mlx-quant-fidelity 0.1.0__tar.gz

mlx_quant_fidelity-0.1.0/.github/workflows/ci.yml

@@ -0,0 +1,21 @@
+name: CI
+on:
+  push:
+    branches: [main, "feature/**"]
+  pull_request:
+    branches: [main]
+
+jobs:
+  test:
+    runs-on: macos-15
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          fetch-depth: 0  # hatch-vcs needs tags/history
+      - uses: astral-sh/setup-uv@v5
+      - run: uv sync --group dev
+      - run: uv run python -c "import mlx.core, mlx_lm; print('mlx import OK')"
+      - run: uv run ruff check .
+      - run: uv run ruff format --check .
+      - run: uv run mypy
+      - run: uv run pytest --cov=mlx_quant_fidelity --cov-report=term-missing --cov-fail-under=85

mlx_quant_fidelity-0.1.0/.github/workflows/release.yml

@@ -0,0 +1,20 @@
+name: release
+on:
+  push:
+    tags: ["v*"]
+
+jobs:
+  build-and-publish:
+    # ubuntu, not macOS: pypa/gh-action-pypi-publish is Linux-only, and the package is a
+    # pure-Python (py3-none-any) wheel, so the build is platform-independent.
+    runs-on: ubuntu-latest
+    environment: pypi
+    permissions:
+      id-token: write  # PyPI Trusted Publishing
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          fetch-depth: 0  # hatch-vcs derives the version from the tag
+      - uses: astral-sh/setup-uv@v5
+      - run: uv build
+      - uses: pypa/gh-action-pypi-publish@release/v1

mlx_quant_fidelity-0.1.0/.gitignore

@@ -0,0 +1,16 @@
+.codegraph/
+.hypothesis/
+docs/superpowers/
+docs/backlog/
+/CLAUDE.md
+/_spike/
+.coverage
+.mypy_cache/
+.pytest_cache/
+.ruff_cache/
+.venv/
+__pycache__/
+build/
+dist/
+*.egg-info/
+src/mlx_quant_fidelity/_version.py

mlx_quant_fidelity-0.1.0/.python-version

@@ -0,0 +1 @@
+3.12

mlx_quant_fidelity-0.1.0/CHANGELOG.md

@@ -0,0 +1,19 @@
+# Changelog
+
+The format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and
+this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
+
+## [0.1.0]
+
+First release. Measures how much quality a KV-cache quantization costs on MLX.
+
+### Added
+
+- KV-cache fidelity probe: teacher-forced paired scoring of a quantized KV cache against full precision on the same model, on identical corpus tokens. Per-chunk reduction to scalars keeps a long corpus from holding full distributions in memory.
+- Metrics in fp32: full-vocab KL divergence (mean, median, p99, max), top-token flip rate, and perplexity delta.
+- `mlx-quant-fidelity kv <model>` CLI and `measure_kv_fidelity(...)` Python API, with JSON and Markdown reports that record the corpus provenance, model revision, and quantize-start mode.
+- Stress mode (quantize from token 0) with an exact-zero guard that refuses to report a silent "perfect fidelity" when quantization never engaged.
+- Cache-capability gate that flags models whose KV cache cannot be quantized instead of crashing mid-run.
+- WikiText-2 test-split loader with a pinned, reproducible chunking contract.
+- Device-derived wired-memory cap installed before any model load, on both the CLI and test paths.
+- Committed sample reports for Llama-3.2-1B, Llama-3.2-3B, and Qwen2.5-7B at 4-bit and 8-bit KV.

mlx_quant_fidelity-0.1.0/LICENSE

@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

mlx_quant_fidelity-0.1.0/NOTICE

@@ -0,0 +1,4 @@
+mlx-quant-fidelity
+Copyright 2026 Denis Ineshin
+
+Licensed under the Apache License, Version 2.0.

mlx_quant_fidelity-0.1.0/PKG-INFO

@@ -0,0 +1,122 @@
+Metadata-Version: 2.4
+Name: mlx-quant-fidelity
+Version: 0.1.0
+Summary: Measure MLX quantization quality loss — KL divergence, perplexity, top-token agreement for KV cache and weights
+Project-URL: Homepage, https://github.com/IonDen/mlx-quant-fidelity
+Project-URL: Source, https://github.com/IonDen/mlx-quant-fidelity
+Project-URL: Issues, https://github.com/IonDen/mlx-quant-fidelity/issues
+Project-URL: Changelog, https://github.com/IonDen/mlx-quant-fidelity/blob/main/CHANGELOG.md
+Project-URL: Roadmap, https://github.com/IonDen/mlx-quant-fidelity/blob/main/ROADMAP.md
+Author-email: Denis Ineshin <denis.ineshin@gmail.com>
+License-Expression: Apache-2.0
+License-File: LICENSE
+License-File: NOTICE
+Keywords: apple-silicon,eval,kl-divergence,kv-cache,mlx,perplexity,quality,quantization
+Classifier: Development Status :: 3 - Alpha
+Classifier: Intended Audience :: Developers
+Classifier: Intended Audience :: Science/Research
+Classifier: License :: OSI Approved :: Apache Software License
+Classifier: Operating System :: MacOS :: MacOS X
+Classifier: Programming Language :: Python :: 3 :: Only
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Classifier: Typing :: Typed
+Requires-Python: >=3.11
+Requires-Dist: huggingface-hub>=0.24
+Requires-Dist: mlx-lm>=0.31.3
+Requires-Dist: mlx>=0.21
+Requires-Dist: numpy>=1.26
+Requires-Dist: pyarrow>=14
+Description-Content-Type: text/markdown
+
+# mlx-quant-fidelity
+
+Measure how much quality a quantization costs on Apple Silicon. `mlx-quant-fidelity` scores a quantized KV cache against full precision on the same model and reports the drift as numbers you can act on: KL divergence, top-token flip rate, perplexity delta. 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.
+
+Version 0.1.0 measures **KV-cache quantization**. Weight-quantization fidelity is next; see the [roadmap](ROADMAP.md).
+
+## Install
+
+```bash
+pip install mlx-quant-fidelity
+```
+
+Apple Silicon (MLX), Python 3.11+.
+
+## Use it
+
+```bash
+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.
+
+```python
+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)
+```
+
+## What a report looks like
+
+```markdown
+# 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/`](_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. The point of the tool is that you can see this for your model before you pick a bit-width.
+
+## 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. Planned for 0.2.0.
+
+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.
+
+## 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
+
+0.1.0, released on PyPI as `mlx-quant-fidelity` — the KV-cache fidelity probe (CLI + Python API, JSON and Markdown reports). Weight-quantization fidelity, downstream-task accuracy, and memory-normalized method ranking are on the [roadmap](ROADMAP.md).
+
+## License
+
+[Apache-2.0](LICENSE).

mlx_quant_fidelity-0.1.0/README.md

@@ -0,0 +1,89 @@
+# mlx-quant-fidelity
+
+Measure how much quality a quantization costs on Apple Silicon. `mlx-quant-fidelity` scores a quantized KV cache against full precision on the same model and reports the drift as numbers you can act on: KL divergence, top-token flip rate, perplexity delta. 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.
+
+Version 0.1.0 measures **KV-cache quantization**. Weight-quantization fidelity is next; see the [roadmap](ROADMAP.md).
+
+## Install
+
+```bash
+pip install mlx-quant-fidelity
+```
+
+Apple Silicon (MLX), Python 3.11+.
+
+## Use it
+
+```bash
+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.
+
+```python
+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)
+```
+
+## What a report looks like
+
+```markdown
+# 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/`](_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. The point of the tool is that you can see this for your model before you pick a bit-width.
+
+## 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. Planned for 0.2.0.
+
+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.
+
+## 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
+
+0.1.0, released on PyPI as `mlx-quant-fidelity` — the KV-cache fidelity probe (CLI + Python API, JSON and Markdown reports). Weight-quantization fidelity, downstream-task accuracy, and memory-normalized method ranking are on the [roadmap](ROADMAP.md).
+
+## License
+
+[Apache-2.0](LICENSE).

mlx_quant_fidelity-0.1.0/ROADMAP.md

@@ -0,0 +1,17 @@
+# Roadmap
+
+`mlx-quant-fidelity` measures how much quality a quantization costs on Apple Silicon / MLX.
+
+## Released
+
+- **0.1.0** — KV-cache quantization fidelity: KL divergence, top-token flip rate, perplexity delta; CLI and Python API; JSON/Markdown reports; WikiText-2 corpus.
+
+## Next
+
+- **Weight-quantization fidelity** — the same drift metrics for q4/q6/q8 weight quantization, with reference-logit handling that respects the full-vocab KLD requirement (stream or recompute rather than persist truncated top-k).
+- **Deployment mode** — `quantize_start > 0`, matching what mlx-lm users run, with the first N tokens kept full-precision.
+- **Downstream-task accuracy** — EleutherAI's lm-evaluation-harness (MMLU, ARC, HellaSwag, GSM8K, HumanEval) alongside distributional drift.
+- **Method comparison** — rank quantization methods on one yardstick, memory-normalized (quality vs KV bytes per token) as a Pareto view, never a raw-metric sort.
+- **Quantizer-only control** — an optional dequantize-then-standard-attention path that separates quantizer error from quantized-kernel numerics.
+- **Wider attention coverage** — MLA and sliding-window caches beyond flag-don't-crash.
+- **Fidelity badge** — a model-card badge driven by an explicit, documented threshold policy.

mlx_quant_fidelity-0.1.0/_artifacts/samples/llama-3.2-1b-4bit-kv4.json

@@ -0,0 +1,37 @@
+{
+  "cache_supported": true,
+  "corpus": {
+    "bos_policy": "none",
+    "chunk_length": 512,
+    "final_chunk_policy": "drop",
+    "n_tokens": 51200,
+    "name": "wikitext-2-raw",
+    "normalization": "raw",
+    "split": "test",
+    "stride": 512,
+    "tokenizer_id": "mlx-community/Llama-3.2-1B-Instruct-4bit"
+  },
+  "flip_rate": 0.20483365949119373,
+  "kl": {
+    "max": 6.249123573303223,
+    "mean": 0.14773540842612012,
+    "median": 0.09692534431815147,
+    "p99": 0.9224590277671838
+  },
+  "kv_bits": 4,
+  "kv_group_size": 64,
+  "mlx_lm_version": "0.31.3",
+  "mlx_version": "0.31.2",
+  "model_id": "mlx-community/Llama-3.2-1B-Instruct-4bit",
+  "model_revision": "08231374eeacb049a0eade7922910865b8fce912",
+  "n_chunks": 100,
+  "n_positions": 51100,
+  "peak_memory_bytes": 2552171857,
+  "perplexity_delta": 4.121097638366695,
+  "perplexity_quant": 27.06214792192995,
+  "perplexity_ref": 22.941050283563253,
+  "quantize_mode": "stress",
+  "quantize_start": 0,
+  "verdict": "bad",
+  "warnings": []
+}

mlx_quant_fidelity-0.1.0/_artifacts/samples/llama-3.2-1b-4bit-kv4.md

@@ -0,0 +1,16 @@
+# KV-fidelity: `mlx-community/Llama-3.2-1B-Instruct-4bit` @ 4-bit (group 64)
+
+**Verdict:** bad · **mode:** stress (quantize_start=0)
+
+| metric | value |
+|---|---|
+| KL mean | 0.1477 nats |
+| KL median | 0.0969 nats |
+| KL p99 | 0.9225 nats |
+| KL max | 6.2491 nats |
+| flip rate | 0.2048 |
+| perplexity Δ | +4.1211 (22.941 → 27.062) |
+
+Measured on **wikitext-2-raw/test**, 51100 positions across 100 chunks of length 512 (tokenizer `mlx-community/Llama-3.2-1B-Instruct-4bit`). Fidelity is corpus- and context-length-specific; short-prose temp-0 drift under-predicts long-context/code degradation.
+
+_mlx 0.31.2, mlx-lm 0.31.3, model rev `08231374eeacb049a0eade7922910865b8fce912`, peak 2.55 GB._

mlx_quant_fidelity-0.1.0/_artifacts/samples/llama-3.2-1b-4bit-kv8.json

@@ -0,0 +1,37 @@
+{
+  "cache_supported": true,
+  "corpus": {
+    "bos_policy": "none",
+    "chunk_length": 512,
+    "final_chunk_policy": "drop",
+    "n_tokens": 51200,
+    "name": "wikitext-2-raw",
+    "normalization": "raw",
+    "split": "test",
+    "stride": 512,
+    "tokenizer_id": "mlx-community/Llama-3.2-1B-Instruct-4bit"
+  },
+  "flip_rate": 0.012583170254403132,
+  "kl": {
+    "max": 0.04656221345067024,
+    "mean": 0.0004305653305290256,
+    "median": 0.00027623827918432653,
+    "p99": 0.002857390865683556
+  },
+  "kv_bits": 8,
+  "kv_group_size": 64,
+  "mlx_lm_version": "0.31.3",
+  "mlx_version": "0.31.2",
+  "model_id": "mlx-community/Llama-3.2-1B-Instruct-4bit",
+  "model_revision": "08231374eeacb049a0eade7922910865b8fce912",
+  "n_chunks": 100,
+  "n_positions": 51100,
+  "peak_memory_bytes": 2556678370,
+  "perplexity_delta": 0.015172139713783395,
+  "perplexity_quant": 22.956222423277037,
+  "perplexity_ref": 22.941050283563253,
+  "quantize_mode": "stress",
+  "quantize_start": 0,
+  "verdict": "marginal",
+  "warnings": []
+}

mlx_quant_fidelity-0.1.0/_artifacts/samples/llama-3.2-1b-4bit-kv8.md

@@ -0,0 +1,16 @@
+# KV-fidelity: `mlx-community/Llama-3.2-1B-Instruct-4bit` @ 8-bit (group 64)
+
+**Verdict:** marginal · **mode:** stress (quantize_start=0)
+
+| metric | value |
+|---|---|
+| KL mean | 0.0004 nats |
+| KL median | 0.0003 nats |
+| KL p99 | 0.0029 nats |
+| KL max | 0.0466 nats |
+| flip rate | 0.0126 |
+| perplexity Δ | +0.0152 (22.941 → 22.956) |
+
+Measured on **wikitext-2-raw/test**, 51100 positions across 100 chunks of length 512 (tokenizer `mlx-community/Llama-3.2-1B-Instruct-4bit`). Fidelity is corpus- and context-length-specific; short-prose temp-0 drift under-predicts long-context/code degradation.
+
+_mlx 0.31.2, mlx-lm 0.31.3, model rev `08231374eeacb049a0eade7922910865b8fce912`, peak 2.56 GB._