Open source libraries for MLX and Apple Silicon
mlx-dfloat — Run DFloat11 checkpoints on Apple Silicon. DFloat11 compresses BF16 weights losslessly to about 70% of their size, and a Metal kernel unpacks each block on the GPU back to the exact original bits just before it runs. FLUX.1, Z-Image, FLUX.2 Klein, Qwen-Image-2.1, ERNIE-Image and Krea 2 generate images on a Mac.
mlx-guard — External runtime safety supervision for MLX workloads. A small native supervisor sits outside the Python process, samples the real memory footprint of the process group it owns, and escalates to a hard stop against a limit you choose — so a runaway run can't take the whole Mac down with it.
mlx-train-perf — Fused, logit-free cross-entropy loss and a flash-attention training path for mlx-lm LoRA/QLoRA fine-tuning on Apple Silicon. Custom Metal kernels never build the full logits or attention-score matrices — about 3900× less loss-layer memory, and half the training peak at 8K context.
mlx-taef — Tiny AutoEncoders for diffusion latents in pure MLX. Decode FLUX.2 latents in ~260 ms or FLUX.1 in ~185 ms on M1 Max, vs ~2 s for the full VAE, with ~4× less peak memory.
mlx-teacache — TeaCache step-skipping for diffusion on Apple Silicon, in pure MLX. Skips redundant denoising steps across FLUX.1, FLUX.2 Klein and Z-Image — 1.46× on FLUX.1-dev at 25 steps.
mlx-model-doctor — Validate an MLX or Hugging Face model repo before you load it. Static checks on config, tokenizer, safetensors index, quantization and memory budget catch a broken repo up front, so it fails fast with a clear reason instead of a confusing crash.
mlx-quant-fidelity — Measure how much quality a quantization costs on Apple Silicon. Scores a quantized KV cache against full precision and reports the drift — KL divergence, top-token flip rate, perplexity delta — so you pick a bit-width by impact, not by file size.
Skills for Claude Code and OpenAI Codex that I use every day. Each one installs with a single command.
Audits Claude Code subagents and Codex custom agents, then cuts what each one costs to launch: a tight tool allowlist, a shorter prompt, and a model sized to the job. One real agent dropped from 20,122 to 13,962 launch tokens with the same result.
npx skills add IonDen/skills --skill subagent-optimizerShrinks a skill's SKILL.md without losing an instruction. It records every rule and literal in the original first, then checks the rewrite against that record and asks when unsure instead of guessing. On five real skills it trimmed 2.3% and kept every requirement it found.
npx skills add IonDen/skills --skill skill-optimizerKeeps tests honest in any language: they run real code instead of mocks and fail on the bug they exist to catch. Given the same prompts, Claude Haiku without the skill mocked both collaborators in 3 of 3 Python suites. With it, none did.
npx skills add IonDen/skills --skill writing-tests-that-can-failOn unified memory a runaway MLX run can end in a paging storm or a kernel panic rather than a catchable allocation error — why the last line of defense can't live inside the process it's meant to stop.
Read →Dropping the logits matrix freed memory but didn't extend context on a 32 GB Mac — the peak had moved to attention backward. A custom Metal FlashAttention for MLX, and two system models that failed.
Read →The obvious way to measure quantization damage writes itself — and every step of it hides a failure mode that buries damage, measures the wrong quantity, or passes without exercising the feature.
Read →