// LoRA fine-tune a causal LM natively on Apple Silicon (M1/M2/M3/M4) via `mlx-lm`. Triggered when the user is on macOS arm64, mentions MLX, or wants better throughput than torch-MPS. Use this instead of `train-sft` when the host is a Mac with no NVIDIA GPU.
Set up training and inference on AMD Ryzen AI Max+ 395 / Strix Halo (gfx1151, RDNA 3.5) with TheRock nightly ROCm wheels. Triggered when the host has gfx1151, when `rocminfo` shows Strix Halo, or when the user mentions Strix Halo / Ryzen AI Max / gfx1151 / 128GB unified memory.
Recreate the ml-intern-kit Python environment on a new machine (laptop, rented GPU box, Docker, fresh checkout). Triggered when the user is on a new host, sees ImportError on a core dep (torch/transformers/trl/peft/accelerate), or wants to install flash-attn / unsloth / bitsandbytes after the fact.
Evaluate a trained or downloaded language model with `lm-eval-harness` standard tasks (arc, hellaswag, gsm8k, mmlu, truthfulqa, ifeval, ...). Triggered when the user wants to benchmark, eval, or compare a model — pre- or post-training.
Audit a Hugging Face dataset before training to confirm splits, columns, format, sample rows, distributions, and duplicates. Triggered before any training/fine-tuning script runs, when a user mentions a new dataset, or when you hit a KeyError / format mismatch in a training job.
Submit a training/inference script to Hugging Face Jobs (`hf jobs run`). Triggered when the user wants to run training in the cloud, scale beyond local hardware, or kick off a multi-hour fine-tune. Enforces pre-flight: hub_model_id required, ≥2h timeout for training, single-job validation before batch.
Run a literature-first crawl before writing ANY ML training/fine-tuning/inference code. Spawns an Explore sub-agent that mines papers, citation graphs, methodology sections, and matched HF datasets to produce a ranked list of training recipes attributed to specific published results. Triggered when the user asks to fine-tune, train, or improve a model, or when the user names a task/benchmark and you need a recipe before coding.
| name | train-mlx |
| description | LoRA fine-tune a causal LM natively on Apple Silicon (M1/M2/M3/M4) via `mlx-lm`. Triggered when the user is on macOS arm64, mentions MLX, or wants better throughput than torch-MPS. Use this instead of `train-sft` when the host is a Mac with no NVIDIA GPU. |
For everything except Apple Silicon, prefer the train-sft skill (torch +
TRL). MLX wins on M-series because the unified memory means a 7B model that
spills VRAM on a 24GB consumer NVIDIA card fits comfortably on a 36GB M-series.
uname -sm returns Darwin arm64.make env-mlx # bootstrap with --accel=mps --mlx
python scripts/dump_mlx_jsonl.py trl-lib/Capybara data/capybara-mlx
train.jsonl + valid.jsonl with the messages (or text)
field. You can skip this if you already have HF JSONL on disk.mlx-community/SmolLM2-1.7B-Instruct-4bitmlx-community/Llama-3.2-3B-Instruct-4bitmlx-community/Qwen2.5-7B-Instruct-4bitpython scripts/mlx_finetune.py --config configs/mlx_default.yaml --iters 50
python scripts/mlx_finetune.py --config configs/mlx_default.yaml --iters 1000 --fuse
--fuse merges the LoRA back into the base weights so you can load the
result with plain mlx_lm.generate or convert back to safetensors.mlx_lm.fuse --upload-repo you/your-mlx-model.| Model | Min unified memory | Notes |
|---|---|---|
| ≤1.7B | 16 GB | works on M1/M2 base |
| 3B | 24 GB | M2 Pro / M3 Pro |
| 7B 4-bit | 36 GB | M3 Max / M4 Pro |
| 13B 4-bit | 64 GB | M2 Ultra / M3 Max |
train-dpo skill at slower speeds).bitsandbytes-style 8-bit optimizers (MLX has its own quantization).scripts/eval_lm.py (lm-eval-harness) loads MLX-fused safetensors fine. For
in-process MLX eval: mlx_lm.generate --model outputs/mlx-lora-fused --prompt "...".