Release 3.15.1 with GPU Docker export fix

anulum · anulum · commit 3535085fa790 · 2026-05-19T03:48:17.000+02:00
Co-Authored-By: Arcane Sapience &lt;protoscience@anulum.li&gt;
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -7,6 +7,13 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [3.15.1] — 2026-05-19
+
+- Add the hash-pinned PyTorch ONNX exporter stack required by the GPU Docker
+  model-builder stage.
+- Update ONNX export supply-chain policy, artefact guidance, and regression
+  coverage for the GPU release image.
+
 ## [3.15.0] — 2026-05-19
 
 ### Added — safety operations retuning (2026-05-06)
diff --git a/CITATION.cff b/CITATION.cff
@@ -14,7 +14,7 @@ authors:
     given-names: "Miroslav"
     orcid: "https://orcid.org/0009-0009-3560-0851"
     affiliation: "ANULUM CH & LI"
-version: "3.15.0"
+version: "3.15.1"
 doi: "10.5281/zenodo.18822167"
 date-released: "2026-05-19"
 license: "AGPL-3.0-or-later"
diff --git a/README.md b/README.md
@@ -72,7 +72,7 @@ graph LR
 
 | Surface | Current inventory |
 |---|---:|
-| Package version | 3.15.0 |
+| Package version | 3.15.1 |
 | Public API exports | 214 |
 | Python capability source modules | 310 |
 | Python capability classes | 701 |
@@ -412,7 +412,7 @@ Kubernetes: [Helm chart](deploy/helm/director-ai/) with GPU toggle, HPA, Sigstor
   title     = {Director-AI: Real-time LLM Hallucination Guardrail},
   year      = {2026},
   url       = {https://github.com/anulum/director-ai},
-  version   = {3.15.0},
+  version   = {3.15.1},
   license   = {AGPL-3.0-or-later}
 }
 ```
diff --git a/docs-site/deployment/onnx-artefacts.md b/docs-site/deployment/onnx-artefacts.md
@@ -101,8 +101,8 @@ read-only.
 ## Wheel Target Matrix
 
 `requirements/onnx_wheel_targets.toml` is the tracked source for these targets.
-`requirements/docker-gpu-export.txt` pins the `onnx` and `onnxruntime`
-build-time wheels.
+`requirements/docker-gpu-export.txt` pins the `onnx`, `onnxruntime`,
+`onnxscript`, `onnx_ir`, `ml_dtypes`, and `protobuf` build-time wheels.
 
 | Target id | Platform | Extra | Runtime package | Execution provider | Status |
 |-----------|----------|-------|-----------------|--------------------|--------|
diff --git a/docs-site/deployment/supply-chain.md b/docs-site/deployment/supply-chain.md
@@ -76,8 +76,12 @@ that are safe to keep hash-pinned outside the default service image:
 
 | Package | File | Controls |
 |---------|------|----------|
+| `ml_dtypes` | `requirements/docker-gpu-export.txt` | hash-pin, isolated-build-stage |
 | `onnx` | `requirements/docker-gpu-export.txt` | hash-pin, isolated-build-stage |
+| `onnx_ir` | `requirements/docker-gpu-export.txt` | hash-pin, isolated-build-stage |
 | `onnxruntime` | `requirements/docker-gpu-export.txt` | hash-pin, isolated-build-stage |
+| `onnxscript` | `requirements/docker-gpu-export.txt` | hash-pin, isolated-build-stage |
+| `protobuf` | `requirements/docker-gpu-export.txt` | hash-pin, isolated-build-stage |
 
 Legacy exporter packages that force `transformers<5` are not installed. That
 keeps the model-loader stack on the audited `transformers>=5.0.0rc3,<6` line.
diff --git a/docs-site/guide/agpl-compliance.md b/docs-site/guide/agpl-compliance.md
@@ -15,7 +15,7 @@ curl https://your-server:8080/v1/source
 ```json
 {
   "license": "AGPL-3.0-or-later",
-  "version": "3.15.0",
+  "version": "3.15.1",
   "repository_url": "https://github.com/anulum/director-ai",
   "instructions": "git clone https://github.com/anulum/director-ai",
   "agpl_section": "13"
diff --git a/docs-site/index.md b/docs-site/index.md
@@ -2,7 +2,7 @@
 
 **Real-time LLM hallucination guardrail** — NLI + RAG fact-checking with token-level streaming halt.
 
-<span class="version-badge">v3.15.0 — 5-tier scoring, 6 advanced RAG strategies, multi-agent swarm guardian, config wizard</span>
+<span class="version-badge">v3.15.1 — 5-tier scoring, 6 advanced RAG strategies, multi-agent swarm guardian, config wizard</span>
 
 [![CI](https://github.com/anulum/director-ai/actions/workflows/ci.yml/badge.svg)](https://github.com/anulum/director-ai/actions/workflows/ci.yml)
 [![Pre-commit](https://github.com/anulum/director-ai/actions/workflows/pre-commit.yml/badge.svg)](https://github.com/anulum/director-ai/actions/workflows/pre-commit.yml)
@@ -145,7 +145,7 @@ AGPL-3.0 for open source / research. [Commercial licensing](licensing.md) availa
 [GitHub Discussions](https://github.com/anulum/director-ai/discussions) |
 [www.anulum.li](https://www.anulum.li)
 
-*Maintained by [Miroslav Šotek](https://orcid.org/0009-0009-3560-0851) at [Anulum](https://www.anulum.li). Current release: v3.15.0.*
+*Maintained by [Miroslav Šotek](https://orcid.org/0009-0009-3560-0851) at [Anulum](https://www.anulum.li). Current release: v3.15.1.*
 
 <p align="center">
   <a href="https://www.anulum.li">
diff --git a/docs/BUILD_JUDGE_DATASET.md b/docs/BUILD_JUDGE_DATASET.md
@@ -1,6 +1,6 @@
 # Director-AI Judge Dataset Builder
 
-> **Module**: `training/build_judge_dataset.py` | **Version**: 3.15.0 | **License**: GNU AGPL v3
+> **Module**: `training/build_judge_dataset.py` | **Version**: 3.15.1 | **License**: GNU AGPL v3
 >
 > © Concepts 1996–2026 Miroslav Šotek. All rights reserved.
 > © Code 2020–2026 Miroslav Šotek. All rights reserved.
diff --git a/docs/DATA_PIPELINE.md b/docs/DATA_PIPELINE.md
@@ -1,6 +1,6 @@
 # Director-AI Training Data Pipeline
 
-> **Module**: `training/data_pipeline.py` | **Version**: 3.15.0 | **License**: GNU AGPL v3
+> **Module**: `training/data_pipeline.py` | **Version**: 3.15.1 | **License**: GNU AGPL v3
 >
 > © Concepts 1996–2026 Miroslav Šotek. All rights reserved.
 > © Code 2020–2026 Miroslav Šotek. All rights reserved.
diff --git a/docs/E2E_EVAL.md b/docs/E2E_EVAL.md
@@ -1,6 +1,6 @@
 # Director-AI End-to-End Guardrail Benchmark
 
-> **Module**: `benchmarks/e2e_eval.py` | **Version**: 3.15.0 | **License**: GNU AGPL v3
+> **Module**: `benchmarks/e2e_eval.py` | **Version**: 3.15.1 | **License**: GNU AGPL v3
 >
 > © Concepts 1996–2026 Miroslav Šotek. All rights reserved.
 > © Code 2020–2026 Miroslav Šotek. All rights reserved.
diff --git a/docs/JUDGE_BENCHMARK.md b/docs/JUDGE_BENCHMARK.md
@@ -1,6 +1,6 @@
 # Director-AI Local Judge Benchmark
 
-> **Module**: `benchmarks/run_judge_benchmark.py` | **Version**: 3.15.0 | **License**: GNU AGPL v3
+> **Module**: `benchmarks/run_judge_benchmark.py` | **Version**: 3.15.1 | **License**: GNU AGPL v3
 >
 > © Concepts 1996–2026 Miroslav Šotek. All rights reserved.
 > © Code 2020–2026 Miroslav Šotek. All rights reserved.
diff --git a/docs/PRODUCTION_CHECKLIST.md b/docs/PRODUCTION_CHECKLIST.md
@@ -1,6 +1,6 @@
 # Production Deployment Checklist
 
-> **Module**: Director-AI | **Version**: 3.15.0 | **License**: GNU AGPL v3
+> **Module**: Director-AI | **Version**: 3.15.1 | **License**: GNU AGPL v3
 >
 > © Concepts 1996–2026 Miroslav Šotek. All rights reserved.
 
diff --git a/docs/RUST_FFI_PIPELINE.md b/docs/RUST_FFI_PIPELINE.md
@@ -1,6 +1,6 @@
 # Director-AI Rust FFI Pipeline (Backfire Kernel)
 
-> **Crate**: `backfire-kernel` | **Version**: 3.15.0 | **License**: GNU AGPL v3
+> **Crate**: `backfire-kernel` | **Version**: 3.15.1 | **License**: GNU AGPL v3
 >
 > © Concepts 1996–2026 Miroslav Šotek. All rights reserved.
 > © Code 2020–2026 Miroslav Šotek. All rights reserved.
diff --git a/docs/_generated/capability_manifest.json b/docs/_generated/capability_manifest.json
@@ -3685,7 +3685,7 @@
     "name": "director-ai",
     "readme": "README.md",
     "requires_python": ">=3.11",
-    "version": "3.15.0"
+    "version": "3.15.1"
   },
   "project_label": "Director-AI",
   "quality_gates": {
diff --git a/docs/_generated/capability_snapshot.md b/docs/_generated/capability_snapshot.md
@@ -5,7 +5,7 @@
 
 | Surface | Current inventory |
 |---|---:|
-| Package version | 3.15.0 |
+| Package version | 3.15.1 |
 | Public API exports | 214 |
 | Python capability source modules | 310 |
 | Python capability classes | 701 |
diff --git a/pyproject.toml b/pyproject.toml
@@ -12,7 +12,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "director-ai"
-version = "3.15.0"
+version = "3.15.1"
 description = "Real-time LLM hallucination guardrail — NLI + RAG fact-checking with token-level streaming halt"
 authors = [
     {name = "Miroslav Šotek", email = "protoscience@anulum.li"}
diff --git a/requirements/docker-gpu-export.txt b/requirements/docker-gpu-export.txt
@@ -1,6 +1,15 @@
+ml_dtypes==0.5.4 \
+    --hash=sha256:19b9a53598f21e453ea2fbda8aa783c20faff8e1eeb0d7ab899309a0053f1483 \
+    --hash=sha256:9ad459e99793fa6e13bd5b7e6792c8f9190b4e5a1b45c63aba14a4d0a7f1d5ff
 onnx==1.21.0 \
     --hash=sha256:10c3185a232089335581fabb98fba4e86d3e8246b8140f2e406082438100ebda \
     --hash=sha256:a9261bd580fb8548c9c37b3c6750387eb8f21ea43c63880d37b2c622e1684285
+onnx_ir==0.2.1 \
+    --hash=sha256:c7285da889312f91882de2092e298a9eeeefbfc1d1951c49d983992967eb09a7
 onnxruntime==1.26.0 \
     --hash=sha256:11a8df4dcfe9ad5ff0bd71a7571dbed019fabc7594676c89fe8b86ea029c246f \
     --hash=sha256:9b6dd70599005bd1bf29779f04a91978b92b5e719c11a20068a8f8e535f725b6
+onnxscript==0.7.0 \
+    --hash=sha256:5b356907d4501e9919f8599c91d8da967406a37b1fac2b40caa55a49acf242ea
+protobuf==7.34.1 \
+    --hash=sha256:bb3812cd53aefea2b028ef42bd780f5b96407247f20c6ef7c679807e9d188f11
diff --git a/requirements/heavy_optional_dependency_policy.toml b/requirements/heavy_optional_dependency_policy.toml
@@ -74,18 +74,42 @@ extra = "ui"
 risk = "web-ui"
 controls = ["upper-bound", "uv-lock", "runtime-isolation", "audit"]
 
+[[export_only]]
+name = "ml_dtypes"
+requirements = "requirements/docker-gpu-export.txt"
+risk = "onnx-export-numeric-types"
+controls = ["hash-pin", "isolated-build-stage"]
+
 [[export_only]]
 name = "onnx"
 requirements = "requirements/docker-gpu-export.txt"
 risk = "model-graph-serialization"
 controls = ["hash-pin", "isolated-build-stage"]
 
+[[export_only]]
+name = "onnx_ir"
+requirements = "requirements/docker-gpu-export.txt"
+risk = "onnx-intermediate-representation"
+controls = ["hash-pin", "isolated-build-stage"]
+
 [[export_only]]
 name = "onnxruntime"
 requirements = "requirements/docker-gpu-export.txt"
 risk = "native-runtime"
 controls = ["hash-pin", "isolated-build-stage"]
 
+[[export_only]]
+name = "onnxscript"
+requirements = "requirements/docker-gpu-export.txt"
+risk = "pytorch-onnx-exporter"
+controls = ["hash-pin", "isolated-build-stage"]
+
+[[export_only]]
+name = "protobuf"
+requirements = "requirements/docker-gpu-export.txt"
+risk = "onnx-schema-serialization"
+controls = ["hash-pin", "isolated-build-stage"]
+
 [[external_runtime]]
 name = "rclpy"
 source = "ROS 2 distribution"
diff --git a/src/director_ai/__init__.py b/src/director_ai/__init__.py
@@ -17,7 +17,7 @@
     from director_ai import guard, score, CoherenceScorer
 """
 
-__version__ = "3.15.0"
+__version__ = "3.15.1"
 
 # Symbol → (module_path, attribute_name)
 # Module paths are relative to director_ai package.
diff --git a/src/director_ai/core/scoring/_nli_export.py b/src/director_ai/core/scoring/_nli_export.py
@@ -12,6 +12,7 @@
 
 from __future__ import annotations
 
+import inspect
 import logging
 import os
 from collections.abc import Callable
@@ -224,16 +225,22 @@ def forward(self, *inputs: Any) -> Any:
     dynamic_axes["logits"] = {0: "batch"}
     model_file = output_path / "model.onnx"
 
+    export_kwargs: dict[str, Any] = {
+        "input_names": input_names,
+        "output_names": ["logits"],
+        "dynamic_axes": dynamic_axes,
+        "opset_version": 17,
+        "do_constant_folding": True,
+    }
+    if "dynamo" in inspect.signature(torch.onnx.export).parameters:
+        export_kwargs["dynamo"] = False
+
     with torch.no_grad():
         torch.onnx.export(
             _SequenceClassifierOnnxWrapper(model, input_names),
             tuple(encoded[name] for name in input_names),
             str(model_file),
-            input_names=input_names,
-            output_names=["logits"],
-            dynamic_axes=dynamic_axes,
-            opset_version=17,
-            do_constant_folding=True,
+            **export_kwargs,
         )
 
     model.config.save_pretrained(output_path)
diff --git a/tests/test_consumer_api.py b/tests/test_consumer_api.py
@@ -27,7 +27,7 @@
 @pytest.mark.consumer
 class TestVersion:
     def test_version_string(self):
-        assert director_ai.__version__ == "3.15.0"
+        assert director_ai.__version__ == "3.15.1"
 
     def test_all_exports_present(self):
         for name in [
diff --git a/tests/test_cost_and_attribution.py b/tests/test_cost_and_attribution.py
@@ -284,4 +284,4 @@ def test_import_from_package(self):
     def test_version_bumped(self):
         from director_ai import __version__
 
-        assert __version__ == "3.15.0"
+        assert __version__ == "3.15.1"
diff --git a/tests/test_onnx_artefact_guidance.py b/tests/test_onnx_artefact_guidance.py
@@ -47,6 +47,7 @@ def test_export_wheel_file_is_pinned() -> None:
 
     assert "onnx==" in text
     assert "onnxruntime==" in text
+    assert "onnxscript==" in text
     assert "--hash=sha256:" in text
 
 
diff --git a/tests/test_onnx_backend.py b/tests/test_onnx_backend.py
@@ -148,6 +148,84 @@ def test_export_uses_torch_onnx_not_legacy_exporter(self, tmp_path, monkeypatch)
         _, _, exported_path = fake_torch.onnx.export.call_args.args
         assert exported_path.endswith("model.onnx")
 
+    def test_export_disables_dynamo_when_dynamic_axes_are_used(
+        self,
+        tmp_path,
+        monkeypatch,
+    ):
+        fake_torch = ModuleType("torch")
+        fake_torch.long = object()
+        fake_torch.no_grad = MagicMock(return_value=nullcontext())
+        fake_torch.ones = MagicMock(
+            side_effect=lambda shape, dtype=None: {"shape": shape, "dtype": dtype},
+        )
+        fake_torch.nn = SimpleNamespace(Module=object)
+        captured: dict[str, object] = {}
+
+        def export(
+            model,
+            args,
+            f,
+            *,
+            input_names,
+            output_names,
+            dynamic_axes,
+            opset_version,
+            do_constant_folding,
+            dynamo,
+        ):
+            captured.update(
+                {
+                    "model": model,
+                    "args": args,
+                    "f": f,
+                    "input_names": input_names,
+                    "output_names": output_names,
+                    "dynamic_axes": dynamic_axes,
+                    "opset_version": opset_version,
+                    "do_constant_folding": do_constant_folding,
+                    "dynamo": dynamo,
+                },
+            )
+
+        fake_torch.onnx = SimpleNamespace(export=export)
+        fake_transformers = ModuleType("transformers")
+        fake_transformers.AutoTokenizer = SimpleNamespace(from_pretrained=MagicMock())
+        fake_transformers.AutoModelForSequenceClassification = SimpleNamespace(
+            from_pretrained=MagicMock()
+        )
+        monkeypatch.setitem(sys.modules, "torch", fake_torch)
+        monkeypatch.setitem(sys.modules, "transformers", fake_transformers)
+
+        model = MagicMock()
+        model.config.save_pretrained = MagicMock()
+        tokenizer = MagicMock()
+        tokenizer.return_value = {
+            "input_ids": fake_torch.ones((1, 4), dtype=fake_torch.long),
+            "attention_mask": fake_torch.ones((1, 4), dtype=fake_torch.long),
+        }
+
+        with (
+            patch.object(
+                fake_transformers.AutoTokenizer,
+                "from_pretrained",
+                return_value=tokenizer,
+            ),
+            patch.object(
+                fake_transformers.AutoModelForSequenceClassification,
+                "from_pretrained",
+                return_value=model,
+            ),
+        ):
+            export_onnx(
+                model_name="test/model",
+                output_dir=str(tmp_path),
+                revision="abc123",
+            )
+
+        assert captured["dynamo"] is False
+        assert captured["dynamic_axes"]["input_ids"] == {0: "batch", 1: "sequence"}
+
 
 # ── ONNX runtime ─────────────────────────────────────────────────
 
diff --git a/tests/test_tools/test_capability_manifest.py b/tests/test_tools/test_capability_manifest.py
@@ -99,7 +99,7 @@ def test_markdown_snapshot_is_readme_safe() -> None:
     snapshot = tool.render_markdown_snapshot(manifest)
 
     assert "do not edit counts by hand" in snapshot
-    assert "| Package version | 3.15.0 |" in snapshot
+    assert "| Package version | 3.15.1 |" in snapshot
     assert "Evidence boundary" in snapshot
 
 
diff --git a/tests/test_v330_hardening.py b/tests/test_v330_hardening.py
@@ -31,7 +31,7 @@ class TestVersionSync:
     def test_version_is_current_release(self):
         from director_ai import __version__
 
-        assert __version__ == "3.15.0"
+        assert __version__ == "3.15.1"
 
 
 # â”€â”€ Item 3: gRPC proto stubs â”€â”€â”€â”€â”€â”€â”€â”€â”€â”€â”€â”€â”€â”€â”€â”€â”€â”€â”€â”€â”€â”€â”€â”€â”€â”€â”€â”€â”€â”€â”€â”€â”€â”€â”€â”€â”€â”€â”€â”€
diff --git a/uv.lock b/uv.lock

Original file line number	Diff line number	Diff line change
`@@ -15,7 +15,7 @@ curl https://your-server:8080/v1/source`
`15`	`15`	```json
`16`	`16`	`{`
`17`	`17`	`"license": "AGPL-3.0-or-later",`
`18`		`- "version": "3.15.0",`
	`18`	`+ "version": "3.15.1",`
`19`	`19`	`"repository_url": "https://github.com/anulum/director-ai",`
`20`	`20`	`"instructions": "git clone https://github.com/anulum/director-ai",`
`21`	`21`	`"agpl_section": "13"`
Original file line number	Diff line number	Diff line change
`@@ -1,6 +1,6 @@`
`1`	`1`	`# Director-AI Judge Dataset Builder`
`2`	`2`
`3`		-> Module: `training/build_judge_dataset.py` \| Version: 3.15.0 \| License: GNU AGPL v3
	`3`	+> Module: `training/build_judge_dataset.py` \| Version: 3.15.1 \| License: GNU AGPL v3
`4`	`4`	`>`
`5`	`5`	`> © Concepts 1996–2026 Miroslav Šotek. All rights reserved.`
`6`	`6`	`> © Code 2020–2026 Miroslav Šotek. All rights reserved.`
Original file line number	Diff line number	Diff line change
`@@ -1,6 +1,6 @@`
`1`	`1`	`# Director-AI Training Data Pipeline`
`2`	`2`
`3`		-> Module: `training/data_pipeline.py` \| Version: 3.15.0 \| License: GNU AGPL v3
	`3`	+> Module: `training/data_pipeline.py` \| Version: 3.15.1 \| License: GNU AGPL v3
`4`	`4`	`>`
`5`	`5`	`> © Concepts 1996–2026 Miroslav Šotek. All rights reserved.`
`6`	`6`	`> © Code 2020–2026 Miroslav Šotek. All rights reserved.`
Original file line number	Diff line number	Diff line change
`@@ -1,6 +1,6 @@`
`1`	`1`	`# Director-AI End-to-End Guardrail Benchmark`
`2`	`2`
`3`		-> Module: `benchmarks/e2e_eval.py` \| Version: 3.15.0 \| License: GNU AGPL v3
	`3`	+> Module: `benchmarks/e2e_eval.py` \| Version: 3.15.1 \| License: GNU AGPL v3
`4`	`4`	`>`
`5`	`5`	`> © Concepts 1996–2026 Miroslav Šotek. All rights reserved.`
`6`	`6`	`> © Code 2020–2026 Miroslav Šotek. All rights reserved.`