mlx-quant-fidelity 0.3.0

Measure MLX quantization quality loss — KL divergence, perplexity, top-token agreement for KV cache and weights

mlx-quant-fidelity

Measure how much quality a quantization costs on Apple Silicon. mlx-quant-fidelity scores a quantized model against a higher-precision reference on the same corpus and reports the drift as numbers you can act on: KL divergence, top-token flip rate, perplexity delta. It measures both KV-cache quantization and weight quantization. No more choosing a bit-width by file size.

The CUDA/GGUF world has had this for years: llama.cpp's --kl-divergence-base, EleutherAI's lm-evaluation-harness. MLX had nothing. This is the MLX version, and it covers the KV-cache and attention angle those tools skip.

Install

pip install mlx-quant-fidelity

Apple Silicon (MLX), Python 3.11+.

Use it

mlx-quant-fidelity kv mlx-community/Llama-3.2-3B-Instruct-4bit --kv-bits 8

Prints a Markdown report. Add --format json for JSON, --kv-bits 4, --kv-group-size 64, or --max-chunks N to bound the corpus.

from mlx_quant_fidelity import measure_kv_fidelity

report = measure_kv_fidelity("mlx-community/Llama-3.2-3B-Instruct-4bit", kv_bits=8)
print(report.kl.mean, report.flip_rate, report.verdict)

Or measure weight quantization — a quantized repo against a higher-precision reference:

mlx-quant-fidelity weights mlx-community/Llama-3.2-3B-Instruct-4bit --reference mlx-community/Llama-3.2-3B-Instruct-bf16

from mlx_quant_fidelity import measure_weight_fidelity

# measure_weight_fidelity(quantized_repo, reference_repo)
report = measure_weight_fidelity(
    "mlx-community/Llama-3.2-3B-Instruct-4bit",  # quantized
    "mlx-community/Llama-3.2-3B-Instruct-bf16",  # reference
)
print(report.kl.mean, report.flip_rate, report.verdict)

What a report looks like

# KV-fidelity: `mlx-community/Llama-3.2-3B-Instruct-4bit` @ 8-bit (group 64)

**Verdict:** good · **mode:** stress (quantize_start=0)

| metric | value |
|---|---|
| KL mean | 0.0002 nats |
| KL median | 0.0001 nats |
| KL p99 | 0.0015 nats |
| KL max | 0.1129 nats |
| flip rate | 0.0065 |
| perplexity Δ | +0.0054 (17.722 → 17.728) |

Measured on **wikitext-2-raw/test**, 51100 positions across 100 chunks of length 512 ...

How much does KV quantization cost?

M1 Max, WikiText-2 test (100 chunks of 512 tokens), stress mode (quantize from token 0). Reproduce any row with mlx-quant-fidelity kv <model> --kv-bits <bits> --max-chunks 100; the full committed reports are under _artifacts/samples/.

Model	KV bits	KL mean (nats)	flip rate	verdict
Llama-3.2-1B	4	0.148	0.20	bad
Llama-3.2-1B	8	0.0004	0.013	marginal
Llama-3.2-3B	4	0.051	0.11	bad
Llama-3.2-3B	8	0.0002	0.007	good
Qwen2.5-7B	4	9.36	0.99	bad
Qwen2.5-7B	8	0.009	0.032	marginal

8-bit KV is near-lossless on all three models. 4-bit is another matter, and Qwen2.5-7B at 4-bit in stress mode falls apart: nearly every token flips. That is the attention sink at work: stress mode quantizes the cache from token 0, including the first tokens attention leans on most, and Qwen2.5 does not tolerate it. mlx-lm's own default keeps the first 5000 tokens full-precision for exactly this reason. Run the tool first and you see it coming.

How much does weight quantization cost?

Same corpus and recipe, but the comparison is now a quantized model repo against a higher-precision reference repo. Reproduce any row with mlx-quant-fidelity weights <quant> --reference <reference> --max-chunks 100; the committed reports are under _artifacts/samples/weights/.

Model	quant	reference	KL mean (nats)	flip rate	perplexity Δ	verdict
Llama-3.2-1B	4-bit	bf16	0.158	0.21	+3.5	marginal
Llama-3.2-1B	8-bit	bf16	0.001	0.023	−0.01	good
Llama-3.2-3B	4-bit	bf16	0.085	0.15	+1.4	marginal
Llama-3.2-3B	8-bit	bf16	0.0009	0.021	0.00	good
Qwen2.5-7B	4-bit	8-bit	0.109	0.16	+0.9	marginal

8-bit weights are near-lossless: about 2% of top tokens flip and perplexity barely moves. 4-bit is a real trade: 15 to 21% of top tokens flip and perplexity climbs a point or more, worst on the small 1B model. The Qwen row compares 4-bit against 8-bit rather than bf16, so its drift is relative to an already-quantized reference, not full precision; the report records that the reference is 8-bit and says so in plain text. The verdict tiers are provisional, anchored to these q8 and q4 reference points on short prose rather than to downstream task accuracy.

Unlike the KV probe, both runs use standard attention, so the drift is the deployed quantized model's weight-quant cost with no quantized-attention kernel folded in. It does still include the quantized-matmul kernel's numerics, which is exactly what you run when you load the model.

Comparing quantizations

compare ranks a set of quantizations on a memory-normalized Pareto frontier: quality (mean KL divergence) on one axis, memory cost on the other. It identifies any configuration that is both worse quality and more expensive than another option on the list — those are dominated and you would never choose them.

# rank weight quantizations against a bf16 reference
mlx-quant-fidelity compare weights q4 q6 q8 --reference fp16

# rank KV configs on a single model
mlx-quant-fidelity compare kv <model> --configs 4:32,4:64,8:64

Add --max-kld 0.05 to get the cheapest configuration whose mean KLD stays under a threshold, or --min-tier good to get the cheapest one that passes the good-tier verdict. docs/ranking-principles.md explains how each axis is computed, what Pareto domination means in practice, and where the ranking has limits.

How it works

Teacher-forced scoring, not generation. For each fixed-length corpus chunk the model runs twice on the same tokens — once with a full-precision KV cache, once with a quantized one — and the two next-token distributions are compared position by position. Generation would let the runs diverge in their own inputs the moment quantization changed a sampled token, turning the measurement into trajectory drift instead of cache cost. Logits collapse to per-position scalars inside the chunk loop and are released before the next chunk, so a long corpus never holds full distributions in memory.

Two modes:

• stress (--quantize-start 0, the default): quantize from token 0. The harsh, apples-to-apples quantizer test.
• deployment (quantize_start > 0): what mlx-lm users actually run, with the first N tokens kept full-precision. Not available yet — see the roadmap.

A run that returns exactly zero drift raises instead of reporting a silent "perfect fidelity." That almost always means quantization never engaged, not that it was free.

The weight probe works the same way with two models instead of two caches: a quantized repo and a reference repo, scored on the same corpus tokens. A compatibility gate refuses a mismatched pair before loading, and a memory pre-flight refuses a pair too large for the device rather than risking a kernel panic.

What the numbers don't say

• A fidelity number is corpus- and context-length-specific. WikiText-2 at temperature 0 measures short-prose distributional drift; the paper this builds on, Accuracy Is Not All You Need, shows that under-predicts task-specific and long-context degradation. Every report records the corpus and the token count so the number is never read as a bare score.
• Perplexity delta is reported for continuity with llama.cpp. It correlates with mean KLD by construction, so treat it as a familiar restatement, not independent corroboration.
• The measured drift bundles the quantizer's error with the quantized-attention kernel's numerics. That is the real end-to-end cost; a quantizer-only control is on the roadmap.

Status

0.3.0, released on PyPI as mlx-quant-fidelity — adds a compare command for memory-normalized Pareto ranking of KV-cache and weight quantizations. 0.2.0 added KV-cache and weight-quantization fidelity probes. Downstream-task accuracy, deployment mode, and more are on the roadmap.

License

Apache-2.0.