Open source libraries for MLX and Apple Silicon
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.
A chain of Metal kernels is not one command buffer. How MLX 0.32.0's operation and element thresholds decide commits — and why a whole-chain duration guard used the wrong scheduling unit.
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 →A fused MLX/Metal kernel computes the loss in vocabulary tiles, never storing the full logits matrix — with measured memory savings, tiling limits, and performance models that failed.
Read →