mlx-triage 0.1.1

MLX Inference Quality Diagnostic Toolkit

mlx-triage

Your MLX model is producing garbage. Is it the weights? A known MLX bug? Your quantization settings?

mlx-triage answers that in 30 seconds — without loading the model into memory.

pip install mlx-triage
mlx-triage check ./my-model

What It Checks

Tested against 13 models across 5 families (Llama, Qwen, Phi, LiquidAI, Nanbeige), 4 quantization levels (bf16 through 4-bit), from 0.6B to 30B parameters. Zero false negatives. Full validation results ->

Tier 0 — Sanity Checks (no MLX needed, < 30 seconds)

Check	What it catches
Dtype Compatibility	BF16->FP16 precision loss, training/storage dtype mismatches
Tokenizer & EOS Config	Missing EOS tokens, chat template issues, Llama 3 dual-stop-token edge cases
Weight File Integrity	NaN/Inf values, all-zero layers, corrupt safetensors headers
MLX Version & Known Bugs	Outdated MLX with documented bugs affecting your model architecture

Tier 1 — Statistical Smoke Tests (MLX required)

Check	What it catches
Determinism	Non-reproducible outputs at temp=0 (infrastructure issue, not model)
Reference Divergence	MLX output diverging from PyTorch/Transformers reference
Quantization Quality	Excessive perplexity indicating broken quantization

Install

Requires Python 3.11+ and macOS on Apple Silicon (M1-M4).

# From PyPI
pip install mlx-triage

# With MLX for Tier 1 checks
pip install "mlx-triage[mlx]"

# With reference comparison (Tier 1, Test 1.2)
pip install "mlx-triage[reference]"

# Development
git clone https://github.com/swaylenhayes/mlx-triage.git
cd mlx-triage
uv sync --extra dev

Usage

# Tier 0 only (default — no MLX needed)
mlx-triage check /path/to/model

# Tier 0 + Tier 1
mlx-triage check /path/to/model --tier 1

# JSON output
mlx-triage check /path/to/model --format json

# Save report to file
mlx-triage check /path/to/model --tier 1 --output report.json

Tier 0 runs in under 30 seconds on any model. Tier 1 requires MLX and takes 5-15 minutes depending on model size.

How It Works

mlx-triage uses a tiered diagnostic protocol — each tier increases in depth and cost:

1. Tier 0 reads model files directly (safetensors headers, config JSON, tokenizer config) without loading the model into memory. This catches the most common issues instantly.

2. Tier 1 loads the model via MLX and runs statistical tests — determinism checks (10 runs at temp=0), perplexity measurement against a fixed eval corpus, and optional comparison against a PyTorch reference backend.

3. Tiers 2-3 (planned) will add isolation tests (batch invariance, memory pressure, context length stress) and deep diagnostics (layer-wise activation comparison, cross-runtime analysis).

If Tier 0 finds critical issues, Tier 1 is skipped — fix the fundamentals first.

Known Bugs Database

mlx-triage ships with a curated database of documented MLX bugs (known_bugs.yaml), cross-referenced against your installed MLX version and model architecture. Running MLX < 0.22.0 with float16 weights? It flags the known qmv kernel overflow. Got a 4-bit Llama model looping on long prompts? There's a documented bug for that. Safetensors file looks valid but weights are numerically garbage? That's a known silent bfloat16 corruption path.

Contributing a bug report to the database is the easiest way to help — see CONTRIBUTING.md.

Research Basis

The diagnostic protocol is grounded in systematic analysis of MLX infrastructure defects across multiple model architectures and quantization levels. See METHODOLOGY.md for the evidence basis, including infrastructure defect taxonomy, first-party experiments, and cross-model synthesis.

Contributing

Contributions welcome — especially to the known bugs database. See CONTRIBUTING.md.

License

MIT

---

If mlx-triage saved you a debugging session, star it — it helps other MLX developers find the tool.