メニュー
Guides agents on exporting fine-tuned Gemma models to multi-format mobile runtimes, comparing memory and hardware target characteristics, and implementing dynamic sandbox download workflows.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
Guides agents on preparing datasets, running parameter-efficient LoRA fine-tuning, and tracking telemetry metrics using the mlxtune CLI on Apple Silicon. Use when an agent wants to fine-tune standard Gemma 4 base models on text instructions.
Guides agents on preparing multimodal datasets in interleaved ChatML format, executing audio/vision fine-tuning with mlx-vlm, targeting specific layers, and performing key-sanitized weights fusion to avoid silent blank weights loading.
Guides agents on fine-tuning Gemma 4 Quantization-Aware Training (QAT) unquantized checkpoints, maintaining mathematical alignment during LoRA, and exporting to GGUF using strict q4_0 parameters to avoid post-training quantization drift on edge devices.
SOC 職業分類に基づく
| name | gemma-model-export |
| description | Guides agents on exporting fine-tuned Gemma models to multi-format mobile runtimes, comparing memory and hardware target characteristics, and implementing dynamic sandbox download workflows. |
| compatibility | Requires Python 3.14+, uv, Apple Silicon, and Xcode/Swift compiler |
This skill provides step-by-step instructions for AI coding assistants and agents to compile, merge, and quantize fine-tuned Gemma models into highly optimized mobile runtimes (GGUF, LiteRT-LM, MLX Swift, and CoreML).
Before compilation, determine the best on-device runtime using the following decision patterns:
llama.cpp: Cross-platform (iOS, Android, Desktop). Supports hot-swapping multiple adapters. Runs on CPU/GPU.The repository includes a premium multi-format model compiler script scripts/export_formats.py that streamlines weight pipelines:
Fuse the base model and adapters dequantizing to 16-bit, copy vocab files, and quantize to GGUF in one command:
uv run scripts/export_formats.py gguf \
--base ./model \
--adapter ./adapters \
--dest ./my_model.gguf \
--outtype q4_k_m
Natively merges the LoRA adapter weights inside PyTorch and compiles them using litert-torch with highly optimized dynamic 8-bit quantization (dynamic_wi8_afp32):
uv run scripts/export_formats.py litert \
--base ./model \
--adapter ./adapters \
--dest ./models/litert
mlxtune fuse, pass --prefused ./fused_model_dequantized to bypass expensive on-the-fly PEFT reloading and save over 5GB of RAM overhead.Bakes adapters back and outputs a clean, MLX-compatible directory:
uv run scripts/export_formats.py mlx \
--base ./model \
--adapter ./adapters \
--dest ./models/mlx_model
When building local mobile applications that load these exported binaries, you must comply with strict Apple Sandbox policies:
FileManager.default.urls(for: .applicationSupportDirectory, in: .userDomainMask)). Apple preserves this directory but automatically excludes it from standard iCloud backups.