Complexity-ML · Complexity-ML · Sep 6, 2026 · Sep 3, 2026 · Sep 5, 2026 · Sep 5, 2026
diff --git a/.github/workflows/onnx-release.yml b/.github/workflows/onnx-release.yml
@@ -12,11 +12,21 @@ on:
         required: false
         default: true
         type: boolean
+      evidence_run_id:
+        description: "Workflow run ID containing vision-v8-quantized-release-inputs"
+        required: false
+        type: string
+      runner:
+        description: "Runner label with CUDA support for FP16 ONNX gates"
+        required: false
+        default: "self-hosted"
+        type: string
   push:
     tags:
       - "onnx-v8-*"
 
 permissions:
+  actions: read
   contents: write
 
 concurrency:
@@ -25,7 +35,7 @@ concurrency:
 
 jobs:
   publish:
-    runs-on: ubuntu-latest
+    runs-on: ${{ inputs.runner || 'self-hosted' }}
     # Two 640px exports plus 50-seed parity gates on both branches.
     timeout-minutes: 60
     steps:
@@ -55,10 +65,26 @@ jobs:
             --index-url https://download.pytorch.org/whl/cpu
           python -m pip install \
             "onnx==${{ steps.pins.outputs.onnx }}" \
-            "onnxruntime==${{ steps.pins.outputs.onnxruntime }}"
+            "onnxruntime-gpu==${{ steps.pins.outputs.onnxruntime }}" \
+            psutil
+
+      - name: Download quantized release evidence
+        env:
+          GH_TOKEN: ${{ github.token }}
+          EVIDENCE_RUN_ID: ${{ github.event.inputs.evidence_run_id || vars.ONNX_RELEASE_EVIDENCE_RUN_ID }}
+        run: |
+          test -n "$EVIDENCE_RUN_ID"
+          mkdir -p artifacts/vision_v8_quantized_eval
+          gh run download "$EVIDENCE_RUN_ID" \
+            --name vision-v8-quantized-release-inputs \
+            --dir artifacts/vision_v8_quantized_eval
+          test -f artifacts/vision_v8_quantized_eval/calibration.json
+          test -d artifacts/vision_v8_quantized_eval/calibration_images
+          test -f artifacts/vision_v8_quantized_eval/accuracy.json
+          test -f artifacts/vision_v8_quantized_eval/accuracy.md
 
       - name: Test the release tooling
-        run: pytest -q tests/test_onnx_release.py
+        run: pytest -q tests/test_onnx_release.py tests/test_onnx_quantization_benchmark.py
 
       - name: Build, gate and verify the release
         env:
@@ -104,9 +130,5 @@ jobs:
                  --notes-file dist/onnx/RELEASE_NOTES.md \
                  $DRAFT
           gh release upload "$TAG" \
-            dist/onnx/manifest.json \
-            dist/onnx/tr_hash_v8_o2m.onnx \
-            dist/onnx/tr_hash_v8_o2m.json \
-            dist/onnx/tr_hash_v8_nms_free.onnx \
-            dist/onnx/tr_hash_v8_nms_free.json \
+            dist/onnx/* \
             --clobber
diff --git a/.github/workflows/vision-v8-coco-accuracy.yml b/.github/workflows/vision-v8-coco-accuracy.yml
@@ -47,6 +47,33 @@ on:
       onnx_nms_free_metadata:
         description: "Local path to the NMS-free Vision v8 ONNX metadata sidecar on the runner"
         required: false
+      onnx_o2m_fp16_model:
+        description: "Local path to the FP16 O2M Vision v8 ONNX model on the runner"
+        required: false
+      onnx_o2m_fp16_metadata:
+        description: "Local path to the FP16 O2M Vision v8 ONNX metadata sidecar"
+        required: false
+      onnx_nms_free_fp16_model:
+        description: "Local path to the FP16 NMS-free Vision v8 ONNX model"
+        required: false
+      onnx_nms_free_fp16_metadata:
+        description: "Local path to the FP16 NMS-free Vision v8 ONNX metadata sidecar"
+        required: false
+      onnx_o2m_int8_model:
+        description: "Local path to the INT8 O2M Vision v8 ONNX model on the runner"
+        required: false
+      onnx_o2m_int8_metadata:
+        description: "Local path to the INT8 O2M Vision v8 ONNX metadata sidecar"
+        required: false
+      onnx_nms_free_int8_model:
+        description: "Local path to the INT8 NMS-free Vision v8 ONNX model"
+        required: false
+      onnx_nms_free_int8_metadata:
+        description: "Local path to the INT8 NMS-free Vision v8 ONNX metadata sidecar"
+        required: false
+      calibration_manifest:
+        description: "Local path to the quantization calibration manifest"
+        required: false
       annotations:
         description: "Local path to instances_val2017.json on the runner"
         required: true
@@ -66,7 +93,19 @@ on:
           - pytorch
           - onnx
       provider:
-        description: "ONNX Runtime provider alias for backend=onnx"
+        description: "Legacy ONNX provider alias; precision-specific provider inputs are used for release evidence"
+        required: false
+        default: "cpu"
+      fp32_provider:
+        description: "ONNX Runtime provider alias for FP32 evaluation"
+        required: true
+        default: "cpu"
+      fp16_provider:
+        description: "ONNX Runtime provider alias for FP16 evaluation"
+        required: true
+        default: "cuda"
+      int8_provider:
+        description: "ONNX Runtime provider alias for INT8 evaluation"
         required: true
         default: "cpu"
       release_tag:
@@ -136,6 +175,8 @@ jobs:
         run: |
           python -m pip install --upgrade pip
           python -m pip install -e ".[detection,export]"
+          ONNXRUNTIME_VERSION=$(python -c "import json; print(json.load(open('docs/onnx/release.json'))['toolchain']['onnxruntime'])")
+          python -m pip install "onnxruntime-gpu==$ONNXRUNTIME_VERSION"
       - name: Run deterministic COCO evaluation twice
         run: |
           if [ "${{ inputs.backend }}" = "pytorch" ]; then
@@ -147,11 +188,11 @@ jobs:
             test -n "${{ inputs.onnx_o2m_metadata }}"
             test -n "${{ inputs.onnx_nms_free_model }}"
             test -n "${{ inputs.onnx_nms_free_metadata }}"
-            python scripts/evaluate_onnx_coco.py --model "${{ inputs.onnx_o2m_model }}" --metadata "${{ inputs.onnx_o2m_metadata }}" --annotations "${{ inputs.annotations }}" --images "${{ inputs.images }}" --output artifacts/vision_v8_coco_eval/run_a_o2m --branch o2m-nms --provider "${{ inputs.provider }}" --seed 0
-            python scripts/evaluate_onnx_coco.py --model "${{ inputs.onnx_nms_free_model }}" --metadata "${{ inputs.onnx_nms_free_metadata }}" --annotations "${{ inputs.annotations }}" --images "${{ inputs.images }}" --output artifacts/vision_v8_coco_eval/run_a_nms_free --branch nms-free --provider "${{ inputs.provider }}" --seed 0
+            python scripts/evaluate_onnx_coco.py --model "${{ inputs.onnx_o2m_model }}" --metadata "${{ inputs.onnx_o2m_metadata }}" --annotations "${{ inputs.annotations }}" --images "${{ inputs.images }}" --output artifacts/vision_v8_coco_eval/run_a_o2m --branch o2m-nms --provider "${{ inputs.fp32_provider }}" --precision fp32 --seed 0
+            python scripts/evaluate_onnx_coco.py --model "${{ inputs.onnx_nms_free_model }}" --metadata "${{ inputs.onnx_nms_free_metadata }}" --annotations "${{ inputs.annotations }}" --images "${{ inputs.images }}" --output artifacts/vision_v8_coco_eval/run_a_nms_free --branch nms-free --provider "${{ inputs.fp32_provider }}" --precision fp32 --seed 0
             python scripts/merge_vision_v8_coco_reports.py artifacts/vision_v8_coco_eval/run_a_o2m/evaluation.json artifacts/vision_v8_coco_eval/run_a_nms_free/evaluation.json --output artifacts/vision_v8_coco_eval/run_a
-            python scripts/evaluate_onnx_coco.py --model "${{ inputs.onnx_o2m_model }}" --metadata "${{ inputs.onnx_o2m_metadata }}" --annotations "${{ inputs.annotations }}" --images "${{ inputs.images }}" --output artifacts/vision_v8_coco_eval/run_b_o2m --branch o2m-nms --provider "${{ inputs.provider }}" --seed 0
-            python scripts/evaluate_onnx_coco.py --model "${{ inputs.onnx_nms_free_model }}" --metadata "${{ inputs.onnx_nms_free_metadata }}" --annotations "${{ inputs.annotations }}" --images "${{ inputs.images }}" --output artifacts/vision_v8_coco_eval/run_b_nms_free --branch nms-free --provider "${{ inputs.provider }}" --seed 0
+            python scripts/evaluate_onnx_coco.py --model "${{ inputs.onnx_o2m_model }}" --metadata "${{ inputs.onnx_o2m_metadata }}" --annotations "${{ inputs.annotations }}" --images "${{ inputs.images }}" --output artifacts/vision_v8_coco_eval/run_b_o2m --branch o2m-nms --provider "${{ inputs.fp32_provider }}" --precision fp32 --seed 0
+            python scripts/evaluate_onnx_coco.py --model "${{ inputs.onnx_nms_free_model }}" --metadata "${{ inputs.onnx_nms_free_metadata }}" --annotations "${{ inputs.annotations }}" --images "${{ inputs.images }}" --output artifacts/vision_v8_coco_eval/run_b_nms_free --branch nms-free --provider "${{ inputs.fp32_provider }}" --precision fp32 --seed 0
             python scripts/merge_vision_v8_coco_reports.py artifacts/vision_v8_coco_eval/run_b_o2m/evaluation.json artifacts/vision_v8_coco_eval/run_b_nms_free/evaluation.json --output artifacts/vision_v8_coco_eval/run_b
           fi
       - name: Gate full COCO report
@@ -160,6 +201,35 @@ jobs:
           artifacts/vision_v8_coco_eval/run_a/evaluation.json
           --repeat-report artifacts/vision_v8_coco_eval/run_b/evaluation.json
           --config configs/vision_v8_coco_accuracy_gate.json
+      - name: Build quantized release evidence artifact
+        if: inputs.backend == 'onnx' && inputs.calibration_manifest != ''
+        env:
+          CALIBRATION_MANIFEST: ${{ inputs.calibration_manifest }}
+        run: |
+          test -n "${{ inputs.onnx_o2m_fp16_model }}"
+          test -n "${{ inputs.onnx_o2m_fp16_metadata }}"
+          test -n "${{ inputs.onnx_nms_free_fp16_model }}"
+          test -n "${{ inputs.onnx_nms_free_fp16_metadata }}"
+          test -n "${{ inputs.onnx_o2m_int8_model }}"
+          test -n "${{ inputs.onnx_o2m_int8_metadata }}"
+          test -n "${{ inputs.onnx_nms_free_int8_model }}"
+          test -n "${{ inputs.onnx_nms_free_int8_metadata }}"
+          mkdir -p artifacts/vision_v8_quantized_eval
+          python -c "import os; from pathlib import Path; from scripts.check_onnx_quantized_artifacts import materialize_calibration_manifest_images; materialize_calibration_manifest_images(Path(os.environ['CALIBRATION_MANIFEST']), artifact_root=Path('artifacts/vision_v8_quantized_eval'), output_manifest_path=Path('artifacts/vision_v8_quantized_eval/calibration.json'))"
+          python scripts/evaluate_onnx_coco.py --model "${{ inputs.onnx_o2m_fp16_model }}" --metadata "${{ inputs.onnx_o2m_fp16_metadata }}" --annotations "${{ inputs.annotations }}" --images "${{ inputs.images }}" --output artifacts/vision_v8_quantized_eval/o2m_fp16 --branch o2m-nms --provider "${{ inputs.fp16_provider }}" --precision fp16 --seed 0
+          python scripts/evaluate_onnx_coco.py --model "${{ inputs.onnx_nms_free_fp16_model }}" --metadata "${{ inputs.onnx_nms_free_fp16_metadata }}" --annotations "${{ inputs.annotations }}" --images "${{ inputs.images }}" --output artifacts/vision_v8_quantized_eval/nms_free_fp16 --branch nms-free --provider "${{ inputs.fp16_provider }}" --precision fp16 --seed 0
+          python scripts/evaluate_onnx_coco.py --model "${{ inputs.onnx_o2m_int8_model }}" --metadata "${{ inputs.onnx_o2m_int8_metadata }}" --annotations "${{ inputs.annotations }}" --images "${{ inputs.images }}" --output artifacts/vision_v8_quantized_eval/o2m_int8 --branch o2m-nms --provider "${{ inputs.int8_provider }}" --precision int8 --seed 0
+          python scripts/evaluate_onnx_coco.py --model "${{ inputs.onnx_nms_free_int8_model }}" --metadata "${{ inputs.onnx_nms_free_int8_metadata }}" --annotations "${{ inputs.annotations }}" --images "${{ inputs.images }}" --output artifacts/vision_v8_quantized_eval/nms_free_int8 --branch nms-free --provider "${{ inputs.int8_provider }}" --precision int8 --seed 0
+          python scripts/merge_vision_v8_coco_reports.py \
+            artifacts/vision_v8_coco_eval/run_a_o2m/evaluation.json \
+            artifacts/vision_v8_coco_eval/run_a_nms_free/evaluation.json \
+            artifacts/vision_v8_quantized_eval/o2m_fp16/evaluation.json \
+            artifacts/vision_v8_quantized_eval/nms_free_fp16/evaluation.json \
+            artifacts/vision_v8_quantized_eval/o2m_int8/evaluation.json \
+            artifacts/vision_v8_quantized_eval/nms_free_int8/evaluation.json \
+            --output artifacts/vision_v8_quantized_eval
+          cp artifacts/vision_v8_quantized_eval/evaluation.json artifacts/vision_v8_quantized_eval/accuracy.json
+          cp artifacts/vision_v8_quantized_eval/evaluation.md artifacts/vision_v8_quantized_eval/accuracy.md
       - uses: actions/upload-artifact@v4
         with:
           name: vision-v8-coco-accuracy-report
@@ -168,6 +238,15 @@ jobs:
             artifacts/vision_v8_coco_eval/run_a/evaluation.md
             artifacts/vision_v8_coco_eval/run_b/evaluation.json
             artifacts/vision_v8_coco_eval/run_b/evaluation.md
+      - uses: actions/upload-artifact@v4
+        if: inputs.backend == 'onnx' && inputs.calibration_manifest != ''
+        with:
+          name: vision-v8-quantized-release-inputs
+          path: |
+            artifacts/vision_v8_quantized_eval/calibration.json
+            artifacts/vision_v8_quantized_eval/calibration_images/**
+            artifacts/vision_v8_quantized_eval/accuracy.json
+            artifacts/vision_v8_quantized_eval/accuracy.md
       - name: Upload gated reports to GitHub Release
         if: inputs.release_tag != ''
         env:

diff --git a/complexity/deploy/onnx_detector/pipeline.py b/complexity/deploy/onnx_detector/pipeline.py
@@ -39,9 +39,19 @@ def from_files(
         model_path: str | Path,
         metadata_path: str | Path,
         providers: Sequence[str] = ("CPUExecutionProvider",),
+        *,
+        warmup_runs: int = 1,
+        intra_op_num_threads: int | None = None,
+        inter_op_num_threads: int | None = None,
     ) -> "OnnxDetectorPipeline":
         metadata = load_metadata(metadata_path)
-        session = cls.create_session(model_path, providers).open()
+        session = cls.create_session(
+            model_path,
+            providers,
+            warmup_runs=warmup_runs,
+            intra_op_num_threads=intra_op_num_threads,
+            inter_op_num_threads=inter_op_num_threads,
+        ).open()
         validate_output_shape(metadata, session._require_session().get_outputs()[0].shape)
         if session.config.warmup_runs:
             session.warmup((1, 3, metadata.image_size, metadata.image_size))
@@ -74,13 +84,11 @@ def postprocess_single_image(
         scores: np.ndarray,
         classes: np.ndarray,
     ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
-        filtered_boxes, filtered_scores, filtered_classes, _ = (
-            postprocess.filter_by_confidence(
-                boxes,
-                scores,
-                classes,
-                self.metadata.confidence_threshold,
-            )
+        filtered_boxes, filtered_scores, filtered_classes, _ = postprocess.filter_by_confidence(
+            boxes,
+            scores,
+            classes,
+            self.metadata.confidence_threshold,
         )
         if self.metadata.branch == "o2m":
             keep = postprocess.class_aware_nms(

diff --git a/configs/vision_v8_quantization_calibration.example.json b/configs/vision_v8_quantization_calibration.example.json
@@ -0,0 +1,19 @@
+{
+  "schema_version": 1,
+  "dataset": {
+    "name": "coco-2017-train-calibration",
+    "split": "train2017-calibration",
+    "image_ids_sha256": "replace-with-calibration-image-id-manifest-sha256",
+    "annotations_sha256": "replace-with-annotations-sha256",
+    "disjoint_from": "coco-2017-val2017"
+  },
+  "quantization": {
+    "calibration_method": "minmax",
+    "per_channel": true,
+    "symmetric_activations": false,
+    "symmetric_weights": true,
+    "activation_type": "quint8",
+    "weight_type": "qint8",
+    "batch_size": 1
+  }
+}
diff --git a/configs/vision_v8_quantization_thresholds.json b/configs/vision_v8_quantization_thresholds.json
@@ -0,0 +1,49 @@
+{
+    "schema_version": 1,
+    "release_policy": {
+        "required_precisions": ["fp32", "fp16", "int8"],
+        "require_artifact_bindings": true,
+        "provider_by_precision": {
+            "fp32": "CPUExecutionProvider",
+            "fp16": "CUDAExecutionProvider",
+            "int8": "CPUExecutionProvider"
+        },
+        "partial_release": "block",
+        "optional_provider_precisions": []
+    },
+    "precisions": {
+        "fp16": {
+            "max_raw_logit_abs_error": 0.01,
+            "max_decoded_box_px_error": 1.0,
+            "max_score_abs_error": 0.01,
+            "max_map50_95_drop": 0.005,
+            "max_map50_drop": 0.01,
+            "unexpected_fp32_nodes": "fail"
+        },
+        "int8": {
+            "max_raw_logit_abs_error": 0.12,
+            "max_decoded_box_px_error": 4.0,
+            "max_score_abs_error": 0.05,
+            "max_map50_95_drop": 0.02,
+            "max_map50_drop": 0.03,
+            "unexpected_fp32_nodes": "allow"
+        }
+    },
+    "providers": {
+        "CPUExecutionProvider": ["fp32", "int8"],
+        "CUDAExecutionProvider": ["fp32", "fp16"],
+        "TensorrtExecutionProvider": ["fp32", "fp16", "int8"]
+    },
+    "benchmark": {
+        "warmup_iterations": 25,
+        "measured_iterations": 100,
+        "report": [
+            "median_ms",
+            "mean_ms",
+            "stddev_ms",
+            "p95_ms",
+            "throughput_images_per_second",
+            "peak_memory_mb"
+        ]
+    }
+}
diff --git a/docs/index.md b/docs/index.md
@@ -92,6 +92,7 @@ GQA + TR-MoE architecture.
 - [TR-Hash object detection and serving](tr-hash-object-detection.md)
 - [TR-Hash Vision ONNX deployment](onnx_deploy.md)
 - [Vision v8 COCO accuracy gates](vision-v8-coco-accuracy-gates.md)
+- [Vision v8 ONNX quantization](vision-v8-onnx-quantization.md)
 - [Detector specialization and ablations](TR_HASH_DETECTOR_SPECIALIZATION.md)
 - [TR-Hash sensor fusion](tr_hash_sensor_fusion.md)
 - [Vision dependency stack](vision-dependency-stack.md)

diff --git a/docs/onnx/release.json b/docs/onnx/release.json
@@ -3,10 +3,23 @@
   "checkpoint_revision": "f3b3e659612e543ca9ff91892c0662d38dc1a1d6",
   "opset": 17,
   "parity_num_tests": 50,
+  "quantization": {
+    "enabled_precisions": ["fp16", "int8"],
+    "calibration_manifest": "artifacts/vision_v8_quantized_eval/calibration.json",
+    "thresholds": "configs/vision_v8_quantization_thresholds.json",
+    "accuracy_report": "artifacts/vision_v8_quantized_eval/accuracy.json",
+    "accuracy_markdown": "artifacts/vision_v8_quantized_eval/accuracy.md",
+    "fp32_op_allowlist": [],
+    "provider_gates": [
+      {"provider": "CPUExecutionProvider", "precision": "fp32"},
+      {"provider": "CUDAExecutionProvider", "precision": "fp16"},
+      {"provider": "CPUExecutionProvider", "precision": "int8"}
+    ]
+  },
   "toolchain": {
     "torch": "2.13.0",
     "onnx": "1.21.0",
-    "onnxruntime": "1.24.4"
+    "onnxruntime": "1.23.2"
   },
   "branches": [
     {