메뉴
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 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
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.
| 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.