axquant 1.3.0__tar.gz

axquant-1.3.0/.github/workflows/ci.yml

@@ -0,0 +1,62 @@
+name: CI
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+    branches: [main]
+
+permissions:
+  contents: read
+
+jobs:
+  lint:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v7
+      - uses: actions/setup-python@v7
+        with:
+          python-version: "3.13"
+      - name: Install dev dependencies
+        run: pip install -e ".[dev]"
+      - name: Ruff lint
+        run: ruff check .
+      - name: Ruff format
+        run: ruff format --check .
+      - name: mypy
+        run: mypy src
+
+  python-compatibility:
+    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        python-version: ["3.11", "3.12", "3.13"]
+    steps:
+      - uses: actions/checkout@v7
+      - uses: actions/setup-python@v7
+        with:
+          python-version: ${{ matrix.python-version }}
+      - name: Install dev dependencies
+        run: pip install -e ".[dev]"
+      - name: Run non-MLX tests
+        run: pytest -m "not integration"
+
+  test:
+    runs-on: macos-14
+    steps:
+      - uses: actions/checkout@v7
+      - uses: actions/setup-python@v7
+        with:
+          python-version: "3.13"
+      - name: Install with MLX extras
+        run: pip install -e ".[dev,mlx]"
+      - name: Run test suite
+        run: pytest -m "not integration"
+      - name: Run real-hardware integration suite
+        run: |
+          if test -f tests/test_integration_mac.py; then
+            pytest -m integration
+          else
+            echo "No tracked real-hardware integration tests yet"
+          fi

axquant-1.3.0/.github/workflows/release.yml

@@ -0,0 +1,112 @@
+name: Release
+
+on:
+  push:
+    tags: ["v*"]
+
+jobs:
+  dist:
+    runs-on: ubuntu-latest
+    permissions:
+      attestations: write
+      contents: write
+      id-token: write
+    steps:
+      - uses: actions/checkout@v7
+      - uses: actions/setup-python@v7
+        with:
+          python-version: "3.13"
+      - name: Verify tag matches package version
+        run: >-
+          python -c 'import os, tomllib; version =
+          tomllib.load(open("pyproject.toml", "rb"))["project"]["version"];
+          tag = os.environ["GITHUB_REF_NAME"];
+          assert tag == f"v{version}", f"tag {tag!r} does not match package v{version}"'
+      - name: Build sdist and wheel
+        run: |
+          pip install build
+          python -m build
+      - name: Generate SHA256SUMS
+        run: |
+          cd dist
+          shasum -a 256 -- * > SHA256SUMS.txt
+      - name: Attest build provenance
+        uses: actions/attest-build-provenance@v4
+        with:
+          subject-path: "dist/*"
+      - name: Extract curated release notes from CHANGELOG.md
+        run: >-
+          python -c 'import os, pathlib, re; version =
+          os.environ["GITHUB_REF_NAME"].removeprefix("v");
+          text = pathlib.Path("CHANGELOG.md").read_text(encoding="utf-8");
+          m = re.search(rf"^## \[{re.escape(version)}\][^\n]*\n(.*?)(?=^## \[|\Z)",
+          text, re.S | re.M);
+          assert m and m.group(1).strip(),
+          f"CHANGELOG.md has no notes section for {version}";
+          pathlib.Path("release-notes.md").write_text(m.group(1).strip() + "\n",
+          encoding="utf-8")'
+      - name: Upload artifacts to the GitHub Release
+        run: |
+          gh release view "${GITHUB_REF_NAME}" >/dev/null 2>&1 || \
+            gh release create "${GITHUB_REF_NAME}" --verify-tag \
+              --title "${GITHUB_REF_NAME}" --notes-file release-notes.md
+          gh release edit "${GITHUB_REF_NAME}" --title "${GITHUB_REF_NAME}"
+          gh release upload "${GITHUB_REF_NAME}" dist/* --clobber
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+      - name: Hand dists to the publish job
+        uses: actions/upload-artifact@v7
+        with:
+          name: dist
+          path: dist/*
+          retention-days: 1
+      - name: Note PyPI publish gate
+        run: |
+          if [ "${{ vars.ENABLE_PYPI_PUBLISH }}" = "true" ]; then
+            echo "ENABLE_PYPI_PUBLISH=true — pypi job will run Trusted Publishing."
+          else
+            echo "ENABLE_PYPI_PUBLISH is not true — pypi job skipped."
+            echo "GitHub Release assets are still published by this dist job."
+            echo "After registering a Trusted Publisher on pypi.org for"
+            echo "defai-digital/axquant workflow release.yml, set the repository"
+            echo "variable ENABLE_PYPI_PUBLISH=true and re-run Release."
+            echo "See docs/ci-root-causes.md."
+          fi
+
+  # OIDC Trusted Publishing requires a matching publisher on pypi.org / test.pypi.org.
+  # Without that operator config, the exchange fails with invalid-publisher and reds
+  # every tag push even though GitHub Release assets already succeeded. Gate the job
+  # on repository variable ENABLE_PYPI_PUBLISH so Releases stay green until configured.
+  pypi:
+    needs: dist
+    if: ${{ vars.ENABLE_PYPI_PUBLISH == 'true' }}
+    runs-on: ubuntu-latest
+    permissions:
+      contents: read
+      id-token: write
+    steps:
+      - uses: actions/download-artifact@v8
+        with:
+          name: dist
+          path: dist
+      - name: Drop non-package files
+        run: rm -f dist/SHA256SUMS.txt
+      # Prerelease tags (vX.Y.ZrcN) go to TestPyPI; final tags go to PyPI.
+      - name: Publish to TestPyPI (prerelease)
+        if: contains(github.ref_name, 'rc')
+        uses: pypa/gh-action-pypi-publish@release/v1
+        with:
+          repository-url: https://test.pypi.org/legacy/
+      - name: Publish to PyPI
+        if: ${{ !contains(github.ref_name, 'rc') }}
+        uses: pypa/gh-action-pypi-publish@release/v1
+      - name: Trusted publisher guidance
+        if: failure()
+        run: |
+          cat <<'EOF'
+          PyPI Trusted Publishing failed (invalid-publisher is usually missing config).
+          Register a Trusted Publisher on pypi.org / test.pypi.org for:
+            owner=defai-digital  repo=axquant  workflow=release.yml
+          Environment: leave empty (this job does not set one).
+          Then keep ENABLE_PYPI_PUBLISH=true. Details: docs/ci-root-causes.md
+          EOF

axquant-1.3.0/.gitignore

@@ -0,0 +1,26 @@
+.DS_Store
+.coverage
+.mypy_cache/
+.pytest_cache/
+.ruff_cache/
+.venv/
+.qoder/
+.vscode/
+.claude/
+__pycache__/
+build/
+dist/
+*.egg-info/
+*.py[cod]
+*.safetensors
+*.gguf
+*.npz
+.env
+credentials.json
+# Local-only product/engineering notes and scratch (never publish).
+.internal/
+**/.internal/
+# Agent convention docs (internal only; not published to GitHub).
+AGENTS.md
+Agents.md
+agents.md

axquant-1.3.0/CHANGELOG.md

@@ -0,0 +1,291 @@
+# Changelog
+
+All notable changes to AXQuant are documented here. The format follows
+[Keep a Changelog](https://keepachangelog.com/en/1.1.0/) and the project uses
+[Semantic Versioning](https://semver.org/spec/v2.0.0.html). The release workflow extracts the
+matching section from this file as the curated GitHub Release notes and fails the release when
+the section is missing — add an entry in the same change as any user-facing modification.
+
+## [Unreleased]
+
+## [1.3.0] - 2026-08-06
+
+### Added
+
+- Group-preserving GPTQ act-order as the new `gptq-act` method (ADR-0002): groups process in
+  descending Hessian-mass order and columns by Hessian diagonal, while group membership and the
+  portable affine packed layout stay byte-compatible with static-grid GPTQ. Registered as a
+  distinct probe/plan/convert method label so the measured frontier — not the literature —
+  decides adoption per tensor.
+- Measured-holdout-safe interaction optimization (`optimize_candidate_interactions`, ADR-0004):
+  complete-candidate selection driven by measured evaluations on development dataset roles,
+  with a fail-closed guard that rejects formal-holdout or missing `dataset_role` provenance and
+  a separate `interaction_measurement_set_sha256` binding so the formal holdout binding stays
+  free for final selection.
+- Kernel-latency-aware planning (ADR-0003): `benchmark-kernels` measures decode/prefill GEMM
+  latency per (bits, group size) on the current host into a host-scoped
+  `axquant.kernel-latency.v1` table; `plan --latency-table` re-ranks candidates by measured
+  kernel speed strictly inside the quality near-tie window and records
+  `cost_model` / `kernel_latency_sha256` / host provenance on the plan. Without a table,
+  planning is bit-identical to before.
+- Method near-tie surfacing (RM-14): plans record `method_near_ties` (capped, most fragile
+  first, with an explicit omitted count) wherever a losing packing method lands within the
+  configurable `method_near_tie_epsilon` of the winner at the same storage key.
+- Report-only measured-KV serving-quality artifact (`axquant.kv-serving-quality.v1`, RM-21):
+  binds the executed per-layer KV plan digest to dual-profile short/long-context quality
+  retention versus BF16 KV, and fails closed unless the mlx-lm-kv runtime check proved exact
+  per-layer execution.
+- Formal MTP A/B admissibility (`formal_mtp_bundle_issues`, RM-20): an MTP off/on evaluation
+  pair is authorizing only when both halves bind the frozen formal-host contract (device, chip,
+  OS, power mode), share identical controls, datasets, seeds, and checkpoint, and report zero
+  kernel fallbacks.
+- Opt-in quantized MTP sidecar (`quantize_qwen36_mtp_sidecar`, ADR-0005): a separate
+  `axquant-portable-affine-u8` artifact emitted alongside the untouched byte-preserved default,
+  gated on a recorded passing AX Engine capability check for the quantized MTP layout.
+- End-to-end CLI and audit wiring for the new evidence paths: `refine-select --interaction`
+  runs the holdout-safe interaction selection; `quantize-mtp-sidecar` executes a live AX Engine
+  capability probe (`--capability-command`, subprocess JSON contract) or consumes a recorded
+  check (`--capability-result`) before emitting the quantized sidecar; `kv-serving-quality`
+  builds the report-only KV artifact from an executed plan, kv_exec summary, and measured
+  results; and the flagship M7 gate now loads both MTP A/B bundles from the release validation
+  chain and enforces `formal_mtp_bundle_issues` against the frozen `mbp-m5` contract and the
+  digest-bound hardware-registry device identity.
+
+### Changed
+
+- Experimental 2/3-bit assignment is restricted to robust-trunk tensor classes (MLP and expert
+  projections, RM-42); attention and all protected roles keep 4-bit-and-up candidates, and
+  experimental plans now list every low-bit tensor with its predicted loss in the plan warnings.
+- The default hardware profile is `ax-engine-apple-silicon-affine-dwq-v3`, adding `gptq-act`
+  to the supported method set. Previously serialized plans keep their recorded v2 profile.
+
+- Qwen3-ASR 1.7B and Qwen3-VL 8B Instruct text-path conversion through public MLX-Audio and
+  MLX-VLM APIs, protected BF16 modality towers, architecture-specific runtime metadata, media
+  generation smokes, and four stable-name AXQ v2 development repositories.
+- Revision-pinned Qwen3-ASR BF16 normalization in `scripts/hf_to_mlx_bf16.py`; the helper forces
+  the public MLX-Audio STT backend and records the key/layout remap in `axquant_source.json`.
+- Explicit development-artifact edition metadata and the completed 28-repository AXQ v2
+  migration from 14 immutable sources. Stable repository names now serve receipt-bound v2
+  revisions on `main` and tag `v2`, while `legacy-pre-v2` preserves each replaced artifact.
+  Every promoted revision passed pre/post-upload MLX-LM checks, reverified Hub trees and hashes,
+  evidence-safe sibling links/model-card metadata, and a verified stable-name catalog entry.
+- Additive `qwen36-mtp-v2` flagship certification contracts: path-neutral `CheckpointKey` and
+  `CandidateKey`, frozen campaigns with disjoint dataset roles, exact-`mbp-m5` preflight,
+  formal-cycle/holdout consumption state, durable archive proof, independent review, clean-host
+  reproduction review, a separate final publication-claim review, and final M0–M8 dispatch.
+- Append-only artifact lifecycle events and semantic impact scans. Certified candidates become
+  ineligible for certified rendering/publication after `superseded` or `revoked`.
+- Deterministic measured-BPW certified naming, bound public metric claims, and generated
+  certified model cards.
+- CLI commands `campaign-freeze`, `campaign-preflight`, `campaign-start-formal`,
+  `campaign-complete-formal`, `campaign-close-no-go`, `campaign-overlap`,
+  `campaign-frontier`, `campaign-record-publication`, `artifact-lifecycle`, and `claim-render`.
+- Checksum-bound cheapest-failure-first frontier/no-go records and formal raw-evidence/custodian
+  attestations, so candidate selection, formal failure, and no-feasible-candidate closure are
+  evidence-backed rather than boolean declarations.
+- Downloaded publication inventory and runtime-verification records; `published` campaign state is
+  unreachable until the exact Hub revision, checkpoint bytes, final audit, claim, lifecycle,
+  AX Engine/MLX-LM smokes, and zero-fallback result are reverified.
+- A fail-closed flagship publication privacy scan for credentials, private host/path data,
+  oversized or invalid text evidence, sensitive filenames, and formal raw holdout files. It
+  rechecks the exact post-packaging tree and scans legitimate large `tokenizer.json` assets
+  without treating them as oversized evidence.
+- A non-symlinked durable-root boundary for the complete campaign/state-transition tree and
+  six-part formal-host evidence whose results bind the exact frozen `mbp-m5` contract.
+
+### Fixed
+
+- Qwen3-ASR inspection now rejects the unnormalized upstream `thinker.*` layout for planning and
+  conversion instead of counting its duplicated tied LM head as an independently convertible
+  tensor. The error points to the pinned BF16 normalization helper.
+- Multimodal runtime smokes require real transcript/image-generation output: ASR no longer passes
+  on log text without a transcript, VLM runs non-verbose, and simple conversion rejects a smoke
+  from the wrong architecture backend before conversion.
+- Qwen3-ASR and Qwen3-VL model-card commands now use MLX-Audio's required `--output-path` and
+  MLX-VLM's supported `--temperature` option.
+- Converted-checkpoint verification now accepts only documented MLX-LM output identities:
+  one-to-one wrapper aliases, Qwen packed gate/up one-to-many splits, contiguous indexed-expert
+  many-to-one stacks, and the exact Qwen 3.5/Qwen3-Next/Nemotron-H `Conv1d` sanitize-axis
+  transforms. Membership, precision, parameter count, metadata, and shapes remain fail-closed.
+- Budget allocation preserves the deterministic marginal-gain order with a priority queue instead
+  of rescanning every tensor after every upgrade, removing quadratic planning time on
+  74,000-tensor Qwen3-Coder-Next checkpoints.
+- The AXQ v2 migration uses immutable Mistral-published BF16 sources and materializes only
+  `model.safetensors.index.json`-bound shards when upstream also ships a redundant consolidated
+  checkpoint; AXQuant's unindexed-Safetensors rejection remains strict.
+- Development cards no longer advertise an AX Engine serve path when the package has no validated
+  native `model-manifest.json`; MLX-LM remains the documented standard text/backbone path.
+- Qwen3 Embedding development cards link their real stable-name 4-bit/8-bit siblings instead of
+  a nonexistent 6-bit variant.
+- Release-bound identity comparisons ignore absolute `local_path` while preserving it as
+  execution provenance; checkpoint/candidate byte digests still fail closed on content drift.
+- Generation-smoke no longer requires the `mlx_lm` Python package; advisory KV / runtime
+  metadata is validated before install gates so Ubuntu non-MLX CI matches Mac developer paths.
+- Gemma4 sharded source prep validates index shard paths and types before importing MLX.
+- mypy on the Ubuntu lint job ignores missing optional `mlx` / `mlx_lm` / `transformers`
+  modules (they are only installed via `axquant[mlx]`).
+
+### Changed
+
+- Existing `4bit`/`6bit` repository names remain development identifiers. A certified flagship
+  must use generated `MP-<measured-main-BPW>bpw[-MTP]` naming and cannot downgrade to the
+  historical Qwen 3.6 v4 audit/publisher path.
+- CI and release workflows use current Node 24 GitHub Actions majors and grant permissions
+  per job, reducing deprecated-runtime warnings while keeping PyPI OIDC and build provenance
+  least-privilege.
+- Documented historical Actions red-run root causes and added `scripts/ci-local.sh` to mirror
+  CI gates (including PATH-isolated non-MLX pytest) before push.
+- Release `pypi` job is gated on repository variable `ENABLE_PYPI_PUBLISH=true` so tag
+  Releases succeed with GitHub assets when Trusted Publishing is not yet configured on
+  pypi.org (set the variable only after registering a publisher).
+
+### Notes
+
+- Real PyPI upload still requires a Trusted Publisher on pypi.org plus
+  `ENABLE_PYPI_PUBLISH=true` (operator steps in docs/ci-root-causes.md).
+- Local-only engineering/product notes stay out of the published tree: force-added
+  certification docs were removed from git (still ignored on developer machines).
+
+## [1.2.0] - 2026-08-04
+
+### Added
+
+- `analyze --capture-points`: measured probing on plain dense backbones (MiniCPM5, Qwen3 dense,
+  Mistral, …) via `--capture-points output`; the default `("output", "hidden")` fails closed
+  with a hint on those layouts.
+- `refine --lm-head-floor`, matching `plan`; the governed `8bit` path still requires measured
+  support.
+- Structured `mtp_sidecar_bits` runtime contracts in copied (byte-preserved) MTP sidecar
+  metadata.
+- New evaluation task categories: JSON, tool-use, multilingual, and long-context.
+
+### Changed
+
+- Calibration evidence is checksum-bound end to end: AWQ/GPTQ analysis records the
+  activation-capture manifest digest, tokenized-cache manifest digest, cache key, and dataset
+  ID; the planner carries the bindings; conversion rejects a mismatched capture and packages
+  `activation_capture_manifest.json` with the checkpoint; publication preparation and release
+  audit independently revalidate it.
+- Completed capture directories are immutable (`completion.json` present → no overwrite), and
+  every committed resume chunk is validated against checkpoint row accounting.
+- `scripts/hf_to_mlx_bf16.py` requires `--revision <40-char SHA>`, converts into a staging
+  directory, atomically renames on success, and writes `axquant_source.json` with the immutable
+  source identity.
+- Mixed-precision plans are labeled by target BPW instead of inheriting a flat storage-class
+  name.
+
+### Fixed
+
+- Qwen3-Next adapter: fused `switch_mlp`/`switch_glu` 3-D weights now classify as `expert`.
+  **Pre-fix artifacts may have retained most expert weights at BF16 and must be regenerated.**
+- Capture resume now fails closed on binding drift (model, cache, `--max-rows`, segments, token
+  budget) instead of silently mixing runs.
+- AWQ/GPTQ convert-time refinement no longer crashes on bfloat16 module weights (the default
+  dtype of real checkpoints).
+- Affine/DWQ/AWQ quantizers build the per-group grid from the float16-stored scale (rounded
+  up, never down), restoring the documented ≤ scale/2 per-element round-trip bound.
+- `plan`/`refine` inject BF16 for policy-preserved tensors (norms, LM head, vision) when 16 is
+  omitted from `--bits`, matching manual planning, instead of failing per tensor; manual
+  tied-weight harmonization now refreshes scale/outlier strategies and stored BPW.
+- Feasibility audits require MTP evidence only from MTP-declaring architectures, making the
+  Qwen3-Next direct-track N0 gate satisfiable with tool-produced reports.
+- Measured probe/KV-probe runs no longer abort when the calibration token count leaves a
+  single-token trailing replay batch.
+- The `kv-exec` runtime check now fails when executed per-layer KV bits differ from the plan
+  (previously reported but not enforced), and benchmark trial timeouts kill the whole
+  AX Engine process group instead of orphaning the generation.
+- MTP benchmark aggregates pair numerators and denominators over counter-reporting trials
+  only, instead of inflating `effective_tokens_per_forward` when some trials lack counters.
+- `prepare_development_model_card` sanitizes the manifest's calibration reference along with
+  the plan's (local-path leak that also broke the release-audit M1 equality invariant).
+- Release audit records issues instead of crashing on wheels missing `axquant/__init__.py`,
+  corrupt measurement-set or calibration files, and non-scalar benchmark metadata; hardware
+  registry checks record all-failed raw benchmarks instead of raising.
+- `--kv-default-bits` rejects sub-policy-floor values (2/3) at parse time instead of failing
+  late in planning.
+- `scripts/hf_to_mlx_bf16.py` writes the lm_head-synthesized shard aside and swaps atomically;
+  saving onto the mmap-backed source silently corrupted every tensor in the shard.
+- Ext4T scripts: union-sync verification compares content only (mtime-only differences no
+  longer deadlock syncs), empty-array expansions no longer crash stock macOS bash 3.2 under
+  `set -u`, shell-config marker blocks no longer corrupt an rc file lacking a trailing
+  newline, live log/publish directories are moved aside before migrate-and-delete, and
+  `--status`/argument handling exit correctly.
+- Architecture-gate errors and inventory warnings name the adapter registry instead of
+  hardcoding Qwen 3.6 (the checks were already registry-wide).
+- Ext4T planning treats `axquant/logs` packages as host-local: layout moves only — no
+  cross-host mirroring, duplicate deletion, or fingerprint-driven renames of mutable log
+  trees; index loading reports the exact file and line for corrupt or incomplete records.
+
+See the [v1.1.x → v1.2.0 migration guide](docs/migration-v1.2.md) for action items.
+
+## [1.1.1] - 2026-08-03
+
+### Added
+
+- Resumable, compressed, shardable activation capture: `capture_progress.json` +
+  `activations/.partial/` resume, deflate-compressed npz output, `--segment-batches` and
+  `--modules-per-shard` controls.
+- Real-hardware end-to-end integration tests (`pytest -m integration` on Apple Silicon).
+- CI workflow (lint on Ubuntu, full MLX suite on `macos-14`) and a signed release workflow
+  (SHA256SUMS + keyless Sigstore attestation).
+- Migration guide, environment compatibility matrix, and known-issues list under `docs/`.
+
+### Changed
+
+- GPTQ peak memory cut ~30% at `in_features = 8192` (6.7 GB → 4.7 GB).
+
+### Fixed
+
+- BF16 activations are cast inside MLX before numpy export in capture.
+- `cosine_distance` is clamped to non-negative against float rounding.
+
+## [1.1.0] - 2026-08-03
+
+### Added
+
+- GPTQ Hessian error-compensated weight refinement, wired through the conversion predicate and
+  the `convert` CLI.
+- End-to-end AWQ: `capture-activations` records checksum-bound per-module Linear input
+  activations; measured AWQ/GPTQ probing via `analyze --calibration-activations`; planner
+  selection with the `gptq-hessian` scale strategy; convert-time refinement via
+  `convert --calibration-activations`.
+- Development model card generation for published checkpoints.
+
+### Changed
+
+- Python API renames: `convert_model(awq_activations=...)` → `calibration_activations=`,
+  `PlanPredicate.awq_metadata` → `method_metadata`; `load_capture_activations` returns a
+  mapping-compatible `LoadedActivationCapture`.
+- Measured-probe backend moved to `axquant-mlx-isolated-probe-v6`; saved probe resume state
+  from older versions is rejected.
+
+## [1.0.2] - 2026-08-03
+
+### Added
+
+- Qwen3-Embedding and Qwen3-Next / Coder-Next conversion support.
+- HF → MLX BF16 helper script and an expanded Hub catalog in the README.
+
+## [1.0.1] - 2026-08-02
+
+### Fixed
+
+- Protection-floor and path-safety gaps found on re-audit; six high-severity and three
+  medium-severity release-safety bugs.
+
+## [1.0.0] - 2026-08-01
+
+### Added
+
+- Initial public release: inspect → calibrate → analyze → plan → convert → validate → publish
+  pipeline with checksum-bound artifacts, evidence gating, fail-closed conversion, family
+  adapters (Qwen 3.6 primary track, Qwen 3.5, Gemma-4, MiniCPM5, Mistral/Devstral, Mistral3,
+  Nemotron 3 Nano), and AX Engine runtime manifests.
+
+[1.3.0]: https://github.com/defai-digital/axquant/compare/v1.2.0...v1.3.0
+[1.2.0]: https://github.com/defai-digital/axquant/compare/v1.1.1...v1.2.0
+[1.1.1]: https://github.com/defai-digital/axquant/compare/v1.1.0...v1.1.1
+[1.1.0]: https://github.com/defai-digital/axquant/compare/v1.0.2...v1.1.0
+[1.0.2]: https://github.com/defai-digital/axquant/compare/v1.0.1...v1.0.2
+[1.0.1]: https://github.com/defai-digital/axquant/compare/v1.0.0...v1.0.1
+[1.0.0]: https://github.com/defai-digital/axquant/releases/tag/v1.0.0

axquant-1.3.0/CONTRIBUTING.md

@@ -0,0 +1,94 @@
+# Contributing to AXQuant
+
+Contributions are warmly welcome. AXQuant exists to help Mac users get more reliable, efficient
+local inference and a better overall experience on Apple Silicon. Bug fixes, documentation,
+tests, usability improvements, runtime compatibility work, architecture adapters, and
+reproducible quantization research can all move that goal forward.
+
+First-time contributors are welcome too. If you are unsure where a change belongs, open a
+[GitHub issue](https://github.com/defai-digital/axquant/issues) and describe what you want to
+improve.
+
+## Fork and submit a pull request
+
+1. Fork [AXQuant](https://github.com/defai-digital/axquant) on GitHub.
+2. Create a focused branch in your fork.
+3. Set up the development environment and make your change.
+4. Add or update tests and documentation where appropriate.
+5. Run the checks below.
+6. Push your branch and open a pull request against `defai-digital/axquant`.
+
+Small, well-scoped fixes can go directly to a pull request. For a substantial design change or a
+new model family, please open an issue first so contributors can agree on scope, runtime support,
+and the evidence needed for promotion.
+
+## Development setup
+
+```bash
+python3.13 -m venv .venv
+source .venv/bin/activate
+python -m pip install -e ".[dev]"
+```
+
+On Apple Silicon, install the optional MLX execution dependencies when your change needs them:
+
+```bash
+python -m pip install -e ".[dev,mlx]"
+```
+
+Most tests use small synthetic Safetensors fixtures and do not require model downloads or real
+weights.
+
+## Run the checks
+
+The local CI mirror is the preferred full check:
+
+```bash
+./scripts/ci-local.sh
+```
+
+You can also run the checks individually:
+
+```bash
+.venv/bin/ruff check .
+.venv/bin/ruff format --check .
+.venv/bin/mypy src
+.venv/bin/pytest -m "not integration"
+```
+
+Real-hardware tests require Apple Silicon and the MLX dependencies:
+
+```bash
+.venv/bin/pytest -m integration
+```
+
+## Pull request guidance
+
+A helpful pull request:
+
+- explains what changed, why it helps users, and any known limitations;
+- stays focused enough to review and test;
+- includes tests for changed behavior and documentation for user-facing changes;
+- reports the commands and hardware used for validation;
+- labels architecture priors, measured evidence, runtime smokes, and certification claims
+  accurately; and
+- does not include credentials, private paths, downloaded model weights, or local-only evidence.
+
+Maintainers may ask for additional evidence when a change affects conversion correctness,
+checkpoint compatibility, quality claims, or release gates.
+
+## Clean-room and evidence boundary
+
+AXQuant is independently implemented against public MLX, MLX-LM, MLX-Audio, and MLX-VLM
+interfaces. Contributions must not import, vendor, translate, or copy mlx-optiq implementation
+code, tests, documentation, calibration data, or generated metadata.
+
+New model families begin at the `inspect-only` support tier. Promotion requires real-checkpoint
+classification, conversion, and runtime evidence. Architecture-prior output must never be
+presented as measured sensitivity, and a successful conversion or runtime smoke must not be
+presented as quality certification.
+
+## License
+
+By submitting a contribution, you agree that it may be distributed under the project's
+[MIT License](LICENSE).

axquant-1.3.0/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 AutomatosX / DEFAI Digital
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.