Mac chip + RAM method recommender

Pick a VeloxQuant-MLX method from your Apple Silicon chip, unified RAM, model size class, and goal. The CLI is the source of truth; the table below mirrors the same heuristics.

Honesty first

• Key accounting ratios (7.5× RVQ, 16× VecInfer) are packed-byte counters.
• Default RVQ / VecInfer often dequantize into fp16 parent storage.
• Resident savings are more likely for full-KV (rabitq) or eviction (streaming_llm / h2o), especially at long context.

CLI

veloxquant recommend \
  --chip M4 \
  --ram-gb 48 \
  --model-class 7B \
  --goal everyday

# Machine-readable JSON
veloxquant recommend --chip M1 --ram-gb 16 --model-class 3B --goal max_context --json

# Static ruleset (for automation / docs sync)
veloxquant recommend --dump-ruleset > docs-site/static/mac-recommender-ruleset.json

Goal → method map

Goal	Method	Typical key accounting	Resident savings likely?
everyday	turboquant_rvq b=1	~7.5×	No (default dequant path)
max_key_accounting	vecinfer 1-bit	~16×	No (default dequant path)
best_quality	spectral	~5.3×	No
max_context	rabitq	~6× full KV	Yes (more likely)
constant_memory	streaming_llm	n/a (eviction)	Yes (bounded tokens)

RAM fit warnings (rule of thumb)

Approximate 4-bit weight footprints used by the recommender:

Model class	~Weights
1B	0.8 GB
3B	2.0 GB
7B	4.5 GB
14B	8.0 GB
32B	18.0 GB

The CLI also subtracts ~4 GB for OS / apps. If headroom is under ~1 GB it warns. Example: 32B on 24 GB is expected to warn (see also docs/MEMORY_CONSTRAINT_FINDINGS.md).

Example (M4 Pro, 48 GB, 7B, everyday)

method=turboquant_rvq
knobs={'bit_width_inlier': 1, 'seed': 42}
key_accounting_ratio≈7.5x
resident_savings_likely=False

See also

• Validation Report for measuring accounting vs peak memory
• Python module: veloxquant_mlx/tools/mac_recommender.py