuner export <run-dir-or-profile>
gemma-macos-tuner export <run-dir-or-profile>
Blacklist bad samples from errors
Blacklist bad samples from errors
gemma-macos-tuner blacklist <profile>
gemma-macos-tuner blacklist <profile>
List training runs
List training runs
gemma-macos-tuner runs list
gemma-macos-tuner runs list
The config is hierarchical INI: defaults → groups → models → datasets → profiles.
ini
[defaults]
output_dir = output
batch_size = 2
gradient_accumulation_steps = 8
learning_rate = 2e-4
num_train_epochs = 3
[model:gemma-3n-e2b-it]
group = gemma
base_model = google/gemma-3n-E2B-it
[model:gemma-4-e2b-it]
group = gemma
base_model = google/gemma-4-E2B-it
[dataset:my-audio-dataset]
data_dir = data/datasets/my-audio-dataset
audio_column = audio_path
text_column = transcript
[profile:my-audio-profile]
model = gemma-3n-e2b-it
dataset = my-audio-dataset
modality = audio
lora_r = 16
lora_alpha = 32
lora_dropout = 0.05
max_seq_length = 512
Use
env var to point to config outside repo root:
bash
export GEMMA_TUNER_CONFIG=/path/to/my/config.ini
配置采用层级化INI结构：默认配置 → 分组配置 → 模型配置 → 数据集配置 → 运行配置。
ini
[defaults]
output_dir = output
batch_size = 2
gradient_accumulation_steps = 8
learning_rate = 2e-4
num_train_epochs = 3
[model:gemma-3n-e2b-it]
group = gemma
base_model = google/gemma-3n-E2B-it
[model:gemma-4-e2b-it]
group = gemma
base_model = google/gemma-4-E2B-it
[dataset:my-audio-dataset]
data_dir = data/datasets/my-audio-dataset
audio_column = audio_path
text_column = transcript
[profile:my-audio-profile]
model = gemma-3n-e2b-it
dataset = my-audio-dataset
modality = audio
lora_r = 16
lora_alpha = 32
lora_dropout = 0.05
max_seq_length = 512
你可以使用
环境变量指定仓库根目录外的配置文件路径：
bash
export GEMMA_TUNER_CONFIG=/path/to/my/config.ini
Modality Configuration
模态配置
Text-Only Fine-Tuning
纯文本微调
Instruction tuning (user/assistant pairs):
ini
[profile:text-instruction]
model = gemma-3n-e2b-it
dataset = my-text-dataset
modality = text
text_sub_mode = instruction
prompt_column = prompt
text_column = response
max_seq_length = 2048
lora_r = 16
lora_alpha = 32
Completion tuning (full sequence trained):
ini
[profile:text-completion]
model = gemma-3n-e2b-it
dataset = my-text-dataset
modality = text
text_sub_mode = completion
text_column = text
max_seq_length = 2048
CSV format for instruction tuning (
data/datasets/my-text-dataset/train.csv
):
csv
prompt,response
"What is photosynthesis?","Photosynthesis is the process by which plants..."
"Explain LoRA fine-tuning","LoRA (Low-Rank Adaptation) is a parameter-efficient..."
指令微调（用户/助手对话对形式）：
ini
[profile:text-instruction]
model = gemma-3n-e2b-it
dataset = my-text-dataset
modality = text
text_sub_mode = instruction
prompt_column = prompt
text_column = response
max_seq_length = 2048
lora_r = 16
lora_alpha = 32
补全微调（训练完整序列）：
ini
[profile:text-completion]
model = gemma-3n-e2b-it
dataset = my-text-dataset
modality = text
text_sub_mode = completion
text_column = text
max_seq_length = 2048
指令微调的
CSV格式（
data/datasets/my-text-dataset/train.csv
）：
csv
prompt,response
"What is photosynthesis?","Photosynthesis is the process by which plants..."
"Explain LoRA fine-tuning","LoRA (Low-Rank Adaptation) is a parameter-efficient..."
ini
[profile:image-caption]
model = gemma-3n-e2b-it
dataset = my-image-dataset
modality = image
image_sub_mode = captioning
image_token_budget = 256
prompt_column = prompt
text_column = caption
max_seq_length = 512
CSV format (
data/datasets/my-image-dataset/train.csv
):
csv
image_path,prompt,caption
/data/images/img1.jpg,Describe this image,A dog sitting on a green lawn...
/data/images/img2.jpg,What is shown here,A bar chart showing quarterly revenue...
ini
[profile:image-caption]
model = gemma-3n-e2b-it
dataset = my-image-dataset
modality = image
image_sub_mode = captioning
image_token_budget = 256
prompt_column = prompt
text_column = caption
max_seq_length = 512
CSV格式（
data/datasets/my-image-dataset/train.csv
）：
csv
image_path,prompt,caption
/data/images/img1.jpg,Describe this image,A dog sitting on a green lawn...
/data/images/img2.jpg,What is shown here,A bar chart showing quarterly revenue...
ini
[profile:audio-asr]
model = gemma-3n-e2b-it
dataset = my-audio-dataset
modality = audio
audio_column = audio_path
text_column = transcript
max_seq_length = 512
lora_r = 16
lora_alpha = 32
lora_dropout = 0.05
CSV format (
data/datasets/my-audio-dataset/train.csv
):
csv
audio_path,transcript
/data/audio/recording1.wav,The patient presents with acute respiratory symptoms
/data/audio/recording2.wav,Counsel objects to the characterization of the evidence
ini
[profile:audio-asr]
model = gemma-3n-e2b-it
dataset = my-audio-dataset
modality = audio
audio_column = audio_path
text_column = transcript
max_seq_length = 512
lora_r = 16
lora_alpha = 32
lora_dropout = 0.05
CSV格式（
data/datasets/my-audio-dataset/train.csv
）：
csv
audio_path,transcript
/data/audio/recording1.wav,The patient presents with acute respiratory symptoms
/data/audio/recording2.wav,Counsel objects to the characterization of the evidence
| Model Key | Hugging Face ID | Notes |
|---|
| | Default, ~2B instruct |
| | ~4B instruct |
| | Needs requirements-gemma4.txt |
| | Needs requirements-gemma4.txt |
| | Base, needs Gemma 4 stack |
| | Base, needs Gemma 4 stack |
Add custom models with a
section using
.
| 模型键名 | Hugging Face ID | 备注 |
|---|
| | 默认模型，约2B参数的对话版 |
| | 约4B参数的对话版 |
| | 需要安装requirements-gemma4.txt依赖 |
| | 需要安装requirements-gemma4.txt依赖 |
| | 基座模型，需要Gemma 4依赖栈 |
| | 基座模型，需要Gemma 4依赖栈 |
Dataset Directory Layout
数据集目录结构
data/
└── datasets/
└── <dataset-name>/
├── train.csv # required
├── validation.csv # optional
└── test.csv # optional
data/
└── datasets/
└── <dataset-name>/
├── train.csv # 必需
├── validation.csv # 可选
└── test.csv # 可选
output/
└── {run-id}-{profile}/
├── metadata.json
├── metrics.json
├── checkpoint-*/
└── adapter_model/ # LoRA artifacts
output/
└── {run-id}-{profile}/
├── metadata.json
├── metrics.json
├── checkpoint-*/
└── adapter_model/ # LoRA产物
Python API Examples
Python API示例
Running Fine-Tuning Programmatically
编程式运行微调
python
from gemma_tuner.core.config import load_config
from gemma_tuner.core.ops import run_finetune
python
from gemma_tuner.core.config import load_config
from gemma_tuner.core.ops import run_finetune
config = load_config("config/config.ini")
config = load_config("config/config.ini")
Run fine-tuning for a profile
Run fine-tuning for a profile
run_finetune(profile="my-audio-profile", config=config, json_logging=True)
run_finetune(profile="my-audio-profile", config=config, json_logging=True)
Using Device Utilities
使用设备工具
python
from gemma_tuner.utils.device import get_device, memory_hint
device = get_device() # Returns "mps", "cuda", or "cpu"
print(f"Training on: {device}")
hint = memory_hint(model_key="gemma-3n-e2b-it")
print(hint)
python
from gemma_tuner.utils.device import get_device, memory_hint
device = get_device() # Returns "mps", "cuda", or "cpu"
print(f"Training on: {device}")
hint = memory_hint(model_key="gemma-3n-e2b-it")
print(hint)
Loading and Inspecting Datasets
加载与查看数据集
python
from gemma_tuner.utils.dataset_utils import load_csv_dataset
train_df, val_df = load_csv_dataset(
data_dir="data/datasets/my-text-dataset",
text_column="response",
prompt_column="prompt"
)
print(f"Train samples: {len(train_df)}, Val samples: {len(val_df)}")
python
from gemma_tuner.utils.dataset_utils import load_csv_dataset
train_df, val_df = load_csv_dataset(
data_dir="data/datasets/my-text-dataset",
text_column="response",
prompt_column="prompt"
)
print(f"Train samples: {len(train_df)}, Val samples: {len(val_df)}")
Custom LoRA Config
自定义LoRA配置
python
from peft import LoraConfig, get_peft_model
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained(
"google/gemma-3n-E2B-it",
torch_dtype="auto",
device_map="mps"
)
lora_config = LoraConfig(
r=16,
lora_alpha=32,
lora_dropout=0.05,
target_modules=["q_proj", "v_proj", "k_proj", "o_proj"],
task_type="CAUSAL_LM"
)
model = get_peft_model(model, lora_config)
model.print_trainable_parameters()
python
from peft import LoraConfig, get_peft_model
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained(
"google/gemma-3n-E2B-it",
torch_dtype="auto",
device_map="mps"
)
lora_config = LoraConfig(
r=16,
lora_alpha=32,
lora_dropout=0.05,
target_modules=["q_proj", "v_proj", "k_proj", "o_proj"],
task_type="CAUSAL_LM"
)
model = get_peft_model(model, lora_config)
model.print_trainable_parameters()
Full Workflow: Text Instruction Tuning
完整流程：文本指令微调
1. Prepare your data
1. 准备数据
mkdir -p data/datasets/my-dataset
cp train.csv data/datasets/my-dataset/
cp validation.csv data/datasets/my-dataset/
mkdir -p data/datasets/my-dataset
cp train.csv data/datasets/my-dataset/
cp validation.csv data/datasets/my-dataset/
2. Add profile to config/config.ini
2. 在config/config.ini中添加运行配置
cat >> config/config.ini << 'EOF'
[dataset:my-dataset]
data_dir = data/datasets/my-dataset
[profile:my-text-run]
model = gemma-3n-e2b-it
dataset = my-dataset
modality = text
text_sub_mode = instruction
prompt_column = prompt
text_column = response
max_seq_length = 2048
lora_r = 16
lora_alpha = 32
EOF
cat >> config/config.ini << 'EOF'
[dataset:my-dataset]
data_dir = data/datasets/my-dataset
[profile:my-text-run]
model = gemma-3n-e2b-it
dataset = my-dataset
modality = text
text_sub_mode = instruction
prompt_column = prompt
text_column = response
max_seq_length = 2048
lora_r = 16
lora_alpha = 32
EOF
3. Prepare dataset
3. 预处理数据集
gemma-macos-tuner prepare my-dataset
gemma-macos-tuner prepare my-dataset
gemma-macos-tuner finetune my-text-run --json-logging
gemma-macos-tuner finetune my-text-run --json-logging
5. Export merged weights
5. 导出合并后的权重
gemma-macos-tuner export my-text-run
gemma-macos-tuner export my-text-run
GCS Streaming for Large Datasets
针对大规模数据集的GCS流式加载
ini
[dataset:large-audio-gcs]
source = gcs
gcs_bucket = my-bucket
gcs_prefix = audio-training-data/
audio_column = audio_path
text_column = transcript
[profile:large-audio-run]
model = gemma-3n-e4b-it
dataset = large-audio-gcs
modality = audio
lora_r = 32
lora_alpha = 64
Set credentials:
bash
export GOOGLE_APPLICATION_CREDENTIALS=/path/to/service-account.json
gemma-macos-tuner finetune large-audio-run
ini
[dataset:large-audio-gcs]
source = gcs
gcs_bucket = my-bucket
gcs_prefix = audio-training-data/
audio_column = audio_path
text_column = transcript
[profile:large-audio-run]
model = gemma-3n-e4b-it
dataset = large-audio-gcs
modality = audio
lora_r = 32
lora_alpha = 64
设置凭证：
bash
export GOOGLE_APPLICATION_CREDENTIALS=/path/to/service-account.json
gemma-macos-tuner finetune large-audio-run
Add a Custom Gemma Checkpoint
添加自定义Gemma检查点
ini
[model:my-custom-gemma]
group = gemma
base_model = my-org/my-gemma-checkpoint
[profile:custom-run]
model = my-custom-gemma
dataset = my-dataset
modality = text
text_sub_mode = instruction
ini
[model:my-custom-gemma]
group = gemma
base_model = my-org/my-gemma-checkpoint
[profile:custom-run]
model = my-custom-gemma
dataset = my-dataset
modality = text
text_sub_mode = instruction
Wrong architecture (x86_64 instead of arm64)
架构错误（显示x86_64而非arm64）
bash
python -c "import platform; print(platform.machine())"
bash
python -c "import platform; print(platform.machine())"
Must be arm64 — if x86_64, reinstall Python natively:
必须返回arm64 — 如果是x86_64，请重新安装原生arm64版本的Python：
brew install python@3.12
python3.12 -m venv .venv && source .venv/bin/activate
brew install python@3.12
python3.12 -m venv .venv && source .venv/bin/activate
- Reduce (try 1)
- Increase
gradient_accumulation_steps
- Use a smaller model ( instead of )
- Reduce
- 减小（可尝试设为1）
- 增大
gradient_accumulation_steps
- 使用更小的模型（选版本而非版本）
- 减小
Gemma 4 model not loading
Gemma 4模型加载失败
Gemma 4 requires the updated Transformers stack
Gemma 4需要更新版本的Transformers栈
pip install -r requirements/requirements-gemma4.txt
pip install -r requirements/requirements-gemma4.txt
Use a separate venv if you also need Gemma 3n
如果你同时需要使用Gemma 3n，建议使用独立的虚拟环境
Config not found outside repo root
无法找到仓库外的配置文件
bash
export GEMMA_TUNER_CONFIG=/absolute/path/to/config/config.ini
gemma-macos-tuner finetune my-profile
bash
export GEMMA_TUNER_CONFIG=/absolute/path/to/config/config.ini
gemma-macos-tuner finetune my-profile
Hugging Face auth errors
Hugging Face认证错误
bash
huggingface-cli login
bash
huggingface-cli login
export HF_TOKEN=your_hf_token
export HF_TOKEN=your_hf_token
System check before debugging anything else
调试前先执行系统检查
bash
gemma-macos-tuner system-check
bash
gemma-macos-tuner system-check
Audio tower loaded even for text-only runs
纯文本运行时也加载了音频塔权重
This is a known v1 issue — USM audio tower weights stay in memory even for
. See
. Workaround: use a smaller model variant to stay within RAM budget.
这是v1版本的已知问题 — 即使设置了
，USM音频塔权重仍会保留在内存中，详见
。临时解决方案：使用更小的模型变体，控制内存占用在RAM预算内。
Architecture Reference
架构参考
| File | Role |
|---|
| Main CLI entrypoint () |
| Dispatches prepare/finetune/evaluate/export |
gemma_tuner/scripts/finetune.py
| Router: Gemma models → |
gemma_tuner/models/gemma/finetune.py
| Core training loop with LoRA |
gemma_tuner/scripts/export.py
| Merges LoRA → HF/SafeTensors tree |
gemma_tuner/utils/device.py
| MPS/CUDA/CPU selection and memory hints |
gemma_tuner/utils/dataset_utils.py
| CSV loading, blacklist/protection semantics |
| Interactive CLI wizard (questionary + Rich) |
| Hierarchical INI configuration |
| 文件路径 | 作用 |
|---|
| CLI主入口（对应命令） |
| 分发预处理/微调/评估/导出任务 |
gemma_tuner/scripts/finetune.py
| 路由：将Gemma模型请求转发到 |
gemma_tuner/models/gemma/finetune.py
| 基于LoRA的核心训练循环实现 |
gemma_tuner/scripts/export.py
| 合并LoRA权重，导出为HF/SafeTensors格式 |
gemma_tuner/utils/device.py
| MPS/CUDA/CPU设备选择与内存提示功能 |
gemma_tuner/utils/dataset_utils.py
| CSV加载、黑名单/样本保护语义实现 |
| 交互式CLI向导实现（基于questionary + Rich） |
| 层级化INI配置文件 |