orbitquant 0.1.2__tar.gz → 0.1.4__tar.gz

{orbitquant-0.1.2 → orbitquant-0.1.4}/.github/workflows/publish-pypi.yml

@@ -6,7 +6,7 @@ on:
       version:
         description: Version to publish
         required: true
-        default: "0.1.2"
+        default: "0.1.4"
 
 jobs:
   build:

{orbitquant-0.1.2 → orbitquant-0.1.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: orbitquant
-Version: 0.1.2
+Version: 0.1.4
 Summary: Clean-room OrbitQuant implementation for diffusion transformers
 Project-URL: Homepage, https://github.com/iamwavecut/OrbitQuant
 Project-URL: Repository, https://github.com/iamwavecut/OrbitQuant
@@ -39,7 +39,7 @@ Requires-Dist: diffusers>=0.35; extra == 'hf'
 Requires-Dist: transformers>=4.53; extra == 'hf'
 Provides-Extra: kernels
 Requires-Dist: kernels>=0.16; extra == 'kernels'
-Requires-Dist: triton>=3.5; extra == 'kernels'
+Requires-Dist: triton>=3.5; (platform_system == 'Linux') and extra == 'kernels'
 Description-Content-Type: text/markdown
 
 # OrbitQuant
@@ -402,6 +402,42 @@ reproduction targets:
 User-facing comparison assets are generated at the native settings above. Small
 low-resolution checks are not accepted as published quality evidence.
 
+To create a local original-vs-OrbitQuant evidence bundle without writing raw
+samples into the model artifact repository, run `compare-native` against a local
+OrbitQuant artifact:
+
+```bash
+orbitquant compare-native \
+  --suite flux2-native \
+  --artifact ./artifacts/flux2-klein-w4a4 \
+  --prompt-id simple-object \
+  --output ./reports/native/flux2-klein-w4a4-comparison \
+  --device cuda \
+  --dtype bfloat16 \
+  --runtime-mode auto_fused
+```
+
+The output directory contains the BF16 sample, the OrbitQuant sample, a
+side-by-side WebP comparison, per-split metadata, and `summary.json` with native
+settings, wall time, peak VRAM, requested runtime mode, and available local
+kernel backends. By default `compare-native` validates the completed bundle
+before returning and includes the validation report in its JSON output. Use
+`--skip-comparison-validation` only when you intentionally want to inspect or
+repair an incomplete bundle.
+
+Validate a copied or previously generated comparison bundle before using it as
+evidence:
+
+```bash
+orbitquant validate-comparison \
+  --input ./reports/native/flux2-klein-w4a4-comparison
+```
+
+The validator accepts bundles copied from a remote GPU host, checks native
+settings and metadata, resolves remote absolute paths by local filename, rejects
+blank-like images, and reports timing, VRAM, and selected OrbitQuant runtime
+counts when the bundle was generated by a recent `compare-native`.
+
 ## Release Metrics
 
 Full GenEval and VBench runs are release evidence for paper reproduction,
@@ -443,6 +479,10 @@ orbitquant external-eval-script \
 For Wan, use the same `external-eval-script` path with `--suite wan-native` to
 run VBench custom-input dimensions against the native 832x480, 81-frame videos.
 
+Imported GenEval `geneval_overall` follows upstream GenEval semantics: average
+over task scores. Image-level and prompt-level hit rates are imported as
+`geneval_image_accuracy` and `geneval_prompt_accuracy` when present.
+
 Final publication gates are tracked in
 [docs/release-gates.md](docs/release-gates.md).
 

{orbitquant-0.1.2 → orbitquant-0.1.4}/README.md

@@ -358,6 +358,42 @@ reproduction targets:
 User-facing comparison assets are generated at the native settings above. Small
 low-resolution checks are not accepted as published quality evidence.
 
+To create a local original-vs-OrbitQuant evidence bundle without writing raw
+samples into the model artifact repository, run `compare-native` against a local
+OrbitQuant artifact:
+
+```bash
+orbitquant compare-native \
+  --suite flux2-native \
+  --artifact ./artifacts/flux2-klein-w4a4 \
+  --prompt-id simple-object \
+  --output ./reports/native/flux2-klein-w4a4-comparison \
+  --device cuda \
+  --dtype bfloat16 \
+  --runtime-mode auto_fused
+```
+
+The output directory contains the BF16 sample, the OrbitQuant sample, a
+side-by-side WebP comparison, per-split metadata, and `summary.json` with native
+settings, wall time, peak VRAM, requested runtime mode, and available local
+kernel backends. By default `compare-native` validates the completed bundle
+before returning and includes the validation report in its JSON output. Use
+`--skip-comparison-validation` only when you intentionally want to inspect or
+repair an incomplete bundle.
+
+Validate a copied or previously generated comparison bundle before using it as
+evidence:
+
+```bash
+orbitquant validate-comparison \
+  --input ./reports/native/flux2-klein-w4a4-comparison
+```
+
+The validator accepts bundles copied from a remote GPU host, checks native
+settings and metadata, resolves remote absolute paths by local filename, rejects
+blank-like images, and reports timing, VRAM, and selected OrbitQuant runtime
+counts when the bundle was generated by a recent `compare-native`.
+
 ## Release Metrics
 
 Full GenEval and VBench runs are release evidence for paper reproduction,
@@ -399,6 +435,10 @@ orbitquant external-eval-script \
 For Wan, use the same `external-eval-script` path with `--suite wan-native` to
 run VBench custom-input dimensions against the native 832x480, 81-frame videos.
 
+Imported GenEval `geneval_overall` follows upstream GenEval semantics: average
+over task scores. Image-level and prompt-level hit rates are imported as
+`geneval_image_accuracy` and `geneval_prompt_accuracy` when present.
+
 Final publication gates are tracked in
 [docs/release-gates.md](docs/release-gates.md).
 

{orbitquant-0.1.2 → orbitquant-0.1.4}/docs/kernel-audit.md

@@ -21,9 +21,9 @@ for the current artifact format and runtime modes.
   the current environment.
 - `scripts/run_cuda_kernel_checks.sh` is the CUDA correctness and benchmark
   gate for GPU hosts. It first tries to load a prebuilt
-  `native_packed_matmul` package through Hugging Face `kernels`,
-  `LOCAL_KERNELS`, or an importable package. If no compatible prebuilt package
-  is available, it fails before starting a source build unless
+  `native_packed_matmul` package from an importable `orbitquant_packed_matmul`
+  module, then through Hugging Face `kernels`/`LOCAL_KERNELS`. If no compatible
+  prebuilt package is available, it fails before starting a source build unless
   `ORBITQUANT_ALLOW_NATIVE_KERNEL_BUILD=1` is set. With that explicit opt-in,
   it builds the exact
   `native-kernels/orbitquant-packed-matmul` kernel-builder redistributable
@@ -284,6 +284,13 @@ for the current artifact format and runtime modes.
   snapshot model repo resolved to commit
   `cb0ceb1a4d070556c52cfba691aba3f6647c246b`. The native loader therefore
   still requires `LOCAL_KERNELS`, an importable package, or Kernel Hub approval.
+- A 2026-07-09T15:06Z live refresh found no new reviewer reply after the
+  2026-07-09T14:00Z model-artifact numbers. Discussion 15 remains open,
+  `repo_info("WaveCut/orbitquant-packed-matmul", repo_type="model")` resolves
+  to commit `cb0ceb1a4d070556c52cfba691aba3f6647c246b`, and
+  `repo_info("WaveCut/orbitquant-packed-matmul", repo_type="kernel")` still
+  returns 404. Do not describe the native package as remotely loadable until a
+  kernel-type repository exists and `get_kernel(...)` has been verified.
 
 ## Packaging Boundary
 

{orbitquant-0.1.2 → orbitquant-0.1.4}/docs/kernel-hub-approval-request.md

@@ -170,6 +170,16 @@ approval blocker.
 On 2026-07-09T14:00Z, `WaveCut` posted those actual model artifact numbers in
 discussion 15.
 
+Live status refresh:
+
+On 2026-07-09T15:06Z, discussion 15 was still open and no newer reviewer reply
+was present after the 2026-07-09T14:00Z model-artifact numbers. The public
+source snapshot repo still resolved as a model repo at commit
+`cb0ceb1a4d070556c52cfba691aba3f6647c246b`; the kernel-type repository still
+returned 404. The remaining action is Hugging Face Kernel Hub publish approval,
+followed by `kernel-builder build-and-upload` and `get_kernel(...)`
+verification.
+
 ## Title
 
 Request Kernel Hub publish access for `WaveCut/orbitquant-packed-matmul`

orbitquant-0.1.4/docs/release-0.1.3.md

@@ -0,0 +1,25 @@
+# OrbitQuant 0.1.3 Release Notes
+
+OrbitQuant 0.1.3 is a packaging patch for cross-platform kernel extras.
+
+## Changes
+
+- The `kernels` extra remains the default install target for optimized
+  packed-weight inference support:
+
+  ```bash
+  pip install "orbitquant[kernels]==0.1.3"
+  ```
+
+- `triton>=3.5` is now constrained to Linux installs of the `kernels` extra.
+  This keeps CUDA/Triton support available on Linux while allowing macOS/MPS
+  users to install `orbitquant[kernels]` for Hugging Face `kernels` package
+  support without trying to resolve an unavailable Triton wheel.
+- The package version, CLI version, and `orbitquant.__version__` are aligned at
+  `0.1.3`.
+
+## Claim Boundary
+
+This release does not change quantization math, artifact format, model cards,
+runtime dispatch policy, or kernel implementations. It only fixes optional
+dependency resolution for platforms where Triton is not published.

orbitquant-0.1.4/docs/release-0.1.4.md

@@ -0,0 +1,36 @@
+# OrbitQuant 0.1.4 Release Notes
+
+OrbitQuant 0.1.4 is a release-metrics patch for external GenEval and VBench
+metric import.
+
+## Install
+
+```bash
+pip install "orbitquant[hf]==0.1.4"
+```
+
+For optimized packed-weight inference dependencies:
+
+```bash
+pip install "orbitquant[kernels]==0.1.4"
+```
+
+## Changes
+
+- Imported GenEval `geneval_overall` now follows upstream GenEval semantics:
+  average over task scores.
+- GenEval image-level and prompt-level hit rates are imported as
+  `geneval_image_accuracy` and `geneval_prompt_accuracy` when they are present
+  in the external results.
+- VBench external eval commands now pass requested dimensions as separate CLI
+  arguments, matching upstream `--dimension` parsing.
+- VBench prompt files now use exported video filenames as keys for custom input
+  evaluation.
+- README and generated model cards document GenEval and VBench release metric
+  semantics.
+
+## Claim Boundary
+
+This release does not change quantization math, artifact format, runtime
+dispatch, kernel implementations, or published model weights. Release-grade
+GenEval/VBench runs remain separate from this package patch.

{orbitquant-0.1.2 → orbitquant-0.1.4}/docs/release-gates.md

@@ -196,7 +196,13 @@ URL, or signed-off audit note.
   source build unless `ORBITQUANT_ALLOW_NATIVE_KERNEL_BUILD=1` is set. A
   2026-07-09 prebuilt-only loader check still returned 404 for the Kernel Hub
   repo while the public source snapshot model repo resolved to commit
-  `cb0ceb1a4d070556c52cfba691aba3f6647c246b`.
+  `cb0ceb1a4d070556c52cfba691aba3f6647c246b`. A 2026-07-09T15:06Z live
+  refresh confirmed discussion 15 remains open, the public source snapshot
+  model repo still resolves to commit
+  `cb0ceb1a4d070556c52cfba691aba3f6647c246b`, and
+  `repo_info("WaveCut/orbitquant-packed-matmul", repo_type="kernel")` still
+  returns 404. The loadable Kernel Hub package therefore remains an external
+  approval gate, not a completed release artifact.
 - [x] Final paper conformance audit is complete against arXiv 2607.02461, with
   documented deviations, implementation notes, and evidence that accepted
   deviations are intentional. The required audit checklist is
@@ -228,7 +234,11 @@ URL, or signed-off audit note.
   metadata, zero remote checksum mismatches, zero README mismatches, and zero
   forbidden remote files. It also reported `public_count=14`,
   `private_count=0`, 14/14 policy-inventory-ready, and 144 missing
-  release-grade metrics.
+  release-grade metrics. A 2026-07-09T15:06Z rerun with `--summary-only` and
+  the same policy inventory root again passed `--fail-on-artifact-regression`
+  with 14/14 artifact-ready, native-smoke-ready, metadata-complete, and
+  policy-inventory-ready; it reported `release_eval_applicable_count=10`,
+  `release_eval_ready_count=0`, and `missing_required_metric_count=144`.
 - [x] Full-model module classification inventories are captured for FLUX.2
   Klein, FLUX.1-schnell, Z-Image-Turbo, and Wan2.1. Raw inventory JSON may
   remain unpublished, but each published artifact manifest must be
@@ -244,9 +254,9 @@ URL, or signed-off audit note.
   `scripts/run_hf_compat_checks.sh --mode all`, using the current Torch base.
   This gate uses registration, pipeline quantization config, and mini
   integration tests; it does not download model weights or generate samples.
-  Evidence: passed on 2026-07-08T15:47Z with Torch 2.12.1, current/release
-  Diffusers 0.39.0 and Transformers 5.13.0, and dev Diffusers 0.40.0.dev0 plus
-  Transformers 5.14.0.dev0.
+  Evidence: passed on 2026-07-09T17:18Z with Torch 2.12.1, current/release
+  Diffusers 0.39.0 and Transformers 5.13.0, and dev Diffusers 0.40.0.dev0
+  at `208704a` plus Transformers 5.14.0.dev0 at `0bc3554`.
 - [x] Checkpoint and model repositories are published with artifact-focused
   model cards, complete file manifests, checksums where applicable, and native
   comparison assets for the advertised targets. Cards must describe the
@@ -258,7 +268,9 @@ URL, or signed-off audit note.
   `repo_count=14`, `artifact_ready_count=14`, `metadata_complete_ready_count=14`,
   `native_smoke_ready_count=14`, `remote_checksum_mismatch_count=0`,
   `readme_mismatch_count=0`, `forbidden_file_count=0`, `public_count=14`, and
-  `private_count=0`.
+  `private_count=0`. A 2026-07-09T15:06Z `--summary-only` rerun preserved the
+  same compact artifact counts and still reported `forbidden_file_count=0`,
+  `remote_checksum_mismatch_count=0`, and `readme_mismatch_count=0`.
 - [x] The GitHub repository is public, tagged, released, and includes the
   release docs, license, source distribution expectations, and reproducible
   verification commands.
@@ -289,7 +301,27 @@ URL, or signed-off audit note.
   `5972b6cfd1d89653fb9ac17668f72818c61a0e3cc5ea1cdd46e59d54405dc1ff` and
   `orbitquant-0.1.1-py3-none-any.whl` SHA256
   `b829c5df00093e697872ca24104e6ef38dbe7e7d70b2c2e33560bfaec224cde1`.
-- [x] The PyPI package is published as `orbitquant==0.1.1`.
+  Patch release notes for compact artifact audit output are
+  [release-0.1.2.md](release-0.1.2.md). GitHub CI for commit `f18feb9`
+  passed as run `29025114082`, tag `v0.1.2` points at commit
+  `f18feb95f7595965f046b3455997d6ce9b8e8e4a`, and GitHub Release
+  `https://github.com/iamwavecut/OrbitQuant/releases/tag/v0.1.2` was published
+  on 2026-07-09 with exact PyPI-matching assets:
+  `orbitquant-0.1.2.tar.gz` SHA256
+  `4afc15589fe713345dfa47be30dc078d943484a1f4bbd4861a413532a6ec8377` and
+  `orbitquant-0.1.2-py3-none-any.whl` SHA256
+  `3bdddebaa46f60307ed50e2ebf4b7ff4fef7817845b512ecfbb6fbf8ba71c91c`.
+  Patch release notes for the cross-platform kernel-extra dependency fix are
+  [release-0.1.3.md](release-0.1.3.md). GitHub CI for commit `02166ec`
+  passed as run `29034308427`, tag `v0.1.3` points at commit
+  `02166ec6bdcd8920b0012a1ff930dce8dd976fdb`, and GitHub Release
+  `https://github.com/iamwavecut/OrbitQuant/releases/tag/v0.1.3` was published
+  on 2026-07-09 with exact PyPI-matching assets:
+  `orbitquant-0.1.3.tar.gz` SHA256
+  `008f949641d00df46f840580c424b9e9fad0d853b4345d454a0b8042c61f3366` and
+  `orbitquant-0.1.3-py3-none-any.whl` SHA256
+  `181e9b532a07312ec47b54091d651b5a62d5aefd32c88d40830bd8529a0fdc53`.
+- [x] The PyPI package is published as `orbitquant==0.1.3`.
   Evidence: commit `ce5c232` added the manual
   `.github/workflows/publish-pypi.yml` Trusted Publishing workflow. A PyPI
   pending publisher was registered for project `orbitquant`, owner
@@ -316,6 +348,30 @@ URL, or signed-off audit note.
   `OrbitQuantConfig().runtime_mode == "auto_fused"`,
   `OrbitQuantConfig().packed_matmul_block_n == 128`, and
   `orbitquant --version == "0.1.1"`.
+  Patch publish run `29025225397` completed successfully on 2026-07-09 from
+  head SHA `f18feb95f7595965f046b3455997d6ce9b8e8e4a`, including full pytest,
+  `ruff check`, package build, `twine check`, wheel smoke, OIDC publication,
+  and PyPI digital attestations. The PyPI JSON API reports version `0.1.2` with
+  `orbitquant-0.1.2.tar.gz` SHA256
+  `4afc15589fe713345dfa47be30dc078d943484a1f4bbd4861a413532a6ec8377` and
+  `orbitquant-0.1.2-py3-none-any.whl` SHA256
+  `3bdddebaa46f60307ed50e2ebf4b7ff4fef7817845b512ecfbb6fbf8ba71c91c`; a fresh
+  venv installed `orbitquant==0.1.2` from PyPI and verified
+  `orbitquant.__version__ == "0.1.2"`,
+  `OrbitQuantConfig().runtime_mode == "auto_fused"`,
+  `orbitquant --version == "0.1.2"`, and `orbitquant audit-hf-artifacts --help`
+  includes `--summary-only`.
+  Patch publish run `29034405916` completed successfully on 2026-07-09 from
+  head SHA `02166ec6bdcd8920b0012a1ff930dce8dd976fdb`, including full pytest,
+  `ruff check`, package build, `twine check`, wheel smoke, OIDC publication,
+  and PyPI digital attestations. The PyPI JSON API reports version `0.1.3` with
+  `orbitquant-0.1.3.tar.gz` SHA256
+  `008f949641d00df46f840580c424b9e9fad0d853b4345d454a0b8042c61f3366` and
+  `orbitquant-0.1.3-py3-none-any.whl` SHA256
+  `181e9b532a07312ec47b54091d651b5a62d5aefd32c88d40830bd8529a0fdc53`; a fresh
+  venv installed `orbitquant[kernels]==0.1.3` from PyPI on macOS and verified
+  `orbitquant.__version__ == "0.1.3"`, installed `kernels==0.16.0`, and did
+  not install Triton because the `triton>=3.5` extra is Linux-only.
 - [x] ComfyUI compatibility is verified after the relevant schema stabilizes,
   including load, graph execution, artifact metadata behavior, and kernel extra
   install guidance for the default `auto_fused` runtime.
@@ -332,7 +388,53 @@ URL, or signed-off audit note.
   `ruff`, package build, and `twine check`, and passed GitHub CI run
   `29020564152` on 2026-07-09. A live GitHub check then reported
   `iamwavecut/ComfyUI-OrbitQuant` as `PUBLIC`, Apache-2.0 licensed, with
-  default branch `main`.
+  default branch `main`. Follow-up commit `85527ee` required
+  `orbitquant>=0.1.1` after the packed-matmul default update and passed
+  GitHub CI run `29022774661`. Commit `4832d4a` requires
+  `orbitquant>=0.1.2`, updates the README install commands for both
+  `orbitquant>=0.1.2` and `orbitquant[kernels]>=0.1.2`, refreshes `uv.lock`
+  to the PyPI `orbitquant` 0.1.2 release hashes, and passed GitHub CI run
+  `29027011708` on 2026-07-09. GitHub Release
+  `https://github.com/iamwavecut/ComfyUI-OrbitQuant/releases/tag/v0.1.0` was
+  published on 2026-07-09T15:12Z from commit
+  `4832d4addb39681e82f38e07627a2bb682e4332d` with assets
+  `comfyui_orbitquant-0.1.0.tar.gz` SHA256
+  `630628c56e5ed35626cd7cca6749c51056cb78a0041b7c1268cf0b5e995d28c0` and
+  `comfyui_orbitquant-0.1.0-py3-none-any.whl` SHA256
+  `0c774c20a6759bea18d5d02b598035c3446a72ee4efaa0be6c1f325f4b3e928b`.
+  Before release, local `uv run pytest -q`, `uv run ruff check .`, package
+  build, and `twine check` passed for the node pack. Follow-up commit
+  `ca62b7a` added the manual `.github/workflows/publish-pypi.yml` Trusted
+  Publishing workflow for a future `comfyui-orbitquant` PyPI release. GitHub
+  reported both `CI` and `Publish PyPI` workflows as active, and ComfyUI CI run
+  `29028857599` passed on 2026-07-09 with node-pack tests, lint, package
+  build, package metadata check, and wheel smoke. The GitHub `pypi`
+  environment was created on 2026-07-09T15:21Z. After PyPI password
+  confirmation, the pending publisher for `comfyui-orbitquant` was registered
+  and workflow run `29029534070` published `comfyui-orbitquant==0.1.0` through
+  Trusted Publishing on 2026-07-09. The PyPI version JSON for `0.1.0` reports
+  `comfyui_orbitquant-0.1.0-py3-none-any.whl` SHA256
+  `0c774c20a6759bea18d5d02b598035c3446a72ee4efaa0be6c1f325f4b3e928b` and
+  `comfyui_orbitquant-0.1.0.tar.gz` SHA256
+  `9a2c1e23e7c7674bb8fb7c350473c9ad636c952c5495b71d805a91996a00b51f`. A fresh
+  venv installed `comfyui-orbitquant==0.1.0` from PyPI and imported
+  `comfyui_orbitquant.NODE_CLASS_MAPPINGS`.
+  Current commit `7e5fc4c` requires `orbitquant>=0.1.3`, exposes the
+  `comfyui-orbitquant[kernels]` extra through `orbitquant[kernels]>=0.1.3`,
+  refreshes `uv.lock` to the PyPI OrbitQuant 0.1.3 release, and passed GitHub
+  CI run `29034705959` on 2026-07-09. Publish run `29034761024` published
+  `comfyui-orbitquant==0.1.2` from head SHA
+  `7e5fc4c0f74dbcb341b755b2b53cb0edf40cb311`. GitHub Release
+  `https://github.com/iamwavecut/ComfyUI-OrbitQuant/releases/tag/v0.1.2` was
+  published on 2026-07-09 with exact PyPI-matching assets:
+  `comfyui_orbitquant-0.1.2.tar.gz` SHA256
+  `48ea3909f96620baba3f2acf52532b15330fc89cb847557f13da915969f4e42b` and
+  `comfyui_orbitquant-0.1.2-py3-none-any.whl` SHA256
+  `2fbcf8a4fecb7d4384d2461001fa33b291a02720d20cf465e7e2669ad145fd83`. A fresh
+  venv installed `comfyui-orbitquant[kernels]==0.1.2` from PyPI on macOS and
+  verified `comfyui-orbitquant==0.1.2`, `orbitquant==0.1.3`,
+  `kernels==0.16.0`, no Triton install, and exported FLUX, Z-Image, Wan,
+  generic component-loader, and artifact-inspector nodes.
 - [x] MPS shader-only gate can be run independently from the native packed
   matmul package gate. Evidence: on 2026-07-09,
   `ORBITQUANT_RUN_NATIVE_KERNEL_PACKAGE_CI=0` with tiny benchmark dimensions

{orbitquant-0.1.2 → orbitquant-0.1.4}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "orbitquant"
-version = "0.1.2"
+version = "0.1.4"
 description = "Clean-room OrbitQuant implementation for diffusion transformers"
 readme = "README.md"
 requires-python = ">=3.11"
@@ -38,7 +38,7 @@ hf = [
 ]
 kernels = [
   "kernels>=0.16",
-  "triton>=3.5",
+  "triton>=3.5; platform_system == 'Linux'",
 ]
 eval = [
   "imageio>=2.37",

{orbitquant-0.1.2 → orbitquant-0.1.4}/scripts/run_cuda_kernel_checks.sh

@@ -237,14 +237,14 @@ PY
       printf '%s\n' \
         "native kernel package CI could not load a compatible prebuilt native packed-matmul kernel." \
         "The script will not start a kernel-builder/Nix source build unless ORBITQUANT_ALLOW_NATIVE_KERNEL_BUILD=1 is set, because uncached CUDA builds can compile the CUDA/NCCL stack on paid GPU hosts." \
-        "Use a loadable Hugging Face Kernel Hub artifact, set LOCAL_KERNELS to a compatible built CUDA variant, or rerun with ORBITQUANT_ALLOW_NATIVE_KERNEL_BUILD=1 only on a host where the build cache/toolchain is intentional." \
+        "Make a compatible orbitquant_packed_matmul package importable, use a loadable Hugging Face Kernel Hub artifact, set LOCAL_KERNELS to a compatible built CUDA variant, or rerun with ORBITQUANT_ALLOW_NATIVE_KERNEL_BUILD=1 only on a host where the build cache/toolchain is intentional." \
         >&2
       exit 1
     fi
     if ! command -v nix >/dev/null 2>&1; then
       printf '%s\n' \
         "native kernel package CI requires either a loadable prebuilt native packed-matmul kernel or nix for kernel-builder." \
-        "Install nix, use a GPU image with nix, set LOCAL_KERNELS to a compatible built variant, wait for the Hugging Face kernel repo to become loadable, or set ORBITQUANT_RUN_NATIVE_KERNEL_PACKAGE_CI=0 only when intentionally skipping native package closure." \
+        "Install nix, make a compatible orbitquant_packed_matmul package importable, set LOCAL_KERNELS to a compatible built variant, wait for the Hugging Face kernel repo to become loadable, or set ORBITQUANT_RUN_NATIVE_KERNEL_PACKAGE_CI=0 only when intentionally skipping native package closure." \
         >&2
       exit 1
     fi

{orbitquant-0.1.2 → orbitquant-0.1.4}/scripts/run_mps_kernel_checks.sh

@@ -94,7 +94,7 @@ if [[ "$RUN_NATIVE_KERNEL_PACKAGE_CI" == "1" && -z "${LOCAL_KERNELS:-}" ]]; then
     stage "native-packed-matmul-local-variant-missing status=$native_kernel_variant_status"
     printf '%s\n' \
       "native packed matmul local variant was not found for this MPS runtime." \
-      "Set LOCAL_KERNELS=$NATIVE_KERNEL_REPO_ID=/absolute/path/to/a/built/variant containing metadata.json, build the matching kernel-builder variant, or set ORBITQUANT_RUN_NATIVE_KERNEL_PACKAGE_CI=0 to run only the inline Metal shader checks." \
+      "Make a compatible orbitquant_packed_matmul package importable, set LOCAL_KERNELS=$NATIVE_KERNEL_REPO_ID=/absolute/path/to/a/built/variant containing metadata.json, build the matching kernel-builder variant, or set ORBITQUANT_RUN_NATIVE_KERNEL_PACKAGE_CI=0 to run only the inline Metal shader checks." \
       >&2
   fi
 fi

{orbitquant-0.1.2 → orbitquant-0.1.4}/src/orbitquant/__init__.py

@@ -12,7 +12,7 @@ from orbitquant.pipeline import (
 )
 from orbitquant.quantizer import register_hf_quantizers
 
-__version__ = "0.1.2"
+__version__ = "0.1.4"
 
 register_hf_quantizers()
 

{orbitquant-0.1.2 → orbitquant-0.1.4}/src/orbitquant/artifacts/benchmark.py

@@ -11,6 +11,7 @@ from orbitquant.artifacts.checksums import (
     write_sha256sums_from_manifest,
 )
 from orbitquant.artifacts.manifest import OrbitQuantManifest
+from orbitquant.artifacts.model_card import render_model_card
 from orbitquant.artifacts.validator import validate_required_artifact_files
 
 _VALID_METRIC_SPLITS = {"original", "orbitquant"}
@@ -126,6 +127,10 @@ def record_artifact_metrics(
     )
     if not refresh_checksums_enabled:
         return record
+    (artifact_path / "README.md").write_text(
+        render_model_card(manifest, benchmark_summary=summary),
+        encoding="utf-8",
+    )
     checksums = _write_manifest_with_refreshed_checksums(
         artifact_path, manifest, changed_paths
     )

{orbitquant-0.1.2 → orbitquant-0.1.4}/src/orbitquant/artifacts/model_card.py

@@ -5,6 +5,31 @@ from typing import Any
 from orbitquant.artifacts.manifest import OrbitQuantManifest
 from orbitquant.eval.native_settings import NativeSuite, list_native_suites
 
+_GENEVAL_METRICS = (
+    "geneval_overall",
+    "geneval_per_task_single_object",
+    "geneval_per_task_two_object",
+    "geneval_per_task_counting",
+    "geneval_per_task_colors",
+    "geneval_per_task_position",
+    "geneval_per_task_color_attr",
+)
+_VBENCH_METRICS = (
+    "vbench_imaging_quality",
+    "vbench_aesthetic_quality",
+    "vbench_motion_smoothness",
+    "vbench_dynamic_degree",
+    "vbench_background_consistency",
+    "vbench_subject_consistency",
+    "vbench_scene",
+    "vbench_overall_consistency",
+)
+_RELEASE_METRICS_BY_MODEL = {
+    "black-forest-labs/FLUX.1-schnell": ("GenEval", _GENEVAL_METRICS),
+    "Tongyi-MAI/Z-Image-Turbo": ("GenEval", _GENEVAL_METRICS),
+    "Wan-AI/Wan2.1-T2V-1.3B-Diffusers": ("VBench", _VBENCH_METRICS),
+}
+
 
 def _comparison_assets(checksums: dict[str, str]) -> list[str]:
     assets = []
@@ -168,18 +193,70 @@ def _native_settings_section(source_model_id: str) -> list[str]:
     return lines
 
 
-def _validation_status_section(source_model_id: str) -> list[str]:
-    release_metric = {
-        "black-forest-labs/FLUX.1-schnell": "GenEval",
-        "Tongyi-MAI/Z-Image-Turbo": "GenEval",
-        "Wan-AI/Wan2.1-T2V-1.3B-Diffusers": "VBench",
-    }.get(source_model_id)
+def _latest_metrics_by_split(
+    benchmark_summary: dict[str, Any] | None,
+) -> dict[str, dict[str, Any]]:
+    if not isinstance(benchmark_summary, dict):
+        return {}
+    split_payload = benchmark_summary.get("metrics")
+    if not isinstance(split_payload, dict):
+        return {}
+    metrics_by_split: dict[str, dict[str, Any]] = {}
+    for split in ("original", "orbitquant"):
+        latest = split_payload.get(split, {}).get("latest")
+        if not isinstance(latest, dict):
+            continue
+        metrics = latest.get("metrics")
+        if isinstance(metrics, dict):
+            metrics_by_split[split] = metrics
+    return metrics_by_split
+
+
+def _format_metric_value(value: Any) -> str:
+    if isinstance(value, bool) or not isinstance(value, int | float):
+        return ""
+    return f"`{value:.4g}`"
+
+
+def _release_metric_rows(
+    source_model_id: str,
+    benchmark_summary: dict[str, Any] | None,
+) -> tuple[str | None, list[tuple[str, str, str]]]:
+    release = _RELEASE_METRICS_BY_MODEL.get(source_model_id)
+    if release is None:
+        return None, []
+    release_metric, metric_names = release
+    metrics_by_split = _latest_metrics_by_split(benchmark_summary)
+    rows = []
+    for metric in metric_names:
+        original = metrics_by_split.get("original", {}).get(metric)
+        orbitquant = metrics_by_split.get("orbitquant", {}).get(metric)
+        if original is None and orbitquant is None:
+            continue
+        rows.append(
+            (
+                metric,
+                _format_metric_value(original),
+                _format_metric_value(orbitquant),
+            )
+        )
+    return release_metric, rows
+
+
+def _validation_status_section(
+    source_model_id: str,
+    benchmark_summary: dict[str, Any] | None,
+) -> list[str]:
+    release_metric, release_rows = _release_metric_rows(source_model_id, benchmark_summary)
     release_line = (
-        f"- Release-grade {release_metric} metrics: not included in this artifact."
+        f"- Release-grade {release_metric} metrics: included below."
+        if release_metric
+        and release_rows
+        else f"- Release-grade {release_metric} metrics: not included in this artifact."
         if release_metric
         else "- Release-grade paper metrics: not applicable to this extra target."
     )
-    return [
+    lines = [
         "## Validation Status",
         "",
         "- Native BF16-vs-OrbitQuant comparison: included when the visual matrix "
@@ -188,6 +265,28 @@ def _validation_status_section(source_model_id: str) -> list[str]:
         "- The model card reports artifact-level validation status only.",
         "",
     ]
+    if release_rows:
+        lines.extend(
+            [
+                f"### Release-Grade {release_metric} Metrics",
+                "",
+                (
+                    "GenEval `geneval_overall` follows upstream GenEval semantics: "
+                    "average over task scores."
+                    if release_metric == "GenEval"
+                    else "VBench metrics follow the upstream dimension names."
+                ),
+                "",
+                "| Metric | BF16 source | OrbitQuant |",
+                "| --- | ---: | ---: |",
+            ]
+        )
+        lines.extend(
+            f"| `{metric}` | {original or '-'} | {orbitquant or '-'} |"
+            for metric, original, orbitquant in release_rows
+        )
+        lines.append("")
+    return lines
 
 
 def _native_validation_proof_section(
@@ -253,7 +352,9 @@ def render_model_card(
     bits = f"W{data['weight_bits']}A{data['activation_bits']}"
     comparison_assets = _comparison_assets(data["checksums"])
     native_settings_lines = _native_settings_section(data["source_model_id"])
-    validation_status_lines = _validation_status_section(data["source_model_id"])
+    validation_status_lines = _validation_status_section(
+        data["source_model_id"], benchmark_summary
+    )
     native_validation_proof_lines = _native_validation_proof_section(benchmark_summary)
     adaln_group_size = int(data.get("adaln_group_size", 64))
     adaln_default_note = (