Add CUA-S1-FORMS FP16 Core ML conversion and verified artifacts

Add the 706,048-parameter CUA-S1-FORMS decision scorer as a portable FP16 Core ML package and compiled FluidAudio bundle, with encoding helper, pinned provenance, licenses, checksums, and conversion reports.

Validation uses the complete pinned 196-row upstream demo, not a new synthetic dataset. PyTorch, Core ML ALL, and Core ML CPU_AND_NE select all 196 labeled options correctly. Maximum probability differences are 0.003099 and 0.002336 respectively, below the preselected 0.005 tolerance. The unmodified Cua evaluator reports zero wrong actions, wrong targets, or unsafe actions on these saved predictions. Seventeen Python regression tests pass.

FluidAudio's Swift runtime independently matches all 196 decisions against the PyTorch probabilities (maximum error 0.002336), including local package loading, shared compiled-model cache loading, UTF-8 encoding, and concurrent/reordered-option checks. The Swift library builds; local XCTest execution is unavailable because this Mac has Command Line Tools without the XCTest framework.

The package is 1,511,163 bytes. Timings and ANE/CPU placement are exploratory measurements on the documented M5 Pro, not general GUI-automation or mobile-device benchmarks. The caller supplies document entities and UI descriptions and owns action validation and execution.

Related code branches:
- https://github.com/FluidInference/mobius/tree/codex/cua-s1-forms
- https://github.com/FluidInference/FluidAudio/tree/codex/cua-s1-forms

Merge this model PR before enabling FluidAudio's default download path. This PR does not merge or release the model automatically.

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2025 Cua AI, Inc.
+
+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.

NOTICES.md

@@ -0,0 +1,18 @@
+# Distribution notices
+
+This repository distributes FP16 Core ML derivatives of the MIT-licensed
+CUA-S1-FORMS checkpoint published by Cua at
+https://huggingface.co/cua-ai/cua-s1-forms/tree/f54adbf447f4ca6ec259f529ee3f2e3e09f8cc71.
+The Cua MIT license is included as LICENSE.
+
+The model implementation includes byte-collation and attention primitives adapted
+from Minimal Labs' MIT-licensed jevlike project. Its copyright and license are
+preserved in UPSTREAM-THIRD-PARTY-NOTICES.md. That file is an unmodified notice
+from the source distribution; its statement that the source package does not
+bundle model weights describes that upstream package. This derived distribution
+does include the converted weights listed above.
+
+The upstream demo dataset, training datasets, source safetensors checkpoint,
+Python environment, and third-party dependencies are not bundled. The local
+preprocessing helper is from the FluidInference Mobius conversion toolkit.
+Model/data revisions and source download hashes are recorded in assets.lock.json.

README.md

@@ -0,0 +1,172 @@
+---
+license: mit
+library_name: coreml
+pipeline_tag: text-classification
+base_model: cua-ai/cua-s1-forms
+base_model_relation: quantized
+datasets:
+- cua-ai/cua-s1-forms
+tags:
+- coreml
+- apple-silicon
+- computer-use
+- classification
+- cua-s1
+- fp16
+---
+
+# CUA-S1-FORMS — Core ML
+
+FP16 Core ML conversion of [Cua's CUA-S1-FORMS](https://huggingface.co/cua-ai/cua-s1-forms),
+a **706,048-parameter** specialist that selects among supplied form actions.
+The portable package is **1,511,163 bytes (1.51 MB)**. No text generation, KV cache,
+or external tokenizer is required.
+
+The model uses small Transformer encoders over UTF-8 bytes and an attention
+readout. It is a classifier, not an autoregressive LLM. The inherited checkpoint
+configuration contains an unused `hf_model` value; this `tinyx` checkpoint does
+not load Qwen weights.
+
+## Files
+
+| File | Purpose |
+| --- | --- |
+| `cua_s1_forms_fp16_options32.mlpackage/` | Portable model; compile locally or add to Xcode |
+| `cua_s1_forms_fp16_options32.mlmodelc/` | Compiled bundle for FluidAudio's model loader |
+| `preprocessing.py` | Upstream-compatible byte encoding and input validation |
+| `conversion.json` | Architecture, conversion versions, and portable package hashes |
+| `assets.lock.json` | Pinned upstream source, model, and demo hashes |
+| `checksums.json` | SHA-256 of each distributed file except this checksum file |
+| `reports/` | Per-row parity, original Cua metrics, and compute-placement report |
+
+The model targets **iOS 17/macOS 14 or newer**. Runtime validation used an Apple
+silicon Mac. Use the portable package for local compilation on other supported
+systems; iPhone performance and compatibility of the precompiled bundle across
+older OS versions have not been measured.
+
+## Python usage
+
+Install `coremltools==9.0`, `numpy==1.26.4`, and `huggingface_hub` on macOS.
+Download this repository, then run from its directory:
+
+```python
+import coremltools as ct
+from preprocessing import InputLimits, prepare_inputs
+
+model = ct.models.MLModel(
+    "cua_s1_forms_fp16_options32.mlpackage",
+    compute_units=ct.ComputeUnit.CPU_AND_NE,
+)
+options = ["fill E-mail: person@example.com", "check", "click", "skip"]
+inputs = prepare_inputs(
+    'TASK fill the form from the document, then submit\n'
+    'FORM Contact details\nELEMENT Edit "Email address" value=""',
+    options,
+    InputLimits(),
+)
+probabilities = model.predict(inputs)["probabilities"][0, :len(options)]
+print(options[int(probabilities.argmax())], probabilities)
+```
+
+During review, download the model PR's revision (for example `--revision refs/pr/1`)
+with `hf download FluidInference/cua-s1-forms-coreml --local-dir ./cua-coreml`.
+After merge, the default `main` revision contains the artifacts.
+
+## Swift usage
+
+The proposed [FluidAudio integration](https://github.com/FluidInference/FluidAudio/tree/codex/cua-s1-forms)
+provides `CuaS1FormsManager`:
+
+```swift
+import FluidAudio
+import Foundation
+
+let manager = try await CuaS1FormsManager.load(
+    from: URL(fileURLWithPath: "/models/cua_s1_forms_fp16_options32.mlpackage"))
+let decision = try await manager.score(
+    context: "TASK fill the form from the document, then submit\nFORM Contact details\nELEMENT Edit \"Email address\" value=\"\"",
+    options: ["fill E-mail: person@example.com", "check", "click", "skip"])
+print(decision.selectedOption, decision.probabilities)
+```
+
+After the model and Swift PRs land, `try await CuaS1FormsManager.load()` downloads
+and caches the compiled artifact automatically.
+
+## Tensor interface
+
+| Name | Type | Shape |
+| --- | --- | --- |
+| `context_ids` | int32 | `[1, 224]` |
+| `option_ids` | int32 | `[1, 32, 96]` |
+| `option_mask` | int32 | `[1, 32]` |
+| `logits` | float32 output | `[1, 32]` |
+| `probabilities` | float32 output | `[1, 32]` |
+
+Encode UTF-8 bytes plus one, pad with zero, and truncate by bytes at 224 for
+context and 96 per option. Supply a nonempty context and 2–32 nonempty options.
+Set option-mask entries to one for supplied options and zero for padding.
+Padded logits are `-10000`; padded probabilities are zero. Inputs above 32
+options must be rejected or use a separately exported larger-capacity model.
+The example helper rejects overflow rather than dropping choices.
+
+## Conversion verification
+
+On the complete pinned **196-row upstream demo**, PyTorch and Core ML both
+selected **196/196 labeled options correctly**, and matched each other's selected
+option on every row. The unmodified upstream Cua evaluator reports 36 fill,
+4 check, 6 click, and 150 skip decisions, with zero wrong actions, wrong targets,
+or unsafe actions on these saved predictions. It counts `skip` as abstention,
+so its 23.47% coverage corresponds to 46 actionable decisions.
+
+| Check | Core ML `ALL` | Core ML `CPU_AND_NE` |
+| --- | ---: | ---: |
+| Selected options matching PyTorch | 196 / 196 | 196 / 196 |
+| Maximum absolute probability error | 0.003099 | 0.002336 |
+| Warm model-call median | 1.85 ms | 0.90 ms |
+| Warm model-call p95 | 2.49 ms | 0.94 ms |
+
+Measured September 19, 2026 on Apple M5 Pro, 24 GB, macOS 27.0, using
+Python 3.11.11, PyTorch 2.7.0, and coremltools 9.0. Timing is exploratory and
+includes Python call overhead; model loading, encoding, document extraction,
+UI observation, and action execution are excluded. It is not an optimized
+PyTorch/MPS speed comparison.
+
+The conversion gates require 100% selected-option agreement, no accuracy loss,
+maximum absolute probability error ≤ 0.005, finite outputs, normalized live
+probabilities, and zero probability for padding. The FP32 export adapter differs
+from the unmodified PyTorch reference by at most 0.00000113. Six reversed-option
+checks also pass on each Core ML configuration. The checked-in placement report
+counts 149 Neural Engine operations, 24 CPU operations, and zero GPU operations;
+operation counts are not a measurement of time spent on each processor.
+
+These results verify conversion on three demo forms and three PDFs. They do not
+establish generalization, live GUI completion rates, or production safety.
+The original demo is not redistributed here; its exact revision and SHA-256 are
+in the asset lock. No training or larger-corpus evaluation was performed for
+this conversion.
+
+## Application responsibilities
+
+The application must extract document entities, describe UI elements, build
+candidate actions, and validate and order the selected actions. Submission and
+other effects require application authorization. Scores are not calibrated
+confidence guarantees. Text outside the byte limits is truncated, and arbitrary
+new forms and languages require their own evaluation.
+
+## Source, reproduction, and license
+
+The [Mobius conversion toolkit](https://github.com/FluidInference/mobius/tree/codex/cua-s1-forms/models/computer-use/cua-s1-forms/coreml)
+contains the conversion code, lockfile, original reference implementation and
+evaluator, tests, and full reproduction instructions. Adaptations are limited
+to export-compatible masking, a floating-point clamp constant, finite padded
+logits, and disabling the fused PyTorch Transformer fast path during tracing.
+All trained layers and checkpoint tensors are retained; internal compute and
+weights are converted to FP16.
+
+- Model: [`cua-ai/cua-s1-forms` at `f54adbf447f4ca6ec259f529ee3f2e3e09f8cc71`](https://huggingface.co/cua-ai/cua-s1-forms/tree/f54adbf447f4ca6ec259f529ee3f2e3e09f8cc71).
+- Demo: [`cua-ai/cua-s1-forms` at `8273f34778b99ac2e12d9f6e7d57dad99ae20845`](https://huggingface.co/datasets/cua-ai/cua-s1-forms/tree/8273f34778b99ac2e12d9f6e7d57dad99ae20845).
+- Code: [`trycua/cua` at `83f142c4290a0f7d9ed545ae8532858c6e4f8145`](https://github.com/trycua/cua/tree/83f142c4290a0f7d9ed545ae8532858c6e4f8145/libs/cua-s1).
+
+The pinned model and dataset cards declare MIT. See [LICENSE](LICENSE),
+[NOTICES.md](NOTICES.md), and the preserved
+[upstream third-party notices](UPSTREAM-THIRD-PARTY-NOTICES.md).

UPSTREAM-THIRD-PARTY-NOTICES.md

@@ -0,0 +1,46 @@
+# Third-party notices
+
+This Cua-S1 component does not bundle third-party model weights, datasets,
+binaries, or media.
+
+## jevlike
+
+Parts of `python/src/cua_s1/model.py` are adapted from the MIT-licensed
+[`jevlike`](https://github.com/vinnylarouge/jevlike) project at commit
+[`94f5fd1b0b11d52bbdfdf4e0ee6aa96b568f8452`](https://github.com/vinnylarouge/jevlike/commit/94f5fd1b0b11d52bbdfdf4e0ee6aa96b568f8452).
+The adapted material is limited to the byte-collation and small attention-model
+primitives in that file.
+
+Copyright (c) 2026 Minimal Labs
+
+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.
+
+The Python package declares Hatchling as a build-system requirement. Hatchling
+is build tooling and is not bundled in the resulting package; it is distributed
+under its own terms.
+
+The Python package declares its runtime dependencies in `python/pyproject.toml`.
+The component-local `python/uv.lock` records the tested development resolution.
+Those dependencies remain distributed under their own licenses and are not
+copied into this repository.
+
+Future checkpoint or software distributions must update this file with the
+applicable notices for all included third-party code, models, datasets, assets,
+and other materials. A reference to Cua-S1 or `cua-s1-form-v0` does not grant
+rights to third-party material.

assets.lock.json

@@ -0,0 +1,54 @@
+{
+  "source_revision": "83f142c4290a0f7d9ed545ae8532858c6e4f8145",
+  "model_revision": "f54adbf447f4ca6ec259f529ee3f2e3e09f8cc71",
+  "dataset_revision": "8273f34778b99ac2e12d9f6e7d57dad99ae20845",
+  "evaluation_file": "artifacts/demo.jsonl",
+  "evaluation_rows": 196,
+  "files": [
+    {
+      "path": "vendor/cua_s1/__init__.py",
+      "url": "https://raw.githubusercontent.com/trycua/cua/83f142c4290a0f7d9ed545ae8532858c6e4f8145/libs/cua-s1/python/src/cua_s1/__init__.py",
+      "sha256": "68ca526b67bdb95401ddaa2ccc09b0717879d16a9102e2014f4f9018df292ddf"
+    },
+    {
+      "path": "vendor/cua_s1/model.py",
+      "url": "https://raw.githubusercontent.com/trycua/cua/83f142c4290a0f7d9ed545ae8532858c6e4f8145/libs/cua-s1/python/src/cua_s1/model.py",
+      "sha256": "7538cb4c9730d70a0b325050f370f981a9e109a1b182b8c66be3ba0fa2474344"
+    },
+    {
+      "path": "vendor/cua_s1/checkpoint.py",
+      "url": "https://raw.githubusercontent.com/trycua/cua/83f142c4290a0f7d9ed545ae8532858c6e4f8145/libs/cua-s1/python/src/cua_s1/checkpoint.py",
+      "sha256": "6d96ab6e946c88af341947d656ff244a8de5846edfd1f6a410466a97d89b375f"
+    },
+    {
+      "path": "vendor/CUA-LICENSE",
+      "url": "https://raw.githubusercontent.com/trycua/cua/83f142c4290a0f7d9ed545ae8532858c6e4f8145/libs/cua-s1/python/LICENSE",
+      "sha256": "c0779290c1d4783169aa3dbfb55feb505e563ef8a004bbf55298ceffcfbda8d9"
+    },
+    {
+      "path": "artifacts/cua-s1-forms.safetensors",
+      "url": "https://huggingface.co/cua-ai/cua-s1-forms/resolve/f54adbf447f4ca6ec259f529ee3f2e3e09f8cc71/cua-s1-forms.safetensors",
+      "sha256": "05954c1caf51c2fb6c13ea4acbfc88a2e7653dea192252bb51dc89e76a356ddc"
+    },
+    {
+      "path": "artifacts/cua-s1-forms.json",
+      "url": "https://huggingface.co/cua-ai/cua-s1-forms/resolve/f54adbf447f4ca6ec259f529ee3f2e3e09f8cc71/cua-s1-forms.json",
+      "sha256": "62d31e2f9a001a8e9b6f8534c5194d07ebdd3f9d62ef1ac281906622992650ca"
+    },
+    {
+      "path": "artifacts/demo.jsonl",
+      "url": "https://huggingface.co/datasets/cua-ai/cua-s1-forms/resolve/8273f34778b99ac2e12d9f6e7d57dad99ae20845/demo.jsonl",
+      "sha256": "4f43b442e79ba2e2ce731e27e9b8e340c2b5dfcaffc92d8ff564c34f115ff1ca"
+    },
+    {
+      "path": "artifacts/model-card.md",
+      "url": "https://huggingface.co/cua-ai/cua-s1-forms/resolve/f54adbf447f4ca6ec259f529ee3f2e3e09f8cc71/README.md",
+      "sha256": "a52b7930d12d59a9f70fba537f02c1683eb4be2b451ec0fd5f407620c5670a93"
+    },
+    {
+      "path": "artifacts/dataset-card.md",
+      "url": "https://huggingface.co/datasets/cua-ai/cua-s1-forms/resolve/8273f34778b99ac2e12d9f6e7d57dad99ae20845/README.md",
+      "sha256": "e5768e2a4b8fee2622ca73fb522bdbb0caca567a49c93266ef0f0fdef5b0af11"
+    }
+  ]
+}

checksums.json

@@ -0,0 +1,19 @@
+{
+  "LICENSE": "c0779290c1d4783169aa3dbfb55feb505e563ef8a004bbf55298ceffcfbda8d9",
+  "NOTICES.md": "027c72741eaa695e60d7b6cebd3666372c81d443a96b673b72a897cf432ddfd0",
+  "README.md": "4008ae7ef610e44c2d6e019a0fa375b3e1705714d5875a8e0c63aba8541b4c96",
+  "UPSTREAM-THIRD-PARTY-NOTICES.md": "4091e69b45c8cc97e30a066fbbd56148dbef66ab716432c048d2333c9c464213",
+  "assets.lock.json": "8e65ad70af6bb814b571cdcfe828ba4bc339147da2d2211cbeac416163ef18ba",
+  "conversion.json": "ac94239751ceb7e19602eb7276940ff53446b1776d4900405af9be0382fe4a94",
+  "cua_s1_forms_fp16_options32.mlmodelc/analytics/coremldata.bin": "a7a11d7772fb6879778e6223d0f3b16a6b2a44df20a7991bf9bcc24042d05542",
+  "cua_s1_forms_fp16_options32.mlmodelc/coremldata.bin": "2ddf9edc046a1e66b3d3a0cd57a1d9e25c75df5f8d937249dc02d9806617d884",
+  "cua_s1_forms_fp16_options32.mlmodelc/model.mil": "e93676bd4f553ffcf9403ae5049185cb55acdcb63908e39952578db0845e0a12",
+  "cua_s1_forms_fp16_options32.mlmodelc/weights/weight.bin": "4da9259f798e44f5a1b50769ee1916fd3747c4d723dd9997b516c7fe238c7895",
+  "cua_s1_forms_fp16_options32.mlpackage/Data/com.apple.CoreML/model.mlmodel": "70485fc18cbb21785df833cbddddc0b5b59acb00d22394b76e55307e2c135dd0",
+  "cua_s1_forms_fp16_options32.mlpackage/Data/com.apple.CoreML/weights/weight.bin": "4da9259f798e44f5a1b50769ee1916fd3747c4d723dd9997b516c7fe238c7895",
+  "cua_s1_forms_fp16_options32.mlpackage/Manifest.json": "2bc0f5f62337b27fb6b0ecde248f1e3dc269e1ba4b65516aaeede2a60e293dcc",
+  "preprocessing.py": "18c74ede43f95f91cd638c7e9631923a353e3a24d3b522ddb8b18b480f9c2846",
+  "reports/ane-fallback.json": "0a5d25b97545bfb2878531f3e409030ae3448a26c3aa3ba7db29f57d7194b2a6",
+  "reports/upstream-metrics.json": "7d6207e430504dc0baec96d8180cb03f8aadc5a6e016b53c4cf03a80baf169ab",
+  "reports/verification.json": "9a5d22ec4b447f0710c12e03c69d7ed802158bb94e9a3c6b059731730ced9f30"
+}

conversion.json

@@ -0,0 +1,35 @@
+{
+  "model": "cua_s1_forms_fp16_options32.mlpackage",
+  "precision": "float16",
+  "minimum_target": "iOS17/macOS14",
+  "limits": {
+    "context_bytes": 224,
+    "option_bytes": 96,
+    "max_options": 32
+  },
+  "model_config": {
+    "context_tokens": 224,
+    "encoder": "tinyx",
+    "heads": 4,
+    "hf_model": "Qwen/Qwen2.5-0.5B",
+    "layers": 2,
+    "option_tokens": 96,
+    "rank": 128,
+    "width": 128
+  },
+  "parameters": 706048,
+  "assets_lock_sha256": "8e65ad70af6bb814b571cdcfe828ba4bc339147da2d2211cbeac416163ef18ba",
+  "model_revision": "f54adbf447f4ca6ec259f529ee3f2e3e09f8cc71",
+  "source_revision": "83f142c4290a0f7d9ed545ae8532858c6e4f8145",
+  "trace_row": 0,
+  "trace_dataset_revision": "8273f34778b99ac2e12d9f6e7d57dad99ae20845",
+  "export_seconds": 0.5374167499830946,
+  "python": "3.11.11",
+  "torch": "2.7.0",
+  "coremltools": "9.0",
+  "package_files": {
+    "Data/com.apple.CoreML/model.mlmodel": "70485fc18cbb21785df833cbddddc0b5b59acb00d22394b76e55307e2c135dd0",
+    "Data/com.apple.CoreML/weights/weight.bin": "4da9259f798e44f5a1b50769ee1916fd3747c4d723dd9997b516c7fe238c7895",
+    "Manifest.json": "2bc0f5f62337b27fb6b0ecde248f1e3dc269e1ba4b65516aaeede2a60e293dcc"
+  }
+}

cua_s1_forms_fp16_options32.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a7a11d7772fb6879778e6223d0f3b16a6b2a44df20a7991bf9bcc24042d05542
+size 243

cua_s1_forms_fp16_options32.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2ddf9edc046a1e66b3d3a0cd57a1d9e25c75df5f8d937249dc02d9806617d884
+size 909

cua_s1_forms_fp16_options32.mlmodelc/model.mil

@@ -0,0 +1,418 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3600.16.1"}, {"coremlc-version", "3600.25.2"}, {"coremltools-component-torch", "2.7.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})]
+{
+    func main<ios17>(tensor<int32, [1, 224]> context_ids, tensor<int32, [1, 32, 96]> option_ids, tensor<int32, [1, 32]> option_mask) {
+            tensor<int32, []> var_14 = const()[name = tensor<string, []>("op_14"), val = tensor<int32, []>(0)];
+            tensor<bool, [1, 32]> option_mask_1 = not_equal(x = option_mask, y = var_14)[name = tensor<string, []>("option_mask")];
+            tensor<int32, []> var_16 = const()[name = tensor<string, []>("op_16"), val = tensor<int32, []>(0)];
+            tensor<bool, [1, 224]> context_mask = not_equal(x = context_ids, y = var_16)[name = tensor<string, []>("context_mask")];
+            tensor<int32, [2]> var_27_begin_0 = const()[name = tensor<string, []>("op_27_begin_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [2]> var_27_end_0 = const()[name = tensor<string, []>("op_27_end_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<bool, [2]> var_27_end_mask_0 = const()[name = tensor<string, []>("op_27_end_mask_0"), val = tensor<bool, [2]>([true, false])];
+            tensor<int32, [1, 1]> var_27 = slice_by_index(begin = var_27_begin_0, end = var_27_end_0, end_mask = var_27_end_mask_0, x = context_ids)[name = tensor<string, []>("op_27")];
+            tensor<fp16, []> fill_like_0_value_0_to_fp16 = const()[name = tensor<string, []>("fill_like_0_value_0_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1, 1]> fill_like_0_cast_fp16 = fill_like(ref_tensor = var_27, value = fill_like_0_value_0_to_fp16)[name = tensor<string, []>("fill_like_0_cast_fp16")];
+            tensor<int32, [2]> var_43_begin_0 = const()[name = tensor<string, []>("op_43_begin_0"), val = tensor<int32, [2]>([0, 1])];
+            tensor<int32, [2]> var_43_end_0 = const()[name = tensor<string, []>("op_43_end_0"), val = tensor<int32, [2]>([1, 224])];
+            tensor<bool, [2]> var_43_end_mask_0 = const()[name = tensor<string, []>("op_43_end_mask_0"), val = tensor<bool, [2]>([true, true])];
+            tensor<int32, [1, 223]> var_43 = slice_by_index(begin = var_43_begin_0, end = var_43_end_0, end_mask = var_43_end_mask_0, x = context_ids)[name = tensor<string, []>("op_43")];
+            tensor<int32, []> var_45 = const()[name = tensor<string, []>("op_45"), val = tensor<int32, []>(1)];
+            tensor<bool, []> safe_context_ids_interleave_0 = const()[name = tensor<string, []>("safe_context_ids_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<string, []> var_43_promoted_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_43_promoted_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1, 223]> var_43_to_fp16 = cast(dtype = var_43_promoted_to_fp16_dtype_0, x = var_43)[name = tensor<string, []>("cast_66")];
+            tensor<fp16, [1, 224]> safe_context_ids_cast_fp16 = concat(axis = var_45, interleave = safe_context_ids_interleave_0, values = (fill_like_0_cast_fp16, var_43_to_fp16))[name = tensor<string, []>("safe_context_ids_cast_fp16")];
+            tensor<int32, []> var_59_batch_dims_0 = const()[name = tensor<string, []>("op_59_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> var_59_validate_indices_0 = const()[name = tensor<string, []>("op_59_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [257, 128]> model_embedding_weight_to_fp16 = const()[name = tensor<string, []>("model_embedding_weight_to_fp16"), val = tensor<fp16, [257, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<string, []> context_ids_to_int16_dtype_0 = const()[name = tensor<string, []>("context_ids_to_int16_dtype_0"), val = tensor<string, []>("int16")];
+            tensor<string, []> cast_54_dtype_0 = const()[name = tensor<string, []>("cast_54_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, []> greater_equal_0_y_0 = const()[name = tensor<string, []>("greater_equal_0_y_0"), val = tensor<int32, []>(0)];
+            tensor<int16, [1, 224]> context_ids_to_int16 = cast(dtype = context_ids_to_int16_dtype_0, x = context_ids)[name = tensor<string, []>("cast_65")];
+            tensor<int32, [1, 224]> cast_54 = cast(dtype = cast_54_dtype_0, x = context_ids_to_int16)[name = tensor<string, []>("cast_64")];
+            tensor<bool, [1, 224]> greater_equal_0 = greater_equal(x = cast_54, y = greater_equal_0_y_0)[name = tensor<string, []>("greater_equal_0")];
+            tensor<int32, []> slice_by_index_18 = const()[name = tensor<string, []>("slice_by_index_18"), val = tensor<int32, []>(257)];
+            tensor<int32, [1, 224]> add_3 = add(x = cast_54, y = slice_by_index_18)[name = tensor<string, []>("add_3")];
+            tensor<int32, [1, 224]> select_0 = select(a = cast_54, b = add_3, cond = greater_equal_0)[name = tensor<string, []>("select_0")];
+            tensor<int32, []> var_59_cast_fp16_cast_uint16_axis_0 = const()[name = tensor<string, []>("op_59_cast_fp16_cast_uint16_axis_0"), val = tensor<int32, []>(0)];
+            tensor<string, []> select_0_to_int16_dtype_0 = const()[name = tensor<string, []>("select_0_to_int16_dtype_0"), val = tensor<string, []>("int16")];
+            tensor<int16, [1, 224]> select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = tensor<string, []>("cast_63")];
+            tensor<fp16, [1, 224, 128]> var_59_cast_fp16_cast_uint16_cast_uint16 = gather(axis = var_59_cast_fp16_cast_uint16_axis_0, batch_dims = var_59_batch_dims_0, indices = select_0_to_int16, validate_indices = var_59_validate_indices_0, x = model_embedding_weight_to_fp16)[name = tensor<string, []>("op_59_cast_fp16_cast_uint16_cast_uint16")];
+            tensor<fp16, [224, 128]> model_position_weight_to_fp16 = const()[name = tensor<string, []>("model_position_weight_to_fp16"), val = tensor<fp16, [224, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65920)))];
+            tensor<fp16, [1, 224, 128]> src_1_cast_fp16 = add(x = var_59_cast_fp16_cast_uint16_cast_uint16, y = model_position_weight_to_fp16)[name = tensor<string, []>("src_1_cast_fp16")];
+            tensor<fp16, []> var_66_promoted_to_fp16 = const()[name = tensor<string, []>("op_66_promoted_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<bool, [1, 224]> mask_1_cast_fp16 = equal(x = safe_context_ids_cast_fp16, y = var_66_promoted_to_fp16)[name = tensor<string, []>("mask_1_cast_fp16")];
+            tensor<fp16, []> var_82_to_fp16 = const()[name = tensor<string, []>("op_82_to_fp16"), val = tensor<fp16, []>(-inf)];
+            tensor<fp16, [1, 224]> var_90_to_fp16 = const()[name = tensor<string, []>("op_90_to_fp16"), val = tensor<fp16, [1, 224]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(123328)))];
+            tensor<fp16, [1, 224]> key_padding_mask_1_cast_fp16 = select(a = var_82_to_fp16, b = var_90_to_fp16, cond = mask_1_cast_fp16)[name = tensor<string, []>("key_padding_mask_1_cast_fp16")];
+            tensor<int32, [1]> query_1_axes_0 = const()[name = tensor<string, []>("query_1_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [128]> model_encoder_layers_0_norm1_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_0_norm1_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(123840)))];
+            tensor<fp16, [128]> model_encoder_layers_0_norm1_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_0_norm1_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(124160)))];
+            tensor<fp16, []> var_69_to_fp16 = const()[name = tensor<string, []>("op_69_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 224, 128]> query_1_cast_fp16 = layer_norm(axes = query_1_axes_0, beta = model_encoder_layers_0_norm1_bias_to_fp16, epsilon = var_69_to_fp16, gamma = model_encoder_layers_0_norm1_weight_to_fp16, x = src_1_cast_fp16)[name = tensor<string, []>("query_1_cast_fp16")];
+            tensor<int32, [3]> query_3_perm_0 = const()[name = tensor<string, []>("query_3_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<fp16, [384, 128]> model_encoder_layers_0_self_attn_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_0_self_attn_in_proj_weight_to_fp16"), val = tensor<fp16, [384, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(124480)))];
+            tensor<fp16, [384]> model_encoder_layers_0_self_attn_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_0_self_attn_in_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(222848)))];
+            tensor<fp16, [224, 1, 128]> query_3_cast_fp16 = transpose(perm = query_3_perm_0, x = query_1_cast_fp16)[name = tensor<string, []>("transpose_24")];
+            tensor<fp16, [224, 1, 384]> linear_0_cast_fp16 = linear(bias = model_encoder_layers_0_self_attn_in_proj_bias_to_fp16, weight = model_encoder_layers_0_self_attn_in_proj_weight_to_fp16, x = query_3_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")];
+            tensor<int32, [4]> concat_0 = const()[name = tensor<string, []>("concat_0"), val = tensor<int32, [4]>([224, 1, 3, 128])];
+            tensor<fp16, [224, 1, 3, 128]> var_124_cast_fp16 = reshape(shape = concat_0, x = linear_0_cast_fp16)[name = tensor<string, []>("op_124_cast_fp16")];
+            tensor<int32, [1]> var_125_axes_0 = const()[name = tensor<string, []>("op_125_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 224, 1, 3, 128]> var_125_cast_fp16 = expand_dims(axes = var_125_axes_0, x = var_124_cast_fp16)[name = tensor<string, []>("op_125_cast_fp16")];
+            tensor<int32, [5]> var_126_perm_0 = const()[name = tensor<string, []>("op_126_perm_0"), val = tensor<int32, [5]>([-2, 1, 2, 0, 4])];
+            tensor<int32, [1]> var_127_axes_0 = const()[name = tensor<string, []>("op_127_axes_0"), val = tensor<int32, [1]>([-2])];
+            tensor<fp16, [3, 224, 1, 1, 128]> var_126_cast_fp16 = transpose(perm = var_126_perm_0, x = var_125_cast_fp16)[name = tensor<string, []>("transpose_23")];
+            tensor<fp16, [3, 224, 1, 128]> var_127_cast_fp16 = squeeze(axes = var_127_axes_0, x = var_126_cast_fp16)[name = tensor<string, []>("op_127_cast_fp16")];
+            tensor<int32, [4]> q_1_begin_0 = const()[name = tensor<string, []>("q_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> q_1_end_0 = const()[name = tensor<string, []>("q_1_end_0"), val = tensor<int32, [4]>([1, 224, 1, 128])];
+            tensor<bool, [4]> q_1_end_mask_0 = const()[name = tensor<string, []>("q_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> q_1_squeeze_mask_0 = const()[name = tensor<string, []>("q_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [224, 1, 128]> q_1_cast_fp16 = slice_by_index(begin = q_1_begin_0, end = q_1_end_0, end_mask = q_1_end_mask_0, squeeze_mask = q_1_squeeze_mask_0, x = var_127_cast_fp16)[name = tensor<string, []>("q_1_cast_fp16")];
+            tensor<int32, [4]> k_1_begin_0 = const()[name = tensor<string, []>("k_1_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
+            tensor<int32, [4]> k_1_end_0 = const()[name = tensor<string, []>("k_1_end_0"), val = tensor<int32, [4]>([2, 224, 1, 128])];
+            tensor<bool, [4]> k_1_end_mask_0 = const()[name = tensor<string, []>("k_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> k_1_squeeze_mask_0 = const()[name = tensor<string, []>("k_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [224, 1, 128]> k_1_cast_fp16 = slice_by_index(begin = k_1_begin_0, end = k_1_end_0, end_mask = k_1_end_mask_0, squeeze_mask = k_1_squeeze_mask_0, x = var_127_cast_fp16)[name = tensor<string, []>("k_1_cast_fp16")];
+            tensor<int32, [4]> v_1_begin_0 = const()[name = tensor<string, []>("v_1_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
+            tensor<int32, [4]> v_1_end_0 = const()[name = tensor<string, []>("v_1_end_0"), val = tensor<int32, [4]>([3, 224, 1, 128])];
+            tensor<bool, [4]> v_1_end_mask_0 = const()[name = tensor<string, []>("v_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> v_1_squeeze_mask_0 = const()[name = tensor<string, []>("v_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [224, 1, 128]> v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = var_127_cast_fp16)[name = tensor<string, []>("v_1_cast_fp16")];
+            tensor<int32, [3]> var_135 = const()[name = tensor<string, []>("op_135"), val = tensor<int32, [3]>([224, 4, 32])];
+            tensor<fp16, [224, 4, 32]> var_136_cast_fp16 = reshape(shape = var_135, x = q_1_cast_fp16)[name = tensor<string, []>("op_136_cast_fp16")];
+            tensor<int32, [3]> q_3_perm_0 = const()[name = tensor<string, []>("q_3_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<int32, [3]> var_142 = const()[name = tensor<string, []>("op_142"), val = tensor<int32, [3]>([224, 4, 32])];
+            tensor<fp16, [224, 4, 32]> var_143_cast_fp16 = reshape(shape = var_142, x = k_1_cast_fp16)[name = tensor<string, []>("op_143_cast_fp16")];
+            tensor<int32, [3]> k_3_perm_0 = const()[name = tensor<string, []>("k_3_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<int32, [3]> var_149 = const()[name = tensor<string, []>("op_149"), val = tensor<int32, [3]>([224, 4, 32])];
+            tensor<fp16, [224, 4, 32]> var_150_cast_fp16 = reshape(shape = var_149, x = v_1_cast_fp16)[name = tensor<string, []>("op_150_cast_fp16")];
+            tensor<int32, [3]> v_3_perm_0 = const()[name = tensor<string, []>("v_3_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<int32, [4]> var_153 = const()[name = tensor<string, []>("op_153"), val = tensor<int32, [4]>([1, 1, 1, 224])];
+            tensor<fp16, [1, 1, 1, 224]> var_154_cast_fp16 = reshape(shape = var_153, x = key_padding_mask_1_cast_fp16)[name = tensor<string, []>("op_154_cast_fp16")];
+            tensor<int32, [4]> var_156_reps_0 = const()[name = tensor<string, []>("op_156_reps_0"), val = tensor<int32, [4]>([1, 4, 1, 1])];
+            tensor<fp16, [1, 4, 1, 224]> var_156_cast_fp16 = tile(reps = var_156_reps_0, x = var_154_cast_fp16)[name = tensor<string, []>("op_156_cast_fp16")];
+            tensor<int32, [4]> var_164 = const()[name = tensor<string, []>("op_164"), val = tensor<int32, [4]>([1, 4, 224, 32])];
+            tensor<fp16, [4, 224, 32]> q_3_cast_fp16 = transpose(perm = q_3_perm_0, x = var_136_cast_fp16)[name = tensor<string, []>("transpose_22")];
+            tensor<fp16, [1, 4, 224, 32]> q_5_cast_fp16 = reshape(shape = var_164, x = q_3_cast_fp16)[name = tensor<string, []>("q_5_cast_fp16")];
+            tensor<int32, [4]> var_166 = const()[name = tensor<string, []>("op_166"), val = tensor<int32, [4]>([1, 4, 224, 32])];
+            tensor<fp16, [4, 224, 32]> k_3_cast_fp16 = transpose(perm = k_3_perm_0, x = var_143_cast_fp16)[name = tensor<string, []>("transpose_21")];
+            tensor<fp16, [1, 4, 224, 32]> k_5_cast_fp16 = reshape(shape = var_166, x = k_3_cast_fp16)[name = tensor<string, []>("k_5_cast_fp16")];
+            tensor<int32, [4]> var_168 = const()[name = tensor<string, []>("op_168"), val = tensor<int32, [4]>([1, 4, 224, 32])];
+            tensor<fp16, [4, 224, 32]> v_3_cast_fp16 = transpose(perm = v_3_perm_0, x = var_150_cast_fp16)[name = tensor<string, []>("transpose_20")];
+            tensor<fp16, [1, 4, 224, 32]> v_5_cast_fp16 = reshape(shape = var_168, x = v_3_cast_fp16)[name = tensor<string, []>("v_5_cast_fp16")];
+            tensor<fp16, []> mul_1_y_0_to_fp16 = const()[name = tensor<string, []>("mul_1_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-3)];
+            tensor<fp16, [1, 4, 224, 32]> mul_1_cast_fp16 = mul(x = q_5_cast_fp16, y = mul_1_y_0_to_fp16)[name = tensor<string, []>("mul_1_cast_fp16")];
+            tensor<bool, []> matmul_0_transpose_y_0 = const()[name = tensor<string, []>("matmul_0_transpose_y_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> matmul_0_transpose_x_0 = const()[name = tensor<string, []>("matmul_0_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 4, 224, 224]> matmul_0_cast_fp16 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_1_cast_fp16, y = k_5_cast_fp16)[name = tensor<string, []>("matmul_0_cast_fp16")];
+            tensor<fp16, [1, 4, 224, 224]> add_0_cast_fp16 = add(x = matmul_0_cast_fp16, y = var_156_cast_fp16)[name = tensor<string, []>("add_0_cast_fp16")];
+            tensor<int32, []> softmax_0_axis_0 = const()[name = tensor<string, []>("softmax_0_axis_0"), val = tensor<int32, []>(-1)];
+            tensor<fp16, [1, 4, 224, 224]> softmax_0_cast_fp16 = softmax(axis = softmax_0_axis_0, x = add_0_cast_fp16)[name = tensor<string, []>("softmax_0_cast_fp16")];
+            tensor<bool, []> attn_output_1_transpose_x_0 = const()[name = tensor<string, []>("attn_output_1_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attn_output_1_transpose_y_0 = const()[name = tensor<string, []>("attn_output_1_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 4, 224, 32]> attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0_cast_fp16, y = v_5_cast_fp16)[name = tensor<string, []>("attn_output_1_cast_fp16")];
+            tensor<int32, [4]> var_171 = const()[name = tensor<string, []>("op_171"), val = tensor<int32, [4]>([2, 0, 1, 3])];
+            tensor<int32, [2]> var_176 = const()[name = tensor<string, []>("op_176"), val = tensor<int32, [2]>([224, 128])];
+            tensor<fp16, [224, 1, 4, 32]> var_172_cast_fp16 = transpose(perm = var_171, x = attn_output_1_cast_fp16)[name = tensor<string, []>("transpose_19")];
+            tensor<fp16, [224, 128]> attn_output_3_cast_fp16 = reshape(shape = var_176, x = var_172_cast_fp16)[name = tensor<string, []>("attn_output_3_cast_fp16")];
+            tensor<fp16, [128, 128]> model_encoder_layers_0_self_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_0_self_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [128, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(223680)))];
+            tensor<fp16, [128]> model_encoder_layers_0_self_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_0_self_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(256512)))];
+            tensor<fp16, [224, 128]> linear_1_cast_fp16 = linear(bias = model_encoder_layers_0_self_attn_out_proj_bias_to_fp16, weight = model_encoder_layers_0_self_attn_out_proj_weight_to_fp16, x = attn_output_3_cast_fp16)[name = tensor<string, []>("linear_1_cast_fp16")];
+            tensor<int32, [3]> var_180 = const()[name = tensor<string, []>("op_180"), val = tensor<int32, [3]>([224, 1, 128])];
+            tensor<fp16, [224, 1, 128]> attn_output_7_cast_fp16 = reshape(shape = var_180, x = linear_1_cast_fp16)[name = tensor<string, []>("attn_output_7_cast_fp16")];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = tensor<string, []>("input_3_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<fp16, [1, 224, 128]> input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = attn_output_7_cast_fp16)[name = tensor<string, []>("transpose_18")];
+            tensor<fp16, [1, 224, 128]> input_5_cast_fp16 = add(x = src_1_cast_fp16, y = input_3_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
+            tensor<int32, [1]> input_7_axes_0 = const()[name = tensor<string, []>("input_7_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [128]> model_encoder_layers_0_norm2_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_0_norm2_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(256832)))];
+            tensor<fp16, [128]> model_encoder_layers_0_norm2_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_0_norm2_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(257152)))];
+            tensor<fp16, [1, 224, 128]> input_7_cast_fp16 = layer_norm(axes = input_7_axes_0, beta = model_encoder_layers_0_norm2_bias_to_fp16, epsilon = var_69_to_fp16, gamma = model_encoder_layers_0_norm2_weight_to_fp16, x = input_5_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
+            tensor<fp16, [512, 128]> model_encoder_layers_0_linear1_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_0_linear1_weight_to_fp16"), val = tensor<fp16, [512, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(257472)))];
+            tensor<fp16, [512]> model_encoder_layers_0_linear1_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_0_linear1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(388608)))];
+            tensor<fp16, [1, 224, 512]> linear_2_cast_fp16 = linear(bias = model_encoder_layers_0_linear1_bias_to_fp16, weight = model_encoder_layers_0_linear1_weight_to_fp16, x = input_7_cast_fp16)[name = tensor<string, []>("linear_2_cast_fp16")];
+            tensor<fp16, [1, 224, 512]> input_11_cast_fp16 = relu(x = linear_2_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")];
+            tensor<fp16, [128, 512]> model_encoder_layers_0_linear2_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_0_linear2_weight_to_fp16"), val = tensor<fp16, [128, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(389696)))];
+            tensor<fp16, [128]> model_encoder_layers_0_linear2_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_0_linear2_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(520832)))];
+            tensor<fp16, [1, 224, 128]> linear_3_cast_fp16 = linear(bias = model_encoder_layers_0_linear2_bias_to_fp16, weight = model_encoder_layers_0_linear2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor<string, []>("linear_3_cast_fp16")];
+            tensor<fp16, [1, 224, 128]> input_17_cast_fp16 = add(x = input_5_cast_fp16, y = linear_3_cast_fp16)[name = tensor<string, []>("input_17_cast_fp16")];
+            tensor<int32, [1]> query_5_axes_0 = const()[name = tensor<string, []>("query_5_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [128]> model_encoder_layers_1_norm1_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_1_norm1_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(521152)))];
+            tensor<fp16, [128]> model_encoder_layers_1_norm1_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_1_norm1_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(521472)))];
+            tensor<fp16, [1, 224, 128]> query_5_cast_fp16 = layer_norm(axes = query_5_axes_0, beta = model_encoder_layers_1_norm1_bias_to_fp16, epsilon = var_69_to_fp16, gamma = model_encoder_layers_1_norm1_weight_to_fp16, x = input_17_cast_fp16)[name = tensor<string, []>("query_5_cast_fp16")];
+            tensor<int32, [3]> query_7_perm_0 = const()[name = tensor<string, []>("query_7_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<fp16, [384, 128]> model_encoder_layers_1_self_attn_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_1_self_attn_in_proj_weight_to_fp16"), val = tensor<fp16, [384, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(521792)))];
+            tensor<fp16, [384]> model_encoder_layers_1_self_attn_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_1_self_attn_in_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(620160)))];
+            tensor<fp16, [224, 1, 128]> query_7_cast_fp16 = transpose(perm = query_7_perm_0, x = query_5_cast_fp16)[name = tensor<string, []>("transpose_17")];
+            tensor<fp16, [224, 1, 384]> linear_4_cast_fp16 = linear(bias = model_encoder_layers_1_self_attn_in_proj_bias_to_fp16, weight = model_encoder_layers_1_self_attn_in_proj_weight_to_fp16, x = query_7_cast_fp16)[name = tensor<string, []>("linear_4_cast_fp16")];
+            tensor<int32, [4]> concat_2 = const()[name = tensor<string, []>("concat_2"), val = tensor<int32, [4]>([224, 1, 3, 128])];
+            tensor<fp16, [224, 1, 3, 128]> var_231_cast_fp16 = reshape(shape = concat_2, x = linear_4_cast_fp16)[name = tensor<string, []>("op_231_cast_fp16")];
+            tensor<int32, [1]> var_232_axes_0 = const()[name = tensor<string, []>("op_232_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 224, 1, 3, 128]> var_232_cast_fp16 = expand_dims(axes = var_232_axes_0, x = var_231_cast_fp16)[name = tensor<string, []>("op_232_cast_fp16")];
+            tensor<int32, [5]> var_233_perm_0 = const()[name = tensor<string, []>("op_233_perm_0"), val = tensor<int32, [5]>([-2, 1, 2, 0, 4])];
+            tensor<int32, [1]> var_234_axes_0 = const()[name = tensor<string, []>("op_234_axes_0"), val = tensor<int32, [1]>([-2])];
+            tensor<fp16, [3, 224, 1, 1, 128]> var_233_cast_fp16 = transpose(perm = var_233_perm_0, x = var_232_cast_fp16)[name = tensor<string, []>("transpose_16")];
+            tensor<fp16, [3, 224, 1, 128]> var_234_cast_fp16 = squeeze(axes = var_234_axes_0, x = var_233_cast_fp16)[name = tensor<string, []>("op_234_cast_fp16")];
+            tensor<int32, [4]> q_7_begin_0 = const()[name = tensor<string, []>("q_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> q_7_end_0 = const()[name = tensor<string, []>("q_7_end_0"), val = tensor<int32, [4]>([1, 224, 1, 128])];
+            tensor<bool, [4]> q_7_end_mask_0 = const()[name = tensor<string, []>("q_7_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> q_7_squeeze_mask_0 = const()[name = tensor<string, []>("q_7_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [224, 1, 128]> q_7_cast_fp16 = slice_by_index(begin = q_7_begin_0, end = q_7_end_0, end_mask = q_7_end_mask_0, squeeze_mask = q_7_squeeze_mask_0, x = var_234_cast_fp16)[name = tensor<string, []>("q_7_cast_fp16")];
+            tensor<int32, [4]> k_7_begin_0 = const()[name = tensor<string, []>("k_7_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
+            tensor<int32, [4]> k_7_end_0 = const()[name = tensor<string, []>("k_7_end_0"), val = tensor<int32, [4]>([2, 224, 1, 128])];
+            tensor<bool, [4]> k_7_end_mask_0 = const()[name = tensor<string, []>("k_7_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> k_7_squeeze_mask_0 = const()[name = tensor<string, []>("k_7_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [224, 1, 128]> k_7_cast_fp16 = slice_by_index(begin = k_7_begin_0, end = k_7_end_0, end_mask = k_7_end_mask_0, squeeze_mask = k_7_squeeze_mask_0, x = var_234_cast_fp16)[name = tensor<string, []>("k_7_cast_fp16")];
+            tensor<int32, [4]> v_7_begin_0 = const()[name = tensor<string, []>("v_7_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
+            tensor<int32, [4]> v_7_end_0 = const()[name = tensor<string, []>("v_7_end_0"), val = tensor<int32, [4]>([3, 224, 1, 128])];
+            tensor<bool, [4]> v_7_end_mask_0 = const()[name = tensor<string, []>("v_7_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> v_7_squeeze_mask_0 = const()[name = tensor<string, []>("v_7_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [224, 1, 128]> v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = var_234_cast_fp16)[name = tensor<string, []>("v_7_cast_fp16")];
+            tensor<int32, [3]> var_242 = const()[name = tensor<string, []>("op_242"), val = tensor<int32, [3]>([224, 4, 32])];
+            tensor<fp16, [224, 4, 32]> var_243_cast_fp16 = reshape(shape = var_242, x = q_7_cast_fp16)[name = tensor<string, []>("op_243_cast_fp16")];
+            tensor<int32, [3]> q_9_perm_0 = const()[name = tensor<string, []>("q_9_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<int32, [3]> var_249 = const()[name = tensor<string, []>("op_249"), val = tensor<int32, [3]>([224, 4, 32])];
+            tensor<fp16, [224, 4, 32]> var_250_cast_fp16 = reshape(shape = var_249, x = k_7_cast_fp16)[name = tensor<string, []>("op_250_cast_fp16")];
+            tensor<int32, [3]> k_9_perm_0 = const()[name = tensor<string, []>("k_9_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<int32, [3]> var_256 = const()[name = tensor<string, []>("op_256"), val = tensor<int32, [3]>([224, 4, 32])];
+            tensor<fp16, [224, 4, 32]> var_257_cast_fp16 = reshape(shape = var_256, x = v_7_cast_fp16)[name = tensor<string, []>("op_257_cast_fp16")];
+            tensor<int32, [3]> v_9_perm_0 = const()[name = tensor<string, []>("v_9_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<int32, [4]> var_271 = const()[name = tensor<string, []>("op_271"), val = tensor<int32, [4]>([1, 4, 224, 32])];
+            tensor<fp16, [4, 224, 32]> q_9_cast_fp16 = transpose(perm = q_9_perm_0, x = var_243_cast_fp16)[name = tensor<string, []>("transpose_15")];
+            tensor<fp16, [1, 4, 224, 32]> q_11_cast_fp16 = reshape(shape = var_271, x = q_9_cast_fp16)[name = tensor<string, []>("q_11_cast_fp16")];
+            tensor<int32, [4]> var_273 = const()[name = tensor<string, []>("op_273"), val = tensor<int32, [4]>([1, 4, 224, 32])];
+            tensor<fp16, [4, 224, 32]> k_9_cast_fp16 = transpose(perm = k_9_perm_0, x = var_250_cast_fp16)[name = tensor<string, []>("transpose_14")];
+            tensor<fp16, [1, 4, 224, 32]> k_11_cast_fp16 = reshape(shape = var_273, x = k_9_cast_fp16)[name = tensor<string, []>("k_11_cast_fp16")];
+            tensor<int32, [4]> var_275 = const()[name = tensor<string, []>("op_275"), val = tensor<int32, [4]>([1, 4, 224, 32])];
+            tensor<fp16, [4, 224, 32]> v_9_cast_fp16 = transpose(perm = v_9_perm_0, x = var_257_cast_fp16)[name = tensor<string, []>("transpose_13")];
+            tensor<fp16, [1, 4, 224, 32]> v_11_cast_fp16 = reshape(shape = var_275, x = v_9_cast_fp16)[name = tensor<string, []>("v_11_cast_fp16")];
+            tensor<fp16, []> mul_3_y_0_to_fp16 = const()[name = tensor<string, []>("mul_3_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-3)];
+            tensor<fp16, [1, 4, 224, 32]> mul_3_cast_fp16 = mul(x = q_11_cast_fp16, y = mul_3_y_0_to_fp16)[name = tensor<string, []>("mul_3_cast_fp16")];
+            tensor<bool, []> matmul_1_transpose_y_0 = const()[name = tensor<string, []>("matmul_1_transpose_y_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> matmul_1_transpose_x_0 = const()[name = tensor<string, []>("matmul_1_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 4, 224, 224]> matmul_1_cast_fp16 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = mul_3_cast_fp16, y = k_11_cast_fp16)[name = tensor<string, []>("matmul_1_cast_fp16")];
+            tensor<fp16, [1, 4, 224, 224]> add_1_cast_fp16 = add(x = matmul_1_cast_fp16, y = var_156_cast_fp16)[name = tensor<string, []>("add_1_cast_fp16")];
+            tensor<int32, []> softmax_1_axis_0 = const()[name = tensor<string, []>("softmax_1_axis_0"), val = tensor<int32, []>(-1)];
+            tensor<fp16, [1, 4, 224, 224]> softmax_1_cast_fp16 = softmax(axis = softmax_1_axis_0, x = add_1_cast_fp16)[name = tensor<string, []>("softmax_1_cast_fp16")];
+            tensor<bool, []> attn_output_9_transpose_x_0 = const()[name = tensor<string, []>("attn_output_9_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attn_output_9_transpose_y_0 = const()[name = tensor<string, []>("attn_output_9_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 4, 224, 32]> attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1_cast_fp16, y = v_11_cast_fp16)[name = tensor<string, []>("attn_output_9_cast_fp16")];
+            tensor<int32, [4]> var_278 = const()[name = tensor<string, []>("op_278"), val = tensor<int32, [4]>([2, 0, 1, 3])];
+            tensor<int32, [2]> var_283 = const()[name = tensor<string, []>("op_283"), val = tensor<int32, [2]>([224, 128])];
+            tensor<fp16, [224, 1, 4, 32]> var_279_cast_fp16 = transpose(perm = var_278, x = attn_output_9_cast_fp16)[name = tensor<string, []>("transpose_12")];
+            tensor<fp16, [224, 128]> attn_output_11_cast_fp16 = reshape(shape = var_283, x = var_279_cast_fp16)[name = tensor<string, []>("attn_output_11_cast_fp16")];
+            tensor<fp16, [128, 128]> model_encoder_layers_1_self_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_1_self_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [128, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(620992)))];
+            tensor<fp16, [128]> model_encoder_layers_1_self_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_1_self_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(653824)))];
+            tensor<fp16, [224, 128]> linear_5_cast_fp16 = linear(bias = model_encoder_layers_1_self_attn_out_proj_bias_to_fp16, weight = model_encoder_layers_1_self_attn_out_proj_weight_to_fp16, x = attn_output_11_cast_fp16)[name = tensor<string, []>("linear_5_cast_fp16")];
+            tensor<int32, [3]> var_287 = const()[name = tensor<string, []>("op_287"), val = tensor<int32, [3]>([224, 1, 128])];
+            tensor<fp16, [224, 1, 128]> attn_output_15_cast_fp16 = reshape(shape = var_287, x = linear_5_cast_fp16)[name = tensor<string, []>("attn_output_15_cast_fp16")];
+            tensor<int32, [3]> input_19_perm_0 = const()[name = tensor<string, []>("input_19_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<fp16, [1, 224, 128]> input_19_cast_fp16 = transpose(perm = input_19_perm_0, x = attn_output_15_cast_fp16)[name = tensor<string, []>("transpose_11")];
+            tensor<fp16, [1, 224, 128]> input_21_cast_fp16 = add(x = input_17_cast_fp16, y = input_19_cast_fp16)[name = tensor<string, []>("input_21_cast_fp16")];
+            tensor<int32, [1]> input_23_axes_0 = const()[name = tensor<string, []>("input_23_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [128]> model_encoder_layers_1_norm2_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_1_norm2_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(654144)))];
+            tensor<fp16, [128]> model_encoder_layers_1_norm2_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_1_norm2_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(654464)))];
+            tensor<fp16, [1, 224, 128]> input_23_cast_fp16 = layer_norm(axes = input_23_axes_0, beta = model_encoder_layers_1_norm2_bias_to_fp16, epsilon = var_69_to_fp16, gamma = model_encoder_layers_1_norm2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("input_23_cast_fp16")];
+            tensor<fp16, [512, 128]> model_encoder_layers_1_linear1_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_1_linear1_weight_to_fp16"), val = tensor<fp16, [512, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(654784)))];
+            tensor<fp16, [512]> model_encoder_layers_1_linear1_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_1_linear1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(785920)))];
+            tensor<fp16, [1, 224, 512]> linear_6_cast_fp16 = linear(bias = model_encoder_layers_1_linear1_bias_to_fp16, weight = model_encoder_layers_1_linear1_weight_to_fp16, x = input_23_cast_fp16)[name = tensor<string, []>("linear_6_cast_fp16")];
+            tensor<fp16, [1, 224, 512]> input_27_cast_fp16 = relu(x = linear_6_cast_fp16)[name = tensor<string, []>("input_27_cast_fp16")];
+            tensor<fp16, [128, 512]> model_encoder_layers_1_linear2_weight_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_1_linear2_weight_to_fp16"), val = tensor<fp16, [128, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(787008)))];
+            tensor<fp16, [128]> model_encoder_layers_1_linear2_bias_to_fp16 = const()[name = tensor<string, []>("model_encoder_layers_1_linear2_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(918144)))];
+            tensor<fp16, [1, 224, 128]> linear_7_cast_fp16 = linear(bias = model_encoder_layers_1_linear2_bias_to_fp16, weight = model_encoder_layers_1_linear2_weight_to_fp16, x = input_27_cast_fp16)[name = tensor<string, []>("linear_7_cast_fp16")];
+            tensor<fp16, [1, 224, 128]> context_cast_fp16 = add(x = input_21_cast_fp16, y = linear_7_cast_fp16)[name = tensor<string, []>("context_cast_fp16")];
+            tensor<int32, [2]> var_320 = const()[name = tensor<string, []>("op_320"), val = tensor<int32, [2]>([32, 96])];
+            tensor<int32, [32, 96]> flat_ids = reshape(shape = var_320, x = option_ids)[name = tensor<string, []>("flat_ids")];
+            tensor<int32, []> var_322 = const()[name = tensor<string, []>("op_322"), val = tensor<int32, []>(0)];
+            tensor<bool, [32, 96]> flat_mask = not_equal(x = flat_ids, y = var_322)[name = tensor<string, []>("flat_mask")];
+            tensor<int32, [2]> var_333_begin_0 = const()[name = tensor<string, []>("op_333_begin_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [2]> var_333_end_0 = const()[name = tensor<string, []>("op_333_end_0"), val = tensor<int32, [2]>([32, 1])];
+            tensor<bool, [2]> var_333_end_mask_0 = const()[name = tensor<string, []>("op_333_end_mask_0"), val = tensor<bool, [2]>([true, false])];
+            tensor<int32, [32, 1]> var_333 = slice_by_index(begin = var_333_begin_0, end = var_333_end_0, end_mask = var_333_end_mask_0, x = flat_ids)[name = tensor<string, []>("op_333")];
+            tensor<fp16, []> fill_like_1_value_0_to_fp16 = const()[name = tensor<string, []>("fill_like_1_value_0_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [32, 1]> fill_like_1_cast_fp16 = fill_like(ref_tensor = var_333, value = fill_like_1_value_0_to_fp16)[name = tensor<string, []>("fill_like_1_cast_fp16")];
+            tensor<int32, [2]> var_349_begin_0 = const()[name = tensor<string, []>("op_349_begin_0"), val = tensor<int32, [2]>([0, 1])];
+            tensor<int32, [2]> var_349_end_0 = const()[name = tensor<string, []>("op_349_end_0"), val = tensor<int32, [2]>([32, 96])];
+            tensor<bool, [2]> var_349_end_mask_0 = const()[name = tensor<string, []>("op_349_end_mask_0"), val = tensor<bool, [2]>([true, true])];
+            tensor<int32, [32, 95]> var_349 = slice_by_index(begin = var_349_begin_0, end = var_349_end_0, end_mask = var_349_end_mask_0, x = flat_ids)[name = tensor<string, []>("op_349")];
+            tensor<int32, []> var_351 = const()[name = tensor<string, []>("op_351"), val = tensor<int32, []>(1)];
+            tensor<bool, []> safe_ids_interleave_0 = const()[name = tensor<string, []>("safe_ids_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<string, []> var_349_promoted_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_349_promoted_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [32, 95]> var_349_to_fp16 = cast(dtype = var_349_promoted_to_fp16_dtype_0, x = var_349)[name = tensor<string, []>("cast_62")];
+            tensor<fp16, [32, 96]> safe_ids_cast_fp16 = concat(axis = var_351, interleave = safe_ids_interleave_0, values = (fill_like_1_cast_fp16, var_349_to_fp16))[name = tensor<string, []>("safe_ids_cast_fp16")];
+            tensor<int32, []> var_365_batch_dims_0 = const()[name = tensor<string, []>("op_365_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> var_365_validate_indices_0 = const()[name = tensor<string, []>("op_365_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<string, []> flat_ids_to_int16_dtype_0 = const()[name = tensor<string, []>("flat_ids_to_int16_dtype_0"), val = tensor<string, []>("int16")];
+            tensor<string, []> cast_55_dtype_0 = const()[name = tensor<string, []>("cast_55_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, []> greater_equal_1_y_0 = const()[name = tensor<string, []>("greater_equal_1_y_0"), val = tensor<int32, []>(0)];
+            tensor<int16, [32, 96]> flat_ids_to_int16 = cast(dtype = flat_ids_to_int16_dtype_0, x = flat_ids)[name = tensor<string, []>("cast_61")];
+            tensor<int32, [32, 96]> cast_55 = cast(dtype = cast_55_dtype_0, x = flat_ids_to_int16)[name = tensor<string, []>("cast_60")];
+            tensor<bool, [32, 96]> greater_equal_1 = greater_equal(x = cast_55, y = greater_equal_1_y_0)[name = tensor<string, []>("greater_equal_1")];
+            tensor<int32, []> slice_by_index_19 = const()[name = tensor<string, []>("slice_by_index_19"), val = tensor<int32, []>(257)];
+            tensor<int32, [32, 96]> add_4 = add(x = cast_55, y = slice_by_index_19)[name = tensor<string, []>("add_4")];
+            tensor<int32, [32, 96]> select_1 = select(a = cast_55, b = add_4, cond = greater_equal_1)[name = tensor<string, []>("select_1")];
+            tensor<int32, []> var_365_cast_fp16_cast_uint16_axis_0 = const()[name = tensor<string, []>("op_365_cast_fp16_cast_uint16_axis_0"), val = tensor<int32, []>(0)];
+            tensor<string, []> select_1_to_int16_dtype_0 = const()[name = tensor<string, []>("select_1_to_int16_dtype_0"), val = tensor<string, []>("int16")];
+            tensor<int16, [32, 96]> select_1_to_int16 = cast(dtype = select_1_to_int16_dtype_0, x = select_1)[name = tensor<string, []>("cast_59")];
+            tensor<fp16, [32, 96, 128]> var_365_cast_fp16_cast_uint16_cast_uint16 = gather(axis = var_365_cast_fp16_cast_uint16_axis_0, batch_dims = var_365_batch_dims_0, indices = select_1_to_int16, validate_indices = var_365_validate_indices_0, x = model_embedding_weight_to_fp16)[name = tensor<string, []>("op_365_cast_fp16_cast_uint16_cast_uint16")];
+            tensor<fp16, [96, 128]> var_369_to_fp16 = const()[name = tensor<string, []>("op_369_to_fp16"), val = tensor<fp16, [96, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(918464)))];
+            tensor<fp16, [32, 96, 128]> src_cast_fp16 = add(x = var_365_cast_fp16_cast_uint16_cast_uint16, y = var_369_to_fp16)[name = tensor<string, []>("src_cast_fp16")];
+            tensor<fp16, []> var_372_promoted_to_fp16 = const()[name = tensor<string, []>("op_372_promoted_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<bool, [32, 96]> mask_cast_fp16 = equal(x = safe_ids_cast_fp16, y = var_372_promoted_to_fp16)[name = tensor<string, []>("mask_cast_fp16")];
+            tensor<fp16, []> var_388_to_fp16 = const()[name = tensor<string, []>("op_388_to_fp16"), val = tensor<fp16, []>(-inf)];
+            tensor<fp16, [32, 96]> var_394_to_fp16 = const()[name = tensor<string, []>("op_394_to_fp16"), val = tensor<fp16, [32, 96]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(943104)))];
+            tensor<fp16, [32, 96]> key_padding_mask_7_cast_fp16 = select(a = var_388_to_fp16, b = var_394_to_fp16, cond = mask_cast_fp16)[name = tensor<string, []>("key_padding_mask_7_cast_fp16")];
+            tensor<int32, [1]> query_9_axes_0 = const()[name = tensor<string, []>("query_9_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [128]> model_option_encoder_layers_0_norm1_weight_to_fp16 = const()[name = tensor<string, []>("model_option_encoder_layers_0_norm1_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(949312)))];
+            tensor<fp16, [128]> model_option_encoder_layers_0_norm1_bias_to_fp16 = const()[name = tensor<string, []>("model_option_encoder_layers_0_norm1_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(949632)))];
+            tensor<fp16, []> var_375_to_fp16 = const()[name = tensor<string, []>("op_375_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [32, 96, 128]> query_9_cast_fp16 = layer_norm(axes = query_9_axes_0, beta = model_option_encoder_layers_0_norm1_bias_to_fp16, epsilon = var_375_to_fp16, gamma = model_option_encoder_layers_0_norm1_weight_to_fp16, x = src_cast_fp16)[name = tensor<string, []>("query_9_cast_fp16")];
+            tensor<int32, [3]> query_11_perm_0 = const()[name = tensor<string, []>("query_11_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<fp16, [384, 128]> model_option_encoder_layers_0_self_attn_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_option_encoder_layers_0_self_attn_in_proj_weight_to_fp16"), val = tensor<fp16, [384, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(949952)))];
+            tensor<fp16, [384]> model_option_encoder_layers_0_self_attn_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_option_encoder_layers_0_self_attn_in_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1048320)))];
+            tensor<fp16, [96, 32, 128]> query_11_cast_fp16 = transpose(perm = query_11_perm_0, x = query_9_cast_fp16)[name = tensor<string, []>("transpose_10")];
+            tensor<fp16, [96, 32, 384]> linear_8_cast_fp16 = linear(bias = model_option_encoder_layers_0_self_attn_in_proj_bias_to_fp16, weight = model_option_encoder_layers_0_self_attn_in_proj_weight_to_fp16, x = query_11_cast_fp16)[name = tensor<string, []>("linear_8_cast_fp16")];
+            tensor<int32, [4]> concat_4 = const()[name = tensor<string, []>("concat_4"), val = tensor<int32, [4]>([96, 32, 3, 128])];
+            tensor<fp16, [96, 32, 3, 128]> var_428_cast_fp16 = reshape(shape = concat_4, x = linear_8_cast_fp16)[name = tensor<string, []>("op_428_cast_fp16")];
+            tensor<int32, [1]> var_429_axes_0 = const()[name = tensor<string, []>("op_429_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 96, 32, 3, 128]> var_429_cast_fp16 = expand_dims(axes = var_429_axes_0, x = var_428_cast_fp16)[name = tensor<string, []>("op_429_cast_fp16")];
+            tensor<int32, [5]> var_430_perm_0 = const()[name = tensor<string, []>("op_430_perm_0"), val = tensor<int32, [5]>([-2, 1, 2, 0, 4])];
+            tensor<int32, [1]> var_431_axes_0 = const()[name = tensor<string, []>("op_431_axes_0"), val = tensor<int32, [1]>([-2])];
+            tensor<fp16, [3, 96, 32, 1, 128]> var_430_cast_fp16 = transpose(perm = var_430_perm_0, x = var_429_cast_fp16)[name = tensor<string, []>("transpose_9")];
+            tensor<fp16, [3, 96, 32, 128]> var_431_cast_fp16 = squeeze(axes = var_431_axes_0, x = var_430_cast_fp16)[name = tensor<string, []>("op_431_cast_fp16")];
+            tensor<int32, [4]> q_13_begin_0 = const()[name = tensor<string, []>("q_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> q_13_end_0 = const()[name = tensor<string, []>("q_13_end_0"), val = tensor<int32, [4]>([1, 96, 32, 128])];
+            tensor<bool, [4]> q_13_end_mask_0 = const()[name = tensor<string, []>("q_13_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> q_13_squeeze_mask_0 = const()[name = tensor<string, []>("q_13_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [96, 32, 128]> q_13_cast_fp16 = slice_by_index(begin = q_13_begin_0, end = q_13_end_0, end_mask = q_13_end_mask_0, squeeze_mask = q_13_squeeze_mask_0, x = var_431_cast_fp16)[name = tensor<string, []>("q_13_cast_fp16")];
+            tensor<int32, [4]> k_13_begin_0 = const()[name = tensor<string, []>("k_13_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
+            tensor<int32, [4]> k_13_end_0 = const()[name = tensor<string, []>("k_13_end_0"), val = tensor<int32, [4]>([2, 96, 32, 128])];
+            tensor<bool, [4]> k_13_end_mask_0 = const()[name = tensor<string, []>("k_13_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> k_13_squeeze_mask_0 = const()[name = tensor<string, []>("k_13_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [96, 32, 128]> k_13_cast_fp16 = slice_by_index(begin = k_13_begin_0, end = k_13_end_0, end_mask = k_13_end_mask_0, squeeze_mask = k_13_squeeze_mask_0, x = var_431_cast_fp16)[name = tensor<string, []>("k_13_cast_fp16")];
+            tensor<int32, [4]> v_13_begin_0 = const()[name = tensor<string, []>("v_13_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
+            tensor<int32, [4]> v_13_end_0 = const()[name = tensor<string, []>("v_13_end_0"), val = tensor<int32, [4]>([3, 96, 32, 128])];
+            tensor<bool, [4]> v_13_end_mask_0 = const()[name = tensor<string, []>("v_13_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> v_13_squeeze_mask_0 = const()[name = tensor<string, []>("v_13_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [96, 32, 128]> v_13_cast_fp16 = slice_by_index(begin = v_13_begin_0, end = v_13_end_0, end_mask = v_13_end_mask_0, squeeze_mask = v_13_squeeze_mask_0, x = var_431_cast_fp16)[name = tensor<string, []>("v_13_cast_fp16")];
+            tensor<int32, [3]> var_439 = const()[name = tensor<string, []>("op_439"), val = tensor<int32, [3]>([96, 128, 32])];
+            tensor<fp16, [96, 128, 32]> var_440_cast_fp16 = reshape(shape = var_439, x = q_13_cast_fp16)[name = tensor<string, []>("op_440_cast_fp16")];
+            tensor<int32, [3]> q_15_perm_0 = const()[name = tensor<string, []>("q_15_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<int32, [3]> var_446 = const()[name = tensor<string, []>("op_446"), val = tensor<int32, [3]>([96, 128, 32])];
+            tensor<fp16, [96, 128, 32]> var_447_cast_fp16 = reshape(shape = var_446, x = k_13_cast_fp16)[name = tensor<string, []>("op_447_cast_fp16")];
+            tensor<int32, [3]> k_15_perm_0 = const()[name = tensor<string, []>("k_15_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<int32, [3]> var_453 = const()[name = tensor<string, []>("op_453"), val = tensor<int32, [3]>([96, 128, 32])];
+            tensor<fp16, [96, 128, 32]> var_454_cast_fp16 = reshape(shape = var_453, x = v_13_cast_fp16)[name = tensor<string, []>("op_454_cast_fp16")];
+            tensor<int32, [3]> v_15_perm_0 = const()[name = tensor<string, []>("v_15_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<int32, [4]> var_457 = const()[name = tensor<string, []>("op_457"), val = tensor<int32, [4]>([32, 1, 1, 96])];
+            tensor<fp16, [32, 1, 1, 96]> var_458_cast_fp16 = reshape(shape = var_457, x = key_padding_mask_7_cast_fp16)[name = tensor<string, []>("op_458_cast_fp16")];
+            tensor<int32, [4]> var_460_reps_0 = const()[name = tensor<string, []>("op_460_reps_0"), val = tensor<int32, [4]>([1, 4, 1, 1])];
+            tensor<fp16, [32, 4, 1, 96]> var_460_cast_fp16 = tile(reps = var_460_reps_0, x = var_458_cast_fp16)[name = tensor<string, []>("op_460_cast_fp16")];
+            tensor<int32, [4]> var_468 = const()[name = tensor<string, []>("op_468"), val = tensor<int32, [4]>([32, 4, 96, 32])];
+            tensor<fp16, [128, 96, 32]> q_15_cast_fp16 = transpose(perm = q_15_perm_0, x = var_440_cast_fp16)[name = tensor<string, []>("transpose_8")];
+            tensor<fp16, [32, 4, 96, 32]> q_cast_fp16 = reshape(shape = var_468, x = q_15_cast_fp16)[name = tensor<string, []>("q_cast_fp16")];
+            tensor<int32, [4]> var_470 = const()[name = tensor<string, []>("op_470"), val = tensor<int32, [4]>([32, 4, 96, 32])];
+            tensor<fp16, [128, 96, 32]> k_15_cast_fp16 = transpose(perm = k_15_perm_0, x = var_447_cast_fp16)[name = tensor<string, []>("transpose_7")];
+            tensor<fp16, [32, 4, 96, 32]> k_cast_fp16 = reshape(shape = var_470, x = k_15_cast_fp16)[name = tensor<string, []>("k_cast_fp16")];
+            tensor<int32, [4]> var_472 = const()[name = tensor<string, []>("op_472"), val = tensor<int32, [4]>([32, 4, 96, 32])];
+            tensor<fp16, [128, 96, 32]> v_15_cast_fp16 = transpose(perm = v_15_perm_0, x = var_454_cast_fp16)[name = tensor<string, []>("transpose_6")];
+            tensor<fp16, [32, 4, 96, 32]> v_cast_fp16 = reshape(shape = var_472, x = v_15_cast_fp16)[name = tensor<string, []>("v_cast_fp16")];
+            tensor<fp16, []> mul_5_y_0_to_fp16 = const()[name = tensor<string, []>("mul_5_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-3)];
+            tensor<fp16, [32, 4, 96, 32]> mul_5_cast_fp16 = mul(x = q_cast_fp16, y = mul_5_y_0_to_fp16)[name = tensor<string, []>("mul_5_cast_fp16")];
+            tensor<bool, []> matmul_2_transpose_y_0 = const()[name = tensor<string, []>("matmul_2_transpose_y_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> matmul_2_transpose_x_0 = const()[name = tensor<string, []>("matmul_2_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [32, 4, 96, 96]> matmul_2_cast_fp16 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = mul_5_cast_fp16, y = k_cast_fp16)[name = tensor<string, []>("matmul_2_cast_fp16")];
+            tensor<fp16, [32, 4, 96, 96]> add_2_cast_fp16 = add(x = matmul_2_cast_fp16, y = var_460_cast_fp16)[name = tensor<string, []>("add_2_cast_fp16")];
+            tensor<int32, []> softmax_2_axis_0 = const()[name = tensor<string, []>("softmax_2_axis_0"), val = tensor<int32, []>(-1)];
+            tensor<fp16, [32, 4, 96, 96]> softmax_2_cast_fp16 = softmax(axis = softmax_2_axis_0, x = add_2_cast_fp16)[name = tensor<string, []>("softmax_2_cast_fp16")];
+            tensor<bool, []> attn_output_17_transpose_x_0 = const()[name = tensor<string, []>("attn_output_17_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attn_output_17_transpose_y_0 = const()[name = tensor<string, []>("attn_output_17_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [32, 4, 96, 32]> attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_0, transpose_y = attn_output_17_transpose_y_0, x = softmax_2_cast_fp16, y = v_cast_fp16)[name = tensor<string, []>("attn_output_17_cast_fp16")];
+            tensor<int32, [4]> var_475 = const()[name = tensor<string, []>("op_475"), val = tensor<int32, [4]>([2, 0, 1, 3])];
+            tensor<int32, [2]> var_480 = const()[name = tensor<string, []>("op_480"), val = tensor<int32, [2]>([3072, 128])];
+            tensor<fp16, [96, 32, 4, 32]> var_476_cast_fp16 = transpose(perm = var_475, x = attn_output_17_cast_fp16)[name = tensor<string, []>("transpose_5")];
+            tensor<fp16, [3072, 128]> attn_output_19_cast_fp16 = reshape(shape = var_480, x = var_476_cast_fp16)[name = tensor<string, []>("attn_output_19_cast_fp16")];
+            tensor<fp16, [128, 128]> model_option_encoder_layers_0_self_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_option_encoder_layers_0_self_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [128, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1049152)))];
+            tensor<fp16, [128]> model_option_encoder_layers_0_self_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("model_option_encoder_layers_0_self_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1081984)))];
+            tensor<fp16, [3072, 128]> linear_9_cast_fp16 = linear(bias = model_option_encoder_layers_0_self_attn_out_proj_bias_to_fp16, weight = model_option_encoder_layers_0_self_attn_out_proj_weight_to_fp16, x = attn_output_19_cast_fp16)[name = tensor<string, []>("linear_9_cast_fp16")];
+            tensor<int32, [3]> var_484 = const()[name = tensor<string, []>("op_484"), val = tensor<int32, [3]>([96, 32, 128])];
+            tensor<fp16, [96, 32, 128]> attn_output_cast_fp16 = reshape(shape = var_484, x = linear_9_cast_fp16)[name = tensor<string, []>("attn_output_cast_fp16")];
+            tensor<int32, [3]> input_35_perm_0 = const()[name = tensor<string, []>("input_35_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<fp16, [32, 96, 128]> input_35_cast_fp16 = transpose(perm = input_35_perm_0, x = attn_output_cast_fp16)[name = tensor<string, []>("transpose_4")];
+            tensor<fp16, [32, 96, 128]> input_37_cast_fp16 = add(x = src_cast_fp16, y = input_35_cast_fp16)[name = tensor<string, []>("input_37_cast_fp16")];
+            tensor<int32, [1]> input_39_axes_0 = const()[name = tensor<string, []>("input_39_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [128]> model_option_encoder_layers_0_norm2_weight_to_fp16 = const()[name = tensor<string, []>("model_option_encoder_layers_0_norm2_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1082304)))];
+            tensor<fp16, [128]> model_option_encoder_layers_0_norm2_bias_to_fp16 = const()[name = tensor<string, []>("model_option_encoder_layers_0_norm2_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1082624)))];
+            tensor<fp16, [32, 96, 128]> input_39_cast_fp16 = layer_norm(axes = input_39_axes_0, beta = model_option_encoder_layers_0_norm2_bias_to_fp16, epsilon = var_375_to_fp16, gamma = model_option_encoder_layers_0_norm2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor<string, []>("input_39_cast_fp16")];
+            tensor<fp16, [512, 128]> model_option_encoder_layers_0_linear1_weight_to_fp16 = const()[name = tensor<string, []>("model_option_encoder_layers_0_linear1_weight_to_fp16"), val = tensor<fp16, [512, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1082944)))];
+            tensor<fp16, [512]> model_option_encoder_layers_0_linear1_bias_to_fp16 = const()[name = tensor<string, []>("model_option_encoder_layers_0_linear1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1214080)))];
+            tensor<fp16, [32, 96, 512]> linear_10_cast_fp16 = linear(bias = model_option_encoder_layers_0_linear1_bias_to_fp16, weight = model_option_encoder_layers_0_linear1_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("linear_10_cast_fp16")];
+            tensor<fp16, [32, 96, 512]> input_43_cast_fp16 = relu(x = linear_10_cast_fp16)[name = tensor<string, []>("input_43_cast_fp16")];
+            tensor<fp16, [128, 512]> model_option_encoder_layers_0_linear2_weight_to_fp16 = const()[name = tensor<string, []>("model_option_encoder_layers_0_linear2_weight_to_fp16"), val = tensor<fp16, [128, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1215168)))];
+            tensor<fp16, [128]> model_option_encoder_layers_0_linear2_bias_to_fp16 = const()[name = tensor<string, []>("model_option_encoder_layers_0_linear2_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1346304)))];
+            tensor<fp16, [32, 96, 128]> linear_11_cast_fp16 = linear(bias = model_option_encoder_layers_0_linear2_bias_to_fp16, weight = model_option_encoder_layers_0_linear2_weight_to_fp16, x = input_43_cast_fp16)[name = tensor<string, []>("linear_11_cast_fp16")];
+            tensor<fp16, [32, 96, 128]> hidden_cast_fp16 = add(x = input_37_cast_fp16, y = linear_11_cast_fp16)[name = tensor<string, []>("hidden_cast_fp16")];
+            tensor<int32, [1]> var_504_axes_0 = const()[name = tensor<string, []>("op_504_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, [32, 96, 1]> var_504 = expand_dims(axes = var_504_axes_0, x = flat_mask)[name = tensor<string, []>("op_504")];
+            tensor<string, []> weights_to_fp16_dtype_0 = const()[name = tensor<string, []>("weights_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [32, 96, 1]> var_504_to_fp16 = cast(dtype = weights_to_fp16_dtype_0, x = var_504)[name = tensor<string, []>("cast_58")];
+            tensor<fp16, [32, 96, 128]> var_510_cast_fp16 = mul(x = hidden_cast_fp16, y = var_504_to_fp16)[name = tensor<string, []>("op_510_cast_fp16")];
+            tensor<int32, [1]> var_515_axes_0 = const()[name = tensor<string, []>("op_515_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, []> var_515_keep_dims_0 = const()[name = tensor<string, []>("op_515_keep_dims_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [32, 128]> var_515_cast_fp16 = reduce_sum(axes = var_515_axes_0, keep_dims = var_515_keep_dims_0, x = var_510_cast_fp16)[name = tensor<string, []>("op_515_cast_fp16")];
+            tensor<int32, [1]> var_520_axes_0 = const()[name = tensor<string, []>("op_520_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, []> var_520_keep_dims_0 = const()[name = tensor<string, []>("op_520_keep_dims_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [32, 1]> var_520_cast_fp16 = reduce_sum(axes = var_520_axes_0, keep_dims = var_520_keep_dims_0, x = var_504_to_fp16)[name = tensor<string, []>("op_520_cast_fp16")];
+            tensor<fp16, []> var_521_to_fp16 = const()[name = tensor<string, []>("op_521_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [32, 1]> var_522_cast_fp16 = maximum(x = var_520_cast_fp16, y = var_521_to_fp16)[name = tensor<string, []>("op_522_cast_fp16")];
+            tensor<fp16, [32, 128]> pooled_cast_fp16 = real_div(x = var_515_cast_fp16, y = var_522_cast_fp16)[name = tensor<string, []>("pooled_cast_fp16")];
+            tensor<int32, [3]> var_525 = const()[name = tensor<string, []>("op_525"), val = tensor<int32, [3]>([1, 32, -1])];
+            tensor<fp16, [1, 32, 128]> options_cast_fp16 = reshape(shape = var_525, x = pooled_cast_fp16)[name = tensor<string, []>("options_cast_fp16")];
+            tensor<int32, []> var_528 = const()[name = tensor<string, []>("op_528"), val = tensor<int32, []>(-1)];
+            tensor<int32, [1]> input_axes_0 = const()[name = tensor<string, []>("input_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [128]> model_head_context_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_head_context_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1346624)))];
+            tensor<fp16, [128]> model_head_context_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_head_context_norm_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1346944)))];
+            tensor<fp16, []> var_536_to_fp16 = const()[name = tensor<string, []>("op_536_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 224, 128]> input_cast_fp16 = layer_norm(axes = input_axes_0, beta = model_head_context_norm_bias_to_fp16, epsilon = var_536_to_fp16, gamma = model_head_context_norm_weight_to_fp16, x = context_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
+            tensor<int32, [1]> input_53_axes_0 = const()[name = tensor<string, []>("input_53_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [128]> model_head_option_norm_weight_to_fp16 = const()[name = tensor<string, []>("model_head_option_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1347264)))];
+            tensor<fp16, [128]> model_head_option_norm_bias_to_fp16 = const()[name = tensor<string, []>("model_head_option_norm_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1347584)))];
+            tensor<fp16, [1, 32, 128]> input_53_cast_fp16 = layer_norm(axes = input_53_axes_0, beta = model_head_option_norm_bias_to_fp16, epsilon = var_536_to_fp16, gamma = model_head_option_norm_weight_to_fp16, x = options_cast_fp16)[name = tensor<string, []>("input_53_cast_fp16")];
+            tensor<fp16, [128, 128]> model_head_query_weight_to_fp16 = const()[name = tensor<string, []>("model_head_query_weight_to_fp16"), val = tensor<fp16, [128, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1347904)))];
+            tensor<fp16, [128]> linear_12_bias_0_to_fp16 = const()[name = tensor<string, []>("linear_12_bias_0_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1380736)))];
+            tensor<fp16, [1, 32, 128]> linear_12_cast_fp16 = linear(bias = linear_12_bias_0_to_fp16, weight = model_head_query_weight_to_fp16, x = input_53_cast_fp16)[name = tensor<string, []>("linear_12_cast_fp16")];
+            tensor<fp16, [128, 128]> model_head_key_weight_to_fp16 = const()[name = tensor<string, []>("model_head_key_weight_to_fp16"), val = tensor<fp16, [128, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1381056)))];
+            tensor<fp16, [1, 224, 128]> linear_13_cast_fp16 = linear(bias = linear_12_bias_0_to_fp16, weight = model_head_key_weight_to_fp16, x = input_cast_fp16)[name = tensor<string, []>("linear_13_cast_fp16")];
+            tensor<fp16, [128, 128]> model_head_value_weight_to_fp16 = const()[name = tensor<string, []>("model_head_value_weight_to_fp16"), val = tensor<fp16, [128, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1413888)))];
+            tensor<fp16, [1, 224, 128]> linear_14_cast_fp16 = linear(bias = linear_12_bias_0_to_fp16, weight = model_head_value_weight_to_fp16, x = input_cast_fp16)[name = tensor<string, []>("linear_14_cast_fp16")];
+            tensor<bool, []> matmul_3_transpose_x_1 = const()[name = tensor<string, []>("matmul_3_transpose_x_1"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matmul_3_transpose_y_1 = const()[name = tensor<string, []>("matmul_3_transpose_y_1"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 32, 224]> matmul_3_cast_fp16 = matmul(transpose_x = matmul_3_transpose_x_1, transpose_y = matmul_3_transpose_y_1, x = linear_12_cast_fp16, y = linear_13_cast_fp16)[name = tensor<string, []>("matmul_3_cast_fp16")];
+            tensor<fp16, []> _inversed_scores_1_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_scores_1_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 32, 224]> _inversed_scores_1_cast_fp16 = mul(x = matmul_3_cast_fp16, y = _inversed_scores_1_y_0_to_fp16)[name = tensor<string, []>("_inversed_scores_1_cast_fp16")];
+            tensor<int32, [1]> var_567_axes_0 = const()[name = tensor<string, []>("op_567_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, 224]> var_567 = expand_dims(axes = var_567_axes_0, x = context_mask)[name = tensor<string, []>("op_567")];
+            tensor<bool, [1, 1, 224]> var_569 = logical_not(x = var_567)[name = tensor<string, []>("op_569")];
+            tensor<fp16, []> var_529_to_fp16 = const()[name = tensor<string, []>("op_529_to_fp16"), val = tensor<fp16, []>(-inf)];
+            tensor<fp16, [1, 32, 224]> scores_cast_fp16 = select(a = var_529_to_fp16, b = _inversed_scores_1_cast_fp16, cond = var_569)[name = tensor<string, []>("scores_cast_fp16")];
+            tensor<fp16, [1, 32, 224]> var_571_cast_fp16 = softmax(axis = var_528, x = scores_cast_fp16)[name = tensor<string, []>("op_571_cast_fp16")];
+            tensor<bool, []> matmul_4_transpose_x_0 = const()[name = tensor<string, []>("matmul_4_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matmul_4_transpose_y_0 = const()[name = tensor<string, []>("matmul_4_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 32, 128]> matmul_4_cast_fp16 = matmul(transpose_x = matmul_4_transpose_x_0, transpose_y = matmul_4_transpose_y_0, x = var_571_cast_fp16, y = linear_14_cast_fp16)[name = tensor<string, []>("matmul_4_cast_fp16")];
+            tensor<fp16, [1, 32, 128]> var_574_cast_fp16 = mul(x = linear_12_cast_fp16, y = matmul_4_cast_fp16)[name = tensor<string, []>("op_574_cast_fp16")];
+            tensor<int32, [1]> var_576_axes_0 = const()[name = tensor<string, []>("op_576_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, []> var_576_keep_dims_0 = const()[name = tensor<string, []>("op_576_keep_dims_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 32]> var_576_cast_fp16 = reduce_sum(axes = var_576_axes_0, keep_dims = var_576_keep_dims_0, x = var_574_cast_fp16)[name = tensor<string, []>("op_576_cast_fp16")];
+            tensor<fp16, []> _inversed_logits_1_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_logits_1_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
+            tensor<fp16, [1, 32]> _inversed_logits_1_cast_fp16 = mul(x = var_576_cast_fp16, y = _inversed_logits_1_y_0_to_fp16)[name = tensor<string, []>("_inversed_logits_1_cast_fp16")];
+            tensor<bool, [1, 32]> var_579 = logical_not(x = option_mask_1)[name = tensor<string, []>("op_579")];
+            tensor<fp16, [1, 32]> logits_3_cast_fp16 = select(a = var_529_to_fp16, b = _inversed_logits_1_cast_fp16, cond = var_579)[name = tensor<string, []>("logits_3_cast_fp16")];
+            tensor<fp16, []> var_587_value_0_to_fp16 = const()[name = tensor<string, []>("op_587_value_0_to_fp16"), val = tensor<fp16, []>(-0x1.388p+13)];
+            tensor<fp16, [1, 32]> var_587_cast_fp16 = fill_like(ref_tensor = logits_3_cast_fp16, value = var_587_value_0_to_fp16)[name = tensor<string, []>("op_587_cast_fp16")];
+            tensor<fp16, [1, 32]> logits_cast_fp16 = select(a = logits_3_cast_fp16, b = var_587_cast_fp16, cond = option_mask_1)[name = tensor<string, []>("logits_cast_fp16")];
+            tensor<string, []> logits_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("logits_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, []> var_589 = const()[name = tensor<string, []>("op_589"), val = tensor<int32, []>(-1)];
+            tensor<fp16, [1, 32]> var_591_cast_fp16 = softmax(axis = var_589, x = logits_cast_fp16)[name = tensor<string, []>("op_591_cast_fp16")];
+            tensor<string, []> var_591_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_591_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<fp32, [1, 32]> probabilities = cast(dtype = var_591_cast_fp16_to_fp32_dtype_0, x = var_591_cast_fp16)[name = tensor<string, []>("cast_56")];
+            tensor<fp32, [1, 32]> logits = cast(dtype = logits_cast_fp16_to_fp32_dtype_0, x = logits_cast_fp16)[name = tensor<string, []>("cast_57")];
+        } -> (logits, probabilities);
+}

cua_s1_forms_fp16_options32.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4da9259f798e44f5a1b50769ee1916fd3747c4d723dd9997b516c7fe238c7895
+size 1446720

cua_s1_forms_fp16_options32.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:70485fc18cbb21785df833cbddddc0b5b59acb00d22394b76e55307e2c135dd0
+size 63826

cua_s1_forms_fp16_options32.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4da9259f798e44f5a1b50769ee1916fd3747c4d723dd9997b516c7fe238c7895
+size 1446720

cua_s1_forms_fp16_options32.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "A4ADBA80-1EEA-488D-A01F-25F39ECC9001": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "B61C4C42-FF5A-4E18-BC04-DD11C9CE38DE": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "B61C4C42-FF5A-4E18-BC04-DD11C9CE38DE"
+}

preprocessing.py

@@ -0,0 +1,46 @@
+"""Host-side UTF-8 byte encoding for the fixed-shape Core ML interface."""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+
+import numpy as np
+
+
+@dataclass(frozen=True)
+class InputLimits:
+    context_bytes: int = 224
+    option_bytes: int = 96
+    max_options: int = 32
+
+    def __post_init__(self):
+        if self.context_bytes < 1 or self.option_bytes < 1 or self.max_options < 2:
+            raise ValueError("Positive byte limits and at least two option slots are required")
+
+
+def byte_ids(text: str, length: int) -> np.ndarray:
+    """Match upstream: truncate UTF-8 bytes, offset byte values by one, zero-pad."""
+    data = text.encode("utf-8", errors="replace")[:length]
+    return np.frombuffer(data, dtype=np.uint8).astype(np.int32) + 1
+
+
+def prepare_inputs(context: str, options: list[str] | tuple[str, ...], limits: InputLimits) -> dict[str, np.ndarray]:
+    """Encode one decision; options are never silently removed to fit the model."""
+    if not isinstance(context, str) or not context:
+        raise ValueError("context must be a nonempty string")
+    if not isinstance(options, (list, tuple)) or len(options) < 2:
+        raise ValueError("At least two options are required")
+    if len(options) > limits.max_options:
+        raise ValueError(f"{len(options)} options exceed {limits.max_options}; re-export with a larger --max-options")
+    if any(not isinstance(option, str) or not option for option in options):
+        raise ValueError("Options must be nonempty strings")
+    context_ids = np.zeros((1, limits.context_bytes), dtype=np.int32)
+    option_ids = np.zeros((1, limits.max_options, limits.option_bytes), dtype=np.int32)
+    option_mask = np.zeros((1, limits.max_options), dtype=np.int32)
+    tokens = byte_ids(context, limits.context_bytes)
+    context_ids[0, : len(tokens)] = tokens
+    for index, option in enumerate(options):
+        tokens = byte_ids(option, limits.option_bytes)
+        option_ids[0, index, : len(tokens)] = tokens
+    option_mask[0, : len(options)] = 1
+    return {"context_ids": context_ids, "option_ids": option_ids, "option_mask": option_mask}

reports/ane-fallback.json

@@ -0,0 +1,84 @@
+{
+  "hardware": {
+    "device": "arm64",
+    "chip": "Apple M5 Pro",
+    "ram": "24GB",
+    "os_version": "macOS 27.0",
+    "timestamp": "2026-09-19T16:51:27.323329+00:00"
+  },
+  "models": [
+    {
+      "model_path": "build/cua_s1_forms_fp16_options32.mlmodelc",
+      "model_name": "cua_s1_forms_fp16_options32",
+      "fallback": {
+        "compute_units": "cpu_and_neural_engine",
+        "total_ops": 173,
+        "ane_ops": 149,
+        "gpu_ops": 0,
+        "cpu_ops": 24,
+        "ane_percent": 86.1,
+        "reasons": [
+          {
+            "reason": "Unsupported tensor data type: int32",
+            "count": 17,
+            "estimated_cpu_runtime_ms": 0.017,
+            "op_types": {
+              "ios17.cast": 8,
+              "ios17.slice_by_index": 4,
+              "ios17.add": 2,
+              "select": 2,
+              "ios17.reshape": 1
+            },
+            "ops": [
+              "var_27",
+              "var_43",
+              "var_43_to_fp16",
+              "context_ids_to_int16",
+              "cast_54",
+              "add_3",
+              "select_0",
+              "select_0_to_int16",
+              "flat_ids",
+              "var_333",
+              "var_349",
+              "var_349_to_fp16",
+              "flat_ids_to_int16",
+              "cast_55",
+              "add_4",
+              "select_1",
+              "select_1_to_int16"
+            ]
+          },
+          {
+            "reason": "Unable to resolve operation input \"y\".",
+            "count": 5,
+            "estimated_cpu_runtime_ms": 0.005,
+            "op_types": {
+              "ios17.not_equal": 3,
+              "ios17.greater_equal": 2
+            },
+            "ops": [
+              "option_mask_1",
+              "context_mask",
+              "greater_equal_0",
+              "flat_mask",
+              "greater_equal_1"
+            ]
+          },
+          {
+            "reason": "Unsupported gather index type",
+            "count": 2,
+            "estimated_cpu_runtime_ms": 0.0144,
+            "op_types": {
+              "ios17.gather": 2
+            },
+            "ops": [
+              "var_59_cast_fp16_cast_uint16_cast_uint16",
+              "var_365_cast_fp16_cast_uint16_cast_uint16"
+            ]
+          }
+        ]
+      }
+    }
+  ]
+}

reports/upstream-metrics.json

@@ -0,0 +1,128 @@
+{
+  "verification_report_sha256": "9a5d22ec4b447f0710c12e03c69d7ed802158bb94e9a3c6b059731730ced9f30",
+  "evaluator_source_revision": "83f142c4290a0f7d9ed545ae8532858c6e4f8145",
+  "evaluator_source_path": "libs/cua-s1/evals/metrics.py",
+  "evaluator_sha256": "49dccdeb4e2b456187cdc4491a6d78fabac6a45384272f8e3ec802a58872aa6e",
+  "adapter_sha256": "c78e45caa86fb59f9a56ef18cc6a35020ea6e281a6a23d23618f56da85a0d95e",
+  "scope": "Offline action/target classification, not GUI execution or task completion",
+  "normalization": "Stable row IDs; fill target is entity option index; upstream treats skip as abstention",
+  "results": {
+    "ALL": {
+      "examples": 196,
+      "accuracy": 1.0,
+      "coverage": 0.23469387755102042,
+      "abstention_rate": 0.7653061224489796,
+      "selective_accuracy": 1.0,
+      "wrong_action_rate": 0.0,
+      "wrong_target_rate": 0.0,
+      "unsafe_action_rate": 0.0,
+      "counts": {
+        "total": 196,
+        "correct": 196,
+        "abstained": 150,
+        "acted": 46,
+        "acted_correct": 46
+      },
+      "per_action": {
+        "abstain": {
+          "accuracy": 1.0,
+          "abstention_rate": 1.0,
+          "examples": 150
+        },
+        "check": {
+          "accuracy": 1.0,
+          "abstention_rate": 0.0,
+          "examples": 4
+        },
+        "click": {
+          "accuracy": 1.0,
+          "abstention_rate": 0.0,
+          "examples": 6
+        },
+        "fill": {
+          "accuracy": 1.0,
+          "abstention_rate": 0.0,
+          "examples": 36
+        }
+      }
+    },
+    "CPU_AND_NE": {
+      "examples": 196,
+      "accuracy": 1.0,
+      "coverage": 0.23469387755102042,
+      "abstention_rate": 0.7653061224489796,
+      "selective_accuracy": 1.0,
+      "wrong_action_rate": 0.0,
+      "wrong_target_rate": 0.0,
+      "unsafe_action_rate": 0.0,
+      "counts": {
+        "total": 196,
+        "correct": 196,
+        "abstained": 150,
+        "acted": 46,
+        "acted_correct": 46
+      },
+      "per_action": {
+        "abstain": {
+          "accuracy": 1.0,
+          "abstention_rate": 1.0,
+          "examples": 150
+        },
+        "check": {
+          "accuracy": 1.0,
+          "abstention_rate": 0.0,
+          "examples": 4
+        },
+        "click": {
+          "accuracy": 1.0,
+          "abstention_rate": 0.0,
+          "examples": 6
+        },
+        "fill": {
+          "accuracy": 1.0,
+          "abstention_rate": 0.0,
+          "examples": 36
+        }
+      }
+    },
+    "upstream_pytorch": {
+      "examples": 196,
+      "accuracy": 1.0,
+      "coverage": 0.23469387755102042,
+      "abstention_rate": 0.7653061224489796,
+      "selective_accuracy": 1.0,
+      "wrong_action_rate": 0.0,
+      "wrong_target_rate": 0.0,
+      "unsafe_action_rate": 0.0,
+      "counts": {
+        "total": 196,
+        "correct": 196,
+        "abstained": 150,
+        "acted": 46,
+        "acted_correct": 46
+      },
+      "per_action": {
+        "abstain": {
+          "accuracy": 1.0,
+          "abstention_rate": 1.0,
+          "examples": 150
+        },
+        "check": {
+          "accuracy": 1.0,
+          "abstention_rate": 0.0,
+          "examples": 4
+        },
+        "click": {
+          "accuracy": 1.0,
+          "abstention_rate": 0.0,
+          "examples": 6
+        },
+        "fill": {
+          "accuracy": 1.0,
+          "abstention_rate": 0.0,
+          "examples": 36
+        }
+      }
+    }
+  }
+}

reports/verification.json

@@ -0,0 +1,3309 @@
+{
+  "purpose": "Local conversion parity on the upstream demo; not a generalization or live GUI benchmark",
+  "created_utc": "2026-09-19T16:49:43.157197+00:00",
+  "model_revision": "f54adbf447f4ca6ec259f529ee3f2e3e09f8cc71",
+  "source_revision": "83f142c4290a0f7d9ed545ae8532858c6e4f8145",
+  "dataset_revision": "8273f34778b99ac2e12d9f6e7d57dad99ae20845",
+  "dataset_file": "demo.jsonl",
+  "dataset_sha256": "4f43b442e79ba2e2ce731e27e9b8e340c2b5dfcaffc92d8ff564c34f115ff1ca",
+  "conversion": {
+    "model": "cua_s1_forms_fp16_options32.mlpackage",
+    "precision": "float16",
+    "minimum_target": "iOS17/macOS14",
+    "limits": {
+      "context_bytes": 224,
+      "option_bytes": 96,
+      "max_options": 32
+    },
+    "model_config": {
+      "context_tokens": 224,
+      "encoder": "tinyx",
+      "heads": 4,
+      "hf_model": "Qwen/Qwen2.5-0.5B",
+      "layers": 2,
+      "option_tokens": 96,
+      "rank": 128,
+      "width": 128
+    },
+    "parameters": 706048,
+    "assets_lock_sha256": "8e65ad70af6bb814b571cdcfe828ba4bc339147da2d2211cbeac416163ef18ba",
+    "model_revision": "f54adbf447f4ca6ec259f529ee3f2e3e09f8cc71",
+    "source_revision": "83f142c4290a0f7d9ed545ae8532858c6e4f8145",
+    "trace_row": 0,
+    "trace_dataset_revision": "8273f34778b99ac2e12d9f6e7d57dad99ae20845",
+    "export_seconds": 0.5374167499830946,
+    "python": "3.11.11",
+    "torch": "2.7.0",
+    "coremltools": "9.0",
+    "package_files": {
+      "Data/com.apple.CoreML/model.mlmodel": "70485fc18cbb21785df833cbddddc0b5b59acb00d22394b76e55307e2c135dd0",
+      "Data/com.apple.CoreML/weights/weight.bin": "4da9259f798e44f5a1b50769ee1916fd3747c4d723dd9997b516c7fe238c7895",
+      "Manifest.json": "2bc0f5f62337b27fb6b0ecde248f1e3dc269e1ba4b65516aaeede2a60e293dcc"
+    }
+  },
+  "environment": {
+    "machine": "Apple M5 Pro",
+    "os": "27.0",
+    "python": "3.11.11",
+    "torch": "2.7.0",
+    "coremltools": "9.0",
+    "torch_cpu_threads": 2
+  },
+  "rows": 196,
+  "upstream": {
+    "correct": 196,
+    "accuracy": 1.0,
+    "inference_latency_ms": {
+      "count": 196,
+      "median_ms": 2.4375420034630224,
+      "p95_ms": 3.0990312443464063,
+      "min_ms": 1.6581250238232315,
+      "max_ms": 3.908416983904317
+    }
+  },
+  "export_adapter_fp32": {
+    "argmax_agreement": 196,
+    "max_abs_probability_error": 1.1324882507324219e-06
+  },
+  "host_preprocessing_latency_ms": {
+    "count": 196,
+    "median_ms": 0.03766651207115501,
+    "p95_ms": 0.04735450784210116,
+    "min_ms": 0.026416993932798505,
+    "max_ms": 0.07812498370185494
+  },
+  "thresholds": {
+    "argmax_agreement": 1.0,
+    "max_abs_probability_error": 0.005,
+    "export_adapter_max_abs_probability_error": 0.0001,
+    "allow_accuracy_loss": false
+  },
+  "backends": {
+    "ALL": {
+      "passed": true,
+      "correct": 196,
+      "accuracy": 1.0,
+      "argmax_agreement": 196,
+      "max_abs_probability_error": 0.0030989646911621094,
+      "mean_row_max_abs_probability_error": 1.6251600213152587e-05,
+      "max_abs_logit_error": 0.03653383255004883,
+      "load_seconds": 0.1422372910019476,
+      "warm_inference_latency_ms": {
+        "count": 196,
+        "median_ms": 1.8487705092411488,
+        "p95_ms": 2.4909165003919043,
+        "min_ms": 1.7662910104263574,
+        "max_ms": 2.893916011089459
+      },
+      "reversed_option_order": {
+        "rows": 6,
+        "max_abs_probability_error": 5.26619260199368e-07
+      },
+      "per_action": {
+        "fill": {
+          "rows": 36,
+          "correct": 36
+        },
+        "skip": {
+          "rows": 150,
+          "correct": 150
+        },
+        "check": {
+          "rows": 4,
+          "correct": 4
+        },
+        "click": {
+          "rows": 6,
+          "correct": 6
+        }
+      },
+      "decisions": [
+        {
+          "row": 0,
+          "label": 22,
+          "upstream": 22,
+          "coreml": 22,
+          "max_abs_probability_error": 5.26619260199368e-07,
+          "max_abs_logit_error": 0.017948150634765625
+        },
+        {
+          "row": 1,
+          "label": 23,
+          "upstream": 23,
+          "coreml": 23,
+          "max_abs_probability_error": 5.320471245795488e-07,
+          "max_abs_logit_error": 0.0173797607421875
+        },
+        {
+          "row": 2,
+          "label": 2,
+          "upstream": 2,
+          "coreml": 2,
+          "max_abs_probability_error": 3.134431381113245e-09,
+          "max_abs_logit_error": 0.02239227294921875
+        },
+        {
+          "row": 3,
+          "label": 5,
+          "upstream": 5,
+          "coreml": 5,
+          "max_abs_probability_error": 2.8436625143513083e-07,
+          "max_abs_logit_error": 0.014525413513183594
+        },
+        {
+          "row": 4,
+          "label": 4,
+          "upstream": 4,
+          "coreml": 4,
+          "max_abs_probability_error": 4.76837158203125e-07,
+          "max_abs_logit_error": 0.013245582580566406
+        },
+        {
+          "row": 5,
+          "label": 6,
+          "upstream": 6,
+          "coreml": 6,
+          "max_abs_probability_error": 3.197783371433616e-08,
+          "max_abs_logit_error": 0.01342916488647461
+        },
+        {
+          "row": 6,
+          "label": 7,
+          "upstream": 7,
+          "coreml": 7,
+          "max_abs_probability_error": 1.7993625078815967e-07,
+          "max_abs_logit_error": 0.011149883270263672
+        },
+        {
+          "row": 7,
+          "label": 8,
+          "upstream": 8,
+          "coreml": 8,
+          "max_abs_probability_error": 6.677582859992981e-07,
+          "max_abs_logit_error": 0.013294458389282227
+        },
+        {
+          "row": 8,
+          "label": 9,
+          "upstream": 9,
+          "coreml": 9,
+          "max_abs_probability_error": 4.183561941317748e-09,
+          "max_abs_logit_error": 0.013601541519165039
+        },
+        {
+          "row": 9,
+          "label": 10,
+          "upstream": 10,
+          "coreml": 10,
+          "max_abs_probability_error": 4.019384505227208e-07,
+          "max_abs_logit_error": 0.029033660888671875
+        },
+        {
+          "row": 10,
+          "label": 11,
+          "upstream": 11,
+          "coreml": 11,
+          "max_abs_probability_error": 2.5259510039177258e-09,
+          "max_abs_logit_error": 0.03653383255004883
+        },
+        {
+          "row": 11,
+          "label": 13,
+          "upstream": 13,
+          "coreml": 13,
+          "max_abs_probability_error": 1.1920928955078125e-07,
+          "max_abs_logit_error": 0.026096343994140625
+        },
+        {
+          "row": 12,
+          "label": 15,
+          "upstream": 15,
+          "coreml": 15,
+          "max_abs_probability_error": 4.578259904519655e-08,
+          "max_abs_logit_error": 0.029152870178222656
+        },
+        {
+          "row": 13,
+          "label": 16,
+          "upstream": 16,
+          "coreml": 16,
+          "max_abs_probability_error": 8.465372957289219e-06,
+          "max_abs_logit_error": 0.01841115951538086
+        },
+        {
+          "row": 14,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 1.1920928955078125e-07,
+          "max_abs_logit_error": 0.020061016082763672
+        },
+        {
+          "row": 15,
+          "label": 24,
+          "upstream": 24,
+          "coreml": 24,
+          "max_abs_probability_error": 8.867027645464987e-08,
+          "max_abs_logit_error": 0.013807952404022217
+        },
+        {
+          "row": 16,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 5.60369528557203e-10,
+          "max_abs_logit_error": 0.011376380920410156
+        },
+        {
+          "row": 17,
+          "label": 25,
+          "upstream": 25,
+          "coreml": 25,
+          "max_abs_probability_error": 3.9392489270539954e-11,
+          "max_abs_logit_error": 0.02391815185546875
+        },
+        {
+          "row": 18,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 5.761169319384862e-11,
+          "max_abs_logit_error": 0.008268356323242188
+        },
+        {
+          "row": 19,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 3.685274307940745e-12,
+          "max_abs_logit_error": 0.012112617492675781
+        },
+        {
+          "row": 20,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 3.197389020215269e-10,
+          "max_abs_logit_error": 0.012657880783081055
+        },
+        {
+          "row": 21,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 1.3100631690576847e-13,
+          "max_abs_logit_error": 0.02386617660522461
+        },
+        {
+          "row": 22,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 2.716319613682572e-09,
+          "max_abs_logit_error": 0.027971267700195312
+        },
+        {
+          "row": 23,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 8.720178357179975e-09,
+          "max_abs_logit_error": 0.026054739952087402
+        },
+        {
+          "row": 24,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 1.5336516778763354e-13,
+          "max_abs_logit_error": 0.012665033340454102
+        },
+        {
+          "row": 25,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 9.446710080851517e-11,
+          "max_abs_logit_error": 0.014183402061462402
+        },
+        {
+          "row": 26,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 4.373890138964498e-12,
+          "max_abs_logit_error": 0.009427070617675781
+        },
+        {
+          "row": 27,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 4.2128106558791956e-13,
+          "max_abs_logit_error": 0.016583919525146484
+        },
+        {
+          "row": 28,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 1.1457235160605705e-10,
+          "max_abs_logit_error": 0.012380242347717285
+        },
+        {
+          "row": 29,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 1.6885971787505838e-13,
+          "max_abs_logit_error": 0.01409292221069336
+        },
+        {
+          "row": 30,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 3.038820750589366e-09,
+          "max_abs_logit_error": 0.023344039916992188
+        },
+        {
+          "row": 31,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 2.535194276731545e-12,
+          "max_abs_logit_error": 0.019451618194580078
+        },
+        {
+          "row": 32,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 1.1524559084818975e-11,
+          "max_abs_logit_error": 0.016178369522094727
+        },
+        {
+          "row": 33,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 7.133360568900571e-11,
+          "max_abs_logit_error": 0.016007661819458008
+        },
+        {
+          "row": 34,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 8.463859558105469e-06,
+          "max_abs_logit_error": 0.021859407424926758
+        },
+        {
+          "row": 35,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 6.966427434917932e-12,
+          "max_abs_logit_error": 0.012809276580810547
+        },
+        {
+          "row": 36,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 3.978734008924789e-12,
+          "max_abs_logit_error": 0.011217594146728516
+        },
+        {
+          "row": 37,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 3.058104880437895e-10,
+          "max_abs_logit_error": 0.015135049819946289
+        },
+        {
+          "row": 38,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 1.1240661179634515e-12,
+          "max_abs_logit_error": 0.013502120971679688
+        },
+        {
+          "row": 39,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 3.3960612100258913e-12,
+          "max_abs_logit_error": 0.015411376953125
+        },
+        {
+          "row": 40,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 3.713369369506836e-05,
+          "max_abs_logit_error": 0.02988719940185547
+        },
+        {
+          "row": 41,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 1.836255592024827e-10,
+          "max_abs_logit_error": 0.017060812562704086
+        },
+        {
+          "row": 42,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 2.5000002068509275e-12,
+          "max_abs_logit_error": 0.009338855743408203
+        },
+        {
+          "row": 43,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 1.031253304972779e-08,
+          "max_abs_logit_error": 0.013645172119140625
+        },
+        {
+          "row": 44,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 3.2869124089174306e-12,
+          "max_abs_logit_error": 0.014712810516357422
+        },
+        {
+          "row": 45,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 2.582760394442829e-13,
+          "max_abs_logit_error": 0.01000899076461792
+        },
+        {
+          "row": 46,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 3.987421504092481e-12,
+          "max_abs_logit_error": 0.008915901184082031
+        },
+        {
+          "row": 47,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 3.5084782301630923e-13,
+          "max_abs_logit_error": 0.012367963790893555
+        },
+        {
+          "row": 48,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 1.1920928955078125e-07,
+          "max_abs_logit_error": 0.020061016082763672
+        },
+        {
+          "row": 49,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 1.312366881833782e-12,
+          "max_abs_logit_error": 0.02675628662109375
+        },
+        {
+          "row": 50,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 5.60369528557203e-10,
+          "max_abs_logit_error": 0.011376380920410156
+        },
+        {
+          "row": 51,
+          "label": 25,
+          "upstream": 25,
+          "coreml": 25,
+          "max_abs_probability_error": 3.9392489270539954e-11,
+          "max_abs_logit_error": 0.02391815185546875
+        },
+        {
+          "row": 52,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 5.761169319384862e-11,
+          "max_abs_logit_error": 0.008268356323242188
+        },
+        {
+          "row": 53,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 3.685274307940745e-12,
+          "max_abs_logit_error": 0.012112617492675781
+        },
+        {
+          "row": 54,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 3.197389020215269e-10,
+          "max_abs_logit_error": 0.012657880783081055
+        },
+        {
+          "row": 55,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 1.3100631690576847e-13,
+          "max_abs_logit_error": 0.02386617660522461
+        },
+        {
+          "row": 56,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 2.716319613682572e-09,
+          "max_abs_logit_error": 0.027971267700195312
+        },
+        {
+          "row": 57,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 8.720178357179975e-09,
+          "max_abs_logit_error": 0.026054739952087402
+        },
+        {
+          "row": 58,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 1.5336516778763354e-13,
+          "max_abs_logit_error": 0.012665033340454102
+        },
+        {
+          "row": 59,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 9.446710080851517e-11,
+          "max_abs_logit_error": 0.014183402061462402
+        },
+        {
+          "row": 60,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 4.373890138964498e-12,
+          "max_abs_logit_error": 0.009427070617675781
+        },
+        {
+          "row": 61,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 4.2128106558791956e-13,
+          "max_abs_logit_error": 0.016583919525146484
+        },
+        {
+          "row": 62,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 1.1457235160605705e-10,
+          "max_abs_logit_error": 0.012380242347717285
+        },
+        {
+          "row": 63,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 1.6885971787505838e-13,
+          "max_abs_logit_error": 0.01409292221069336
+        },
+        {
+          "row": 64,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 3.038820750589366e-09,
+          "max_abs_logit_error": 0.023344039916992188
+        },
+        {
+          "row": 65,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 2.535194276731545e-12,
+          "max_abs_logit_error": 0.019451618194580078
+        },
+        {
+          "row": 66,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 1.1524559084818975e-11,
+          "max_abs_logit_error": 0.016178369522094727
+        },
+        {
+          "row": 67,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 7.133360568900571e-11,
+          "max_abs_logit_error": 0.016007661819458008
+        },
+        {
+          "row": 68,
+          "label": 0,
+          "upstream": 0,
+          "coreml": 0,
+          "max_abs_probability_error": 5.960464477539062e-07,
+          "max_abs_logit_error": 0.02204132080078125
+        },
+        {
+          "row": 69,
+          "label": 1,
+          "upstream": 1,
+          "coreml": 1,
+          "max_abs_probability_error": 5.992624210193753e-07,
+          "max_abs_logit_error": 0.01074361801147461
+        },
+        {
+          "row": 70,
+          "label": 2,
+          "upstream": 2,
+          "coreml": 2,
+          "max_abs_probability_error": 1.6200374375330284e-09,
+          "max_abs_logit_error": 0.012182235717773438
+        },
+        {
+          "row": 71,
+          "label": 3,
+          "upstream": 3,
+          "coreml": 3,
+          "max_abs_probability_error": 7.649177860002965e-07,
+          "max_abs_logit_error": 0.021416664123535156
+        },
+        {
+          "row": 72,
+          "label": 5,
+          "upstream": 5,
+          "coreml": 5,
+          "max_abs_probability_error": 1.716807673801668e-08,
+          "max_abs_logit_error": 0.015079259872436523
+        },
+        {
+          "row": 73,
+          "label": 6,
+          "upstream": 6,
+          "coreml": 6,
+          "max_abs_probability_error": 2.384185791015625e-07,
+          "max_abs_logit_error": 0.022822856903076172
+        },
+        {
+          "row": 74,
+          "label": 7,
+          "upstream": 7,
+          "coreml": 7,
+          "max_abs_probability_error": 1.3113021850585938e-06,
+          "max_abs_logit_error": 0.02725982666015625
+        },
+        {
+          "row": 75,
+          "label": 8,
+          "upstream": 8,
+          "coreml": 8,
+          "max_abs_probability_error": 8.482402336085215e-08,
+          "max_abs_logit_error": 0.01822662353515625
+        },
+        {
+          "row": 76,
+          "label": 9,
+          "upstream": 9,
+          "coreml": 9,
+          "max_abs_probability_error": 4.19224761571968e-08,
+          "max_abs_logit_error": 0.027760028839111328
+        },
+        {
+          "row": 77,
+          "label": 13,
+          "upstream": 13,
+          "coreml": 13,
+          "max_abs_probability_error": 6.794929504394531e-06,
+          "max_abs_logit_error": 0.02509307861328125
+        },
+        {
+          "row": 78,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.1920928955078125e-07,
+          "max_abs_logit_error": 0.013262271881103516
+        },
+        {
+          "row": 79,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 7.902155630290508e-06,
+          "max_abs_logit_error": 0.019384384155273438
+        },
+        {
+          "row": 80,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 4.386720320326276e-09,
+          "max_abs_logit_error": 0.020812511444091797
+        },
+        {
+          "row": 81,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 3.5509287954482716e-08,
+          "max_abs_logit_error": 0.033596038818359375
+        },
+        {
+          "row": 82,
+          "label": 19,
+          "upstream": 19,
+          "coreml": 19,
+          "max_abs_probability_error": 1.1210374850634253e-09,
+          "max_abs_logit_error": 0.021764755249023438
+        },
+        {
+          "row": 83,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 2.9104718635153404e-11,
+          "max_abs_logit_error": 0.017470359802246094
+        },
+        {
+          "row": 84,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 3.615042487581377e-09,
+          "max_abs_logit_error": 0.011733055114746094
+        },
+        {
+          "row": 85,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 2.5568436257117355e-12,
+          "max_abs_logit_error": 0.018376827239990234
+        },
+        {
+          "row": 86,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 3.652814939414384e-09,
+          "max_abs_logit_error": 0.013185739517211914
+        },
+        {
+          "row": 87,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 8.637570658720506e-10,
+          "max_abs_logit_error": 0.010789871215820312
+        },
+        {
+          "row": 88,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 2.6595947666407938e-11,
+          "max_abs_logit_error": 0.022840499877929688
+        },
+        {
+          "row": 89,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.6561418902938385e-11,
+          "max_abs_logit_error": 0.009737014770507812
+        },
+        {
+          "row": 90,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.1621037465658901e-11,
+          "max_abs_logit_error": 0.01908111572265625
+        },
+        {
+          "row": 91,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 8.295308884243013e-13,
+          "max_abs_logit_error": 0.016066551208496094
+        },
+        {
+          "row": 92,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 2.485815372987865e-14,
+          "max_abs_logit_error": 0.01638507843017578
+        },
+        {
+          "row": 93,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 2.78640222006743e-10,
+          "max_abs_logit_error": 0.024122238159179688
+        },
+        {
+          "row": 94,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 3.6049052631881295e-11,
+          "max_abs_logit_error": 0.0227203369140625
+        },
+        {
+          "row": 95,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 7.636086207796211e-12,
+          "max_abs_logit_error": 0.021586894989013672
+        },
+        {
+          "row": 96,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 2.5383029012004954e-12,
+          "max_abs_logit_error": 0.019191741943359375
+        },
+        {
+          "row": 97,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 3.6573055695043877e-09,
+          "max_abs_logit_error": 0.016190528869628906
+        },
+        {
+          "row": 98,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 3.6824872040597256e-09,
+          "max_abs_logit_error": 0.016640663146972656
+        },
+        {
+          "row": 99,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 7.239464139274787e-09,
+          "max_abs_logit_error": 0.012330055236816406
+        },
+        {
+          "row": 100,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 9.094208053284092e-09,
+          "max_abs_logit_error": 0.024295806884765625
+        },
+        {
+          "row": 101,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 4.488498461796553e-11,
+          "max_abs_logit_error": 0.010948657989501953
+        },
+        {
+          "row": 102,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 6.78730405212491e-11,
+          "max_abs_logit_error": 0.0258026123046875
+        },
+        {
+          "row": 103,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 4.291534423828125e-06,
+          "max_abs_logit_error": 0.013744831085205078
+        },
+        {
+          "row": 104,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 4.948583409714047e-08,
+          "max_abs_logit_error": 0.02979755401611328
+        },
+        {
+          "row": 105,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 9.98279237052202e-12,
+          "max_abs_logit_error": 0.012142181396484375
+        },
+        {
+          "row": 106,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 2.4030777368011513e-11,
+          "max_abs_logit_error": 0.013262748718261719
+        },
+        {
+          "row": 107,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 2.511102437097179e-12,
+          "max_abs_logit_error": 0.018053054809570312
+        },
+        {
+          "row": 108,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 2.8072548730051494e-09,
+          "max_abs_logit_error": 0.024975299835205078
+        },
+        {
+          "row": 109,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.1920928955078125e-07,
+          "max_abs_logit_error": 0.013262271881103516
+        },
+        {
+          "row": 110,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 2.9665898182429373e-08,
+          "max_abs_logit_error": 0.02109050750732422
+        },
+        {
+          "row": 111,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.1920928955078125e-07,
+          "max_abs_logit_error": 0.0117645263671875
+        },
+        {
+          "row": 112,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 3.5509287954482716e-08,
+          "max_abs_logit_error": 0.033596038818359375
+        },
+        {
+          "row": 113,
+          "label": 19,
+          "upstream": 19,
+          "coreml": 19,
+          "max_abs_probability_error": 1.1210374850634253e-09,
+          "max_abs_logit_error": 0.021764755249023438
+        },
+        {
+          "row": 114,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 2.9104718635153404e-11,
+          "max_abs_logit_error": 0.017470359802246094
+        },
+        {
+          "row": 115,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 3.615042487581377e-09,
+          "max_abs_logit_error": 0.011733055114746094
+        },
+        {
+          "row": 116,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 2.5568436257117355e-12,
+          "max_abs_logit_error": 0.018376827239990234
+        },
+        {
+          "row": 117,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 3.652814939414384e-09,
+          "max_abs_logit_error": 0.013185739517211914
+        },
+        {
+          "row": 118,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 8.637570658720506e-10,
+          "max_abs_logit_error": 0.010789871215820312
+        },
+        {
+          "row": 119,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 2.6595947666407938e-11,
+          "max_abs_logit_error": 0.022840499877929688
+        },
+        {
+          "row": 120,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.6561418902938385e-11,
+          "max_abs_logit_error": 0.009737014770507812
+        },
+        {
+          "row": 121,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.1621037465658901e-11,
+          "max_abs_logit_error": 0.01908111572265625
+        },
+        {
+          "row": 122,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 8.295308884243013e-13,
+          "max_abs_logit_error": 0.016066551208496094
+        },
+        {
+          "row": 123,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 2.485815372987865e-14,
+          "max_abs_logit_error": 0.01638507843017578
+        },
+        {
+          "row": 124,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 2.78640222006743e-10,
+          "max_abs_logit_error": 0.024122238159179688
+        },
+        {
+          "row": 125,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 3.6049052631881295e-11,
+          "max_abs_logit_error": 0.0227203369140625
+        },
+        {
+          "row": 126,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 7.636086207796211e-12,
+          "max_abs_logit_error": 0.021586894989013672
+        },
+        {
+          "row": 127,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 2.5383029012004954e-12,
+          "max_abs_logit_error": 0.019191741943359375
+        },
+        {
+          "row": 128,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 3.6573055695043877e-09,
+          "max_abs_logit_error": 0.016190528869628906
+        },
+        {
+          "row": 129,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 3.6824872040597256e-09,
+          "max_abs_logit_error": 0.016640663146972656
+        },
+        {
+          "row": 130,
+          "label": 1,
+          "upstream": 1,
+          "coreml": 1,
+          "max_abs_probability_error": 4.76837158203125e-07,
+          "max_abs_logit_error": 0.027278900146484375
+        },
+        {
+          "row": 131,
+          "label": 2,
+          "upstream": 2,
+          "coreml": 2,
+          "max_abs_probability_error": 1.3382816632656613e-08,
+          "max_abs_logit_error": 0.016567230224609375
+        },
+        {
+          "row": 132,
+          "label": 3,
+          "upstream": 3,
+          "coreml": 3,
+          "max_abs_probability_error": 4.5742126530967653e-07,
+          "max_abs_logit_error": 0.025929927825927734
+        },
+        {
+          "row": 133,
+          "label": 4,
+          "upstream": 4,
+          "coreml": 4,
+          "max_abs_probability_error": 1.1920928955078125e-07,
+          "max_abs_logit_error": 0.017045140266418457
+        },
+        {
+          "row": 134,
+          "label": 5,
+          "upstream": 5,
+          "coreml": 5,
+          "max_abs_probability_error": 8.344650268554688e-07,
+          "max_abs_logit_error": 0.017194747924804688
+        },
+        {
+          "row": 135,
+          "label": 6,
+          "upstream": 6,
+          "coreml": 6,
+          "max_abs_probability_error": 3.5762786865234375e-07,
+          "max_abs_logit_error": 0.026395797729492188
+        },
+        {
+          "row": 136,
+          "label": 7,
+          "upstream": 7,
+          "coreml": 7,
+          "max_abs_probability_error": 2.384185791015625e-07,
+          "max_abs_logit_error": 0.020420074462890625
+        },
+        {
+          "row": 137,
+          "label": 8,
+          "upstream": 8,
+          "coreml": 8,
+          "max_abs_probability_error": 0.0030989646911621094,
+          "max_abs_logit_error": 0.021968841552734375
+        },
+        {
+          "row": 138,
+          "label": 9,
+          "upstream": 9,
+          "coreml": 9,
+          "max_abs_probability_error": 1.7425772966817021e-06,
+          "max_abs_logit_error": 0.010594367980957031
+        },
+        {
+          "row": 139,
+          "label": 10,
+          "upstream": 10,
+          "coreml": 10,
+          "max_abs_probability_error": 1.1920928955078125e-07,
+          "max_abs_logit_error": 0.015566825866699219
+        },
+        {
+          "row": 140,
+          "label": 11,
+          "upstream": 11,
+          "coreml": 11,
+          "max_abs_probability_error": 2.733497694862308e-08,
+          "max_abs_logit_error": 0.020360946655273438
+        },
+        {
+          "row": 141,
+          "label": 12,
+          "upstream": 12,
+          "coreml": 12,
+          "max_abs_probability_error": 8.882998372428119e-07,
+          "max_abs_logit_error": 0.01713848114013672
+        },
+        {
+          "row": 142,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 7.821654435247183e-11,
+          "max_abs_logit_error": 0.016475677490234375
+        },
+        {
+          "row": 143,
+          "label": 16,
+          "upstream": 16,
+          "coreml": 16,
+          "max_abs_probability_error": 6.903178473294247e-09,
+          "max_abs_logit_error": 0.02091217041015625
+        },
+        {
+          "row": 144,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.9354274627403356e-08,
+          "max_abs_logit_error": 0.014428257942199707
+        },
+        {
+          "row": 145,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 3.937020665034652e-08,
+          "max_abs_logit_error": 0.013593673706054688
+        },
+        {
+          "row": 146,
+          "label": 17,
+          "upstream": 17,
+          "coreml": 17,
+          "max_abs_probability_error": 1.295774154641549e-09,
+          "max_abs_logit_error": 0.03515338897705078
+        },
+        {
+          "row": 147,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 3.638334078459593e-11,
+          "max_abs_logit_error": 0.016902923583984375
+        },
+        {
+          "row": 148,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.354898415684147e-10,
+          "max_abs_logit_error": 0.007921218872070312
+        },
+        {
+          "row": 149,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 4.4544090638254374e-12,
+          "max_abs_logit_error": 0.018396615982055664
+        },
+        {
+          "row": 150,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 2.2619059403261588e-14,
+          "max_abs_logit_error": 0.019584178924560547
+        },
+        {
+          "row": 151,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 5.77121683775772e-12,
+          "max_abs_logit_error": 0.01009368896484375
+        },
+        {
+          "row": 152,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.4402368186949843e-12,
+          "max_abs_logit_error": 0.016597270965576172
+        },
+        {
+          "row": 153,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 2.3069191001923173e-10,
+          "max_abs_logit_error": 0.023260116577148438
+        },
+        {
+          "row": 154,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 4.5346018606728933e-13,
+          "max_abs_logit_error": 0.023482322692871094
+        },
+        {
+          "row": 155,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.2430889650971721e-12,
+          "max_abs_logit_error": 0.013761520385742188
+        },
+        {
+          "row": 156,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 3.12792326762068e-14,
+          "max_abs_logit_error": 0.013568401336669922
+        },
+        {
+          "row": 157,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.5832196664788967e-12,
+          "max_abs_logit_error": 0.013137876987457275
+        },
+        {
+          "row": 158,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 3.681888127715638e-12,
+          "max_abs_logit_error": 0.026842832565307617
+        },
+        {
+          "row": 159,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.731503829205394e-11,
+          "max_abs_logit_error": 0.015740394592285156
+        },
+        {
+          "row": 160,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.0416112417033219e-10,
+          "max_abs_logit_error": 0.01055908203125
+        },
+        {
+          "row": 161,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 4.944752163282828e-09,
+          "max_abs_logit_error": 0.01996326446533203
+        },
+        {
+          "row": 162,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 3.288525007860699e-11,
+          "max_abs_logit_error": 0.011535167694091797
+        },
+        {
+          "row": 163,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 6.579597977562912e-13,
+          "max_abs_logit_error": 0.010066986083984375
+        },
+        {
+          "row": 164,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 3.8768099841490766e-11,
+          "max_abs_logit_error": 0.014286994934082031
+        },
+        {
+          "row": 165,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.6796633528493032e-13,
+          "max_abs_logit_error": 0.014059066772460938
+        },
+        {
+          "row": 166,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.0951239914902544e-11,
+          "max_abs_logit_error": 0.012742996215820312
+        },
+        {
+          "row": 167,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 5.0552340091769565e-12,
+          "max_abs_logit_error": 0.017006397247314453
+        },
+        {
+          "row": 168,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 4.349431037553586e-09,
+          "max_abs_logit_error": 0.02781963348388672
+        },
+        {
+          "row": 169,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.7824977605052084e-13,
+          "max_abs_logit_error": 0.022431373596191406
+        },
+        {
+          "row": 170,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 3.2377225900326323e-13,
+          "max_abs_logit_error": 0.006749153137207031
+        },
+        {
+          "row": 171,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 3.13179953081999e-08,
+          "max_abs_logit_error": 0.015783309936523438
+        },
+        {
+          "row": 172,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 4.681899312686255e-11,
+          "max_abs_logit_error": 0.013919830322265625
+        },
+        {
+          "row": 173,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 2.2486568163060383e-11,
+          "max_abs_logit_error": 0.016487598419189453
+        },
+        {
+          "row": 174,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 3.743627630115043e-11,
+          "max_abs_logit_error": 0.02295398712158203
+        },
+        {
+          "row": 175,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 7.821654435247183e-11,
+          "max_abs_logit_error": 0.016475677490234375
+        },
+        {
+          "row": 176,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.1420979717513546e-09,
+          "max_abs_logit_error": 0.028388500213623047
+        },
+        {
+          "row": 177,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.5013057463875157e-10,
+          "max_abs_logit_error": 0.015703678131103516
+        },
+        {
+          "row": 178,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 3.937020665034652e-08,
+          "max_abs_logit_error": 0.013593673706054688
+        },
+        {
+          "row": 179,
+          "label": 17,
+          "upstream": 17,
+          "coreml": 17,
+          "max_abs_probability_error": 1.295774154641549e-09,
+          "max_abs_logit_error": 0.03515338897705078
+        },
+        {
+          "row": 180,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 3.638334078459593e-11,
+          "max_abs_logit_error": 0.016902923583984375
+        },
+        {
+          "row": 181,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.354898415684147e-10,
+          "max_abs_logit_error": 0.007921218872070312
+        },
+        {
+          "row": 182,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 4.4544090638254374e-12,
+          "max_abs_logit_error": 0.018396615982055664
+        },
+        {
+          "row": 183,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 2.2619059403261588e-14,
+          "max_abs_logit_error": 0.019584178924560547
+        },
+        {
+          "row": 184,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 5.77121683775772e-12,
+          "max_abs_logit_error": 0.01009368896484375
+        },
+        {
+          "row": 185,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.4402368186949843e-12,
+          "max_abs_logit_error": 0.016597270965576172
+        },
+        {
+          "row": 186,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 2.3069191001923173e-10,
+          "max_abs_logit_error": 0.023260116577148438
+        },
+        {
+          "row": 187,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 4.5346018606728933e-13,
+          "max_abs_logit_error": 0.023482322692871094
+        },
+        {
+          "row": 188,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.2430889650971721e-12,
+          "max_abs_logit_error": 0.013761520385742188
+        },
+        {
+          "row": 189,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 3.12792326762068e-14,
+          "max_abs_logit_error": 0.013568401336669922
+        },
+        {
+          "row": 190,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.5832196664788967e-12,
+          "max_abs_logit_error": 0.013137876987457275
+        },
+        {
+          "row": 191,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 3.681888127715638e-12,
+          "max_abs_logit_error": 0.026842832565307617
+        },
+        {
+          "row": 192,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.731503829205394e-11,
+          "max_abs_logit_error": 0.015740394592285156
+        },
+        {
+          "row": 193,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.0416112417033219e-10,
+          "max_abs_logit_error": 0.01055908203125
+        },
+        {
+          "row": 194,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 4.944752163282828e-09,
+          "max_abs_logit_error": 0.01996326446533203
+        },
+        {
+          "row": 195,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 3.288525007860699e-11,
+          "max_abs_logit_error": 0.011535167694091797
+        }
+      ]
+    },
+    "CPU_AND_NE": {
+      "passed": true,
+      "correct": 196,
+      "accuracy": 1.0,
+      "argmax_agreement": 196,
+      "max_abs_probability_error": 0.00233614444732666,
+      "mean_row_max_abs_probability_error": 2.934475840582507e-05,
+      "max_abs_logit_error": 0.07230281829833984,
+      "load_seconds": 0.578660333994776,
+      "warm_inference_latency_ms": {
+        "count": 196,
+        "median_ms": 0.9031249937834218,
+        "p95_ms": 0.9396872410434298,
+        "min_ms": 0.8796670008450747,
+        "max_ms": 0.9663329983595759
+      },
+      "reversed_option_order": {
+        "rows": 6,
+        "max_abs_probability_error": 3.1828880310058594e-05
+      },
+      "per_action": {
+        "fill": {
+          "rows": 36,
+          "correct": 36
+        },
+        "skip": {
+          "rows": 150,
+          "correct": 150
+        },
+        "check": {
+          "rows": 4,
+          "correct": 4
+        },
+        "click": {
+          "rows": 6,
+          "correct": 6
+        }
+      },
+      "decisions": [
+        {
+          "row": 0,
+          "label": 22,
+          "upstream": 22,
+          "coreml": 22,
+          "max_abs_probability_error": 3.1828880310058594e-05,
+          "max_abs_logit_error": 0.03086566925048828
+        },
+        {
+          "row": 1,
+          "label": 23,
+          "upstream": 23,
+          "coreml": 23,
+          "max_abs_probability_error": 3.814697265625e-05,
+          "max_abs_logit_error": 0.020841598510742188
+        },
+        {
+          "row": 2,
+          "label": 2,
+          "upstream": 2,
+          "coreml": 2,
+          "max_abs_probability_error": 1.1920928955078125e-07,
+          "max_abs_logit_error": 0.024377822875976562
+        },
+        {
+          "row": 3,
+          "label": 5,
+          "upstream": 5,
+          "coreml": 5,
+          "max_abs_probability_error": 3.218650817871094e-05,
+          "max_abs_logit_error": 0.030775785446166992
+        },
+        {
+          "row": 4,
+          "label": 4,
+          "upstream": 4,
+          "coreml": 4,
+          "max_abs_probability_error": 2.944469451904297e-05,
+          "max_abs_logit_error": 0.03573417663574219
+        },
+        {
+          "row": 5,
+          "label": 6,
+          "upstream": 6,
+          "coreml": 6,
+          "max_abs_probability_error": 1.2636184692382812e-05,
+          "max_abs_logit_error": 0.026729822158813477
+        },
+        {
+          "row": 6,
+          "label": 7,
+          "upstream": 7,
+          "coreml": 7,
+          "max_abs_probability_error": 2.3365020751953125e-05,
+          "max_abs_logit_error": 0.023685455322265625
+        },
+        {
+          "row": 7,
+          "label": 8,
+          "upstream": 8,
+          "coreml": 8,
+          "max_abs_probability_error": 0.00014638900756835938,
+          "max_abs_logit_error": 0.04038810729980469
+        },
+        {
+          "row": 8,
+          "label": 9,
+          "upstream": 9,
+          "coreml": 9,
+          "max_abs_probability_error": 1.1920928955078125e-06,
+          "max_abs_logit_error": 0.02460002899169922
+        },
+        {
+          "row": 9,
+          "label": 10,
+          "upstream": 10,
+          "coreml": 10,
+          "max_abs_probability_error": 2.1457672119140625e-05,
+          "max_abs_logit_error": 0.018064022064208984
+        },
+        {
+          "row": 10,
+          "label": 11,
+          "upstream": 11,
+          "coreml": 11,
+          "max_abs_probability_error": 5.960464477539062e-07,
+          "max_abs_logit_error": 0.047691673040390015
+        },
+        {
+          "row": 11,
+          "label": 13,
+          "upstream": 13,
+          "coreml": 13,
+          "max_abs_probability_error": 3.2186508178710938e-06,
+          "max_abs_logit_error": 0.03340768814086914
+        },
+        {
+          "row": 12,
+          "label": 15,
+          "upstream": 15,
+          "coreml": 15,
+          "max_abs_probability_error": 2.384185791015625e-06,
+          "max_abs_logit_error": 0.053191184997558594
+        },
+        {
+          "row": 13,
+          "label": 16,
+          "upstream": 16,
+          "coreml": 16,
+          "max_abs_probability_error": 0.0005360841751098633,
+          "max_abs_logit_error": 0.03362560272216797
+        },
+        {
+          "row": 14,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 1.8596649169921875e-05,
+          "max_abs_logit_error": 0.024669170379638672
+        },
+        {
+          "row": 15,
+          "label": 24,
+          "upstream": 24,
+          "coreml": 24,
+          "max_abs_probability_error": 1.6689300537109375e-05,
+          "max_abs_logit_error": 0.034501075744628906
+        },
+        {
+          "row": 16,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 1.440092489701783e-08,
+          "max_abs_logit_error": 0.033290743827819824
+        },
+        {
+          "row": 17,
+          "label": 25,
+          "upstream": 25,
+          "coreml": 25,
+          "max_abs_probability_error": 1.1920928955078125e-07,
+          "max_abs_logit_error": 0.03323173522949219
+        },
+        {
+          "row": 18,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 8.430758313693332e-09,
+          "max_abs_logit_error": 0.022179126739501953
+        },
+        {
+          "row": 19,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 3.0266471551243512e-09,
+          "max_abs_logit_error": 0.029880523681640625
+        },
+        {
+          "row": 20,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 2.4500966588902884e-08,
+          "max_abs_logit_error": 0.031145095825195312
+        },
+        {
+          "row": 21,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 1.647379281599637e-11,
+          "max_abs_logit_error": 0.03861522674560547
+        },
+        {
+          "row": 22,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 7.152557373046875e-07,
+          "max_abs_logit_error": 0.027164459228515625
+        },
+        {
+          "row": 23,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 5.960464477539062e-07,
+          "max_abs_logit_error": 0.058716535568237305
+        },
+        {
+          "row": 24,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 2.5055844551924444e-11,
+          "max_abs_logit_error": 0.035381317138671875
+        },
+        {
+          "row": 25,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 6.995705792434137e-09,
+          "max_abs_logit_error": 0.04862499237060547
+        },
+        {
+          "row": 26,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 6.470584068551943e-10,
+          "max_abs_logit_error": 0.029034733772277832
+        },
+        {
+          "row": 27,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 6.973024935241767e-11,
+          "max_abs_logit_error": 0.03808116912841797
+        },
+        {
+          "row": 28,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 5.994321927715873e-09,
+          "max_abs_logit_error": 0.03319191932678223
+        },
+        {
+          "row": 29,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 2.4426039316183257e-11,
+          "max_abs_logit_error": 0.030045509338378906
+        },
+        {
+          "row": 30,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 2.384185791015625e-07,
+          "max_abs_logit_error": 0.026231765747070312
+        },
+        {
+          "row": 31,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 7.78299757975276e-10,
+          "max_abs_logit_error": 0.033112093806266785
+        },
+        {
+          "row": 32,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 6.185003176284454e-09,
+          "max_abs_logit_error": 0.022128582000732422
+        },
+        {
+          "row": 33,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 3.4556297823229443e-09,
+          "max_abs_logit_error": 0.038306236267089844
+        },
+        {
+          "row": 34,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 0.0004820823669433594,
+          "max_abs_logit_error": 0.032683372497558594
+        },
+        {
+          "row": 35,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 1.462750809366753e-09,
+          "max_abs_logit_error": 0.029730796813964844
+        },
+        {
+          "row": 36,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 2.7510341227277024e-10,
+          "max_abs_logit_error": 0.02333354949951172
+        },
+        {
+          "row": 37,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 2.657592190757896e-08,
+          "max_abs_logit_error": 0.022855758666992188
+        },
+        {
+          "row": 38,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 9.552543478452691e-11,
+          "max_abs_logit_error": 0.023164749145507812
+        },
+        {
+          "row": 39,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 4.04716482549361e-10,
+          "max_abs_logit_error": 0.02880859375
+        },
+        {
+          "row": 40,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 0.00021719932556152344,
+          "max_abs_logit_error": 0.04735088348388672
+        },
+        {
+          "row": 41,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 6.146233744175333e-09,
+          "max_abs_logit_error": 0.03671276569366455
+        },
+        {
+          "row": 42,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 1.8819494851385343e-09,
+          "max_abs_logit_error": 0.023815155029296875
+        },
+        {
+          "row": 43,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 7.152557373046875e-07,
+          "max_abs_logit_error": 0.02039623260498047
+        },
+        {
+          "row": 44,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 1.780525865635596e-10,
+          "max_abs_logit_error": 0.022927284240722656
+        },
+        {
+          "row": 45,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 3.414331317674879e-11,
+          "max_abs_logit_error": 0.030417919158935547
+        },
+        {
+          "row": 46,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 5.599257724142603e-10,
+          "max_abs_logit_error": 0.025608062744140625
+        },
+        {
+          "row": 47,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 5.212314607705437e-11,
+          "max_abs_logit_error": 0.022733211517333984
+        },
+        {
+          "row": 48,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 1.8596649169921875e-05,
+          "max_abs_logit_error": 0.024669170379638672
+        },
+        {
+          "row": 49,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 1.932842080831776e-10,
+          "max_abs_logit_error": 0.030317306518554688
+        },
+        {
+          "row": 50,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 1.440092489701783e-08,
+          "max_abs_logit_error": 0.033290743827819824
+        },
+        {
+          "row": 51,
+          "label": 25,
+          "upstream": 25,
+          "coreml": 25,
+          "max_abs_probability_error": 1.1920928955078125e-07,
+          "max_abs_logit_error": 0.03323173522949219
+        },
+        {
+          "row": 52,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 8.430758313693332e-09,
+          "max_abs_logit_error": 0.022179126739501953
+        },
+        {
+          "row": 53,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 3.0266471551243512e-09,
+          "max_abs_logit_error": 0.029880523681640625
+        },
+        {
+          "row": 54,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 2.4500966588902884e-08,
+          "max_abs_logit_error": 0.031145095825195312
+        },
+        {
+          "row": 55,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 1.647379281599637e-11,
+          "max_abs_logit_error": 0.03861522674560547
+        },
+        {
+          "row": 56,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 7.152557373046875e-07,
+          "max_abs_logit_error": 0.027164459228515625
+        },
+        {
+          "row": 57,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 5.960464477539062e-07,
+          "max_abs_logit_error": 0.058716535568237305
+        },
+        {
+          "row": 58,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 2.5055844551924444e-11,
+          "max_abs_logit_error": 0.035381317138671875
+        },
+        {
+          "row": 59,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 6.995705792434137e-09,
+          "max_abs_logit_error": 0.04862499237060547
+        },
+        {
+          "row": 60,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 6.470584068551943e-10,
+          "max_abs_logit_error": 0.029034733772277832
+        },
+        {
+          "row": 61,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 6.973024935241767e-11,
+          "max_abs_logit_error": 0.03808116912841797
+        },
+        {
+          "row": 62,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 5.994321927715873e-09,
+          "max_abs_logit_error": 0.03319191932678223
+        },
+        {
+          "row": 63,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 2.4426039316183257e-11,
+          "max_abs_logit_error": 0.030045509338378906
+        },
+        {
+          "row": 64,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 2.384185791015625e-07,
+          "max_abs_logit_error": 0.026231765747070312
+        },
+        {
+          "row": 65,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 7.78299757975276e-10,
+          "max_abs_logit_error": 0.033112093806266785
+        },
+        {
+          "row": 66,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 6.185003176284454e-09,
+          "max_abs_logit_error": 0.022128582000732422
+        },
+        {
+          "row": 67,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 3.4556297823229443e-09,
+          "max_abs_logit_error": 0.038306236267089844
+        },
+        {
+          "row": 68,
+          "label": 0,
+          "upstream": 0,
+          "coreml": 0,
+          "max_abs_probability_error": 3.147125244140625e-05,
+          "max_abs_logit_error": 0.03265953063964844
+        },
+        {
+          "row": 69,
+          "label": 1,
+          "upstream": 1,
+          "coreml": 1,
+          "max_abs_probability_error": 5.030632019042969e-05,
+          "max_abs_logit_error": 0.038794517517089844
+        },
+        {
+          "row": 70,
+          "label": 2,
+          "upstream": 2,
+          "coreml": 2,
+          "max_abs_probability_error": 1.7881393432617188e-06,
+          "max_abs_logit_error": 0.040470123291015625
+        },
+        {
+          "row": 71,
+          "label": 3,
+          "upstream": 3,
+          "coreml": 3,
+          "max_abs_probability_error": 5.2928924560546875e-05,
+          "max_abs_logit_error": 0.040035247802734375
+        },
+        {
+          "row": 72,
+          "label": 5,
+          "upstream": 5,
+          "coreml": 5,
+          "max_abs_probability_error": 4.649162292480469e-06,
+          "max_abs_logit_error": 0.042186737060546875
+        },
+        {
+          "row": 73,
+          "label": 6,
+          "upstream": 6,
+          "coreml": 6,
+          "max_abs_probability_error": 2.1576881408691406e-05,
+          "max_abs_logit_error": 0.033161163330078125
+        },
+        {
+          "row": 74,
+          "label": 7,
+          "upstream": 7,
+          "coreml": 7,
+          "max_abs_probability_error": 4.470348358154297e-05,
+          "max_abs_logit_error": 0.0293731689453125
+        },
+        {
+          "row": 75,
+          "label": 8,
+          "upstream": 8,
+          "coreml": 8,
+          "max_abs_probability_error": 1.1563301086425781e-05,
+          "max_abs_logit_error": 0.03249359130859375
+        },
+        {
+          "row": 76,
+          "label": 9,
+          "upstream": 9,
+          "coreml": 9,
+          "max_abs_probability_error": 2.86102294921875e-06,
+          "max_abs_logit_error": 0.03498554229736328
+        },
+        {
+          "row": 77,
+          "label": 13,
+          "upstream": 13,
+          "coreml": 13,
+          "max_abs_probability_error": 0.0003635883331298828,
+          "max_abs_logit_error": 0.023473739624023438
+        },
+        {
+          "row": 78,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 9.5367431640625e-07,
+          "max_abs_logit_error": 0.034421443939208984
+        },
+        {
+          "row": 79,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 0.0003609657287597656,
+          "max_abs_logit_error": 0.03937721252441406
+        },
+        {
+          "row": 80,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 2.384185791015625e-07,
+          "max_abs_logit_error": 0.03752422332763672
+        },
+        {
+          "row": 81,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.7881393432617188e-06,
+          "max_abs_logit_error": 0.025417327880859375
+        },
+        {
+          "row": 82,
+          "label": 19,
+          "upstream": 19,
+          "coreml": 19,
+          "max_abs_probability_error": 2.384185791015625e-07,
+          "max_abs_logit_error": 0.04453086853027344
+        },
+        {
+          "row": 83,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 5.0942090545902374e-09,
+          "max_abs_logit_error": 0.023950576782226562
+        },
+        {
+          "row": 84,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 2.384185791015625e-07,
+          "max_abs_logit_error": 0.022706031799316406
+        },
+        {
+          "row": 85,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 3.5405053888659666e-10,
+          "max_abs_logit_error": 0.035375118255615234
+        },
+        {
+          "row": 86,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 4.76837158203125e-07,
+          "max_abs_logit_error": 0.03717994689941406
+        },
+        {
+          "row": 87,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.1920928955078125e-07,
+          "max_abs_logit_error": 0.04058122634887695
+        },
+        {
+          "row": 88,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.3188037328859537e-09,
+          "max_abs_logit_error": 0.032048702239990234
+        },
+        {
+          "row": 89,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.9970605169561395e-09,
+          "max_abs_logit_error": 0.030427932739257812
+        },
+        {
+          "row": 90,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.7617375336342889e-09,
+          "max_abs_logit_error": 0.0281219482421875
+        },
+        {
+          "row": 91,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.612549260787688e-10,
+          "max_abs_logit_error": 0.030226707458496094
+        },
+        {
+          "row": 92,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 6.506330630512425e-12,
+          "max_abs_logit_error": 0.04007530212402344
+        },
+        {
+          "row": 93,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.5150824594911683e-08,
+          "max_abs_logit_error": 0.02728891372680664
+        },
+        {
+          "row": 94,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.3666674458789885e-09,
+          "max_abs_logit_error": 0.0217437744140625
+        },
+        {
+          "row": 95,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 3.969476136678196e-10,
+          "max_abs_logit_error": 0.03214454650878906
+        },
+        {
+          "row": 96,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 2.657726538846106e-10,
+          "max_abs_logit_error": 0.03669023513793945
+        },
+        {
+          "row": 97,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.9073486328125e-06,
+          "max_abs_logit_error": 0.019012451171875
+        },
+        {
+          "row": 98,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 2.384185791015625e-07,
+          "max_abs_logit_error": 0.026651382446289062
+        },
+        {
+          "row": 99,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 2.1457672119140625e-06,
+          "max_abs_logit_error": 0.04161381721496582
+        },
+        {
+          "row": 100,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 9.5367431640625e-07,
+          "max_abs_logit_error": 0.04680633544921875
+        },
+        {
+          "row": 101,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.789079107084035e-08,
+          "max_abs_logit_error": 0.029542922973632812
+        },
+        {
+          "row": 102,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 2.5862225694339713e-09,
+          "max_abs_logit_error": 0.023795127868652344
+        },
+        {
+          "row": 103,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 0.00024586915969848633,
+          "max_abs_logit_error": 0.030111312866210938
+        },
+        {
+          "row": 104,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 5.841255187988281e-06,
+          "max_abs_logit_error": 0.03126239776611328
+        },
+        {
+          "row": 105,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 4.928799035575082e-10,
+          "max_abs_logit_error": 0.026149272918701172
+        },
+        {
+          "row": 106,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 2.4136979082101107e-09,
+          "max_abs_logit_error": 0.024585723876953125
+        },
+        {
+          "row": 107,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 6.302839361538304e-10,
+          "max_abs_logit_error": 0.024765491485595703
+        },
+        {
+          "row": 108,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.1920928955078125e-07,
+          "max_abs_logit_error": 0.03492164611816406
+        },
+        {
+          "row": 109,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 9.5367431640625e-07,
+          "max_abs_logit_error": 0.034421443939208984
+        },
+        {
+          "row": 110,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 9.5367431640625e-07,
+          "max_abs_logit_error": 0.0320439338684082
+        },
+        {
+          "row": 111,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.811981201171875e-05,
+          "max_abs_logit_error": 0.032814979553222656
+        },
+        {
+          "row": 112,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.7881393432617188e-06,
+          "max_abs_logit_error": 0.025417327880859375
+        },
+        {
+          "row": 113,
+          "label": 19,
+          "upstream": 19,
+          "coreml": 19,
+          "max_abs_probability_error": 2.384185791015625e-07,
+          "max_abs_logit_error": 0.04453086853027344
+        },
+        {
+          "row": 114,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 5.0942090545902374e-09,
+          "max_abs_logit_error": 0.023950576782226562
+        },
+        {
+          "row": 115,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 2.384185791015625e-07,
+          "max_abs_logit_error": 0.022706031799316406
+        },
+        {
+          "row": 116,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 3.5405053888659666e-10,
+          "max_abs_logit_error": 0.035375118255615234
+        },
+        {
+          "row": 117,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 4.76837158203125e-07,
+          "max_abs_logit_error": 0.03717994689941406
+        },
+        {
+          "row": 118,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.1920928955078125e-07,
+          "max_abs_logit_error": 0.04058122634887695
+        },
+        {
+          "row": 119,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.3188037328859537e-09,
+          "max_abs_logit_error": 0.032048702239990234
+        },
+        {
+          "row": 120,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.9970605169561395e-09,
+          "max_abs_logit_error": 0.030427932739257812
+        },
+        {
+          "row": 121,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.7617375336342889e-09,
+          "max_abs_logit_error": 0.0281219482421875
+        },
+        {
+          "row": 122,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.612549260787688e-10,
+          "max_abs_logit_error": 0.030226707458496094
+        },
+        {
+          "row": 123,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 6.506330630512425e-12,
+          "max_abs_logit_error": 0.04007530212402344
+        },
+        {
+          "row": 124,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.5150824594911683e-08,
+          "max_abs_logit_error": 0.02728891372680664
+        },
+        {
+          "row": 125,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.3666674458789885e-09,
+          "max_abs_logit_error": 0.0217437744140625
+        },
+        {
+          "row": 126,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 3.969476136678196e-10,
+          "max_abs_logit_error": 0.03214454650878906
+        },
+        {
+          "row": 127,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 2.657726538846106e-10,
+          "max_abs_logit_error": 0.03669023513793945
+        },
+        {
+          "row": 128,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.9073486328125e-06,
+          "max_abs_logit_error": 0.019012451171875
+        },
+        {
+          "row": 129,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 2.384185791015625e-07,
+          "max_abs_logit_error": 0.026651382446289062
+        },
+        {
+          "row": 130,
+          "label": 1,
+          "upstream": 1,
+          "coreml": 1,
+          "max_abs_probability_error": 3.170967102050781e-05,
+          "max_abs_logit_error": 0.071075439453125
+        },
+        {
+          "row": 131,
+          "label": 2,
+          "upstream": 2,
+          "coreml": 2,
+          "max_abs_probability_error": 8.344650268554688e-07,
+          "max_abs_logit_error": 0.051357269287109375
+        },
+        {
+          "row": 132,
+          "label": 3,
+          "upstream": 3,
+          "coreml": 3,
+          "max_abs_probability_error": 2.7894973754882812e-05,
+          "max_abs_logit_error": 0.033801641315221786
+        },
+        {
+          "row": 133,
+          "label": 4,
+          "upstream": 4,
+          "coreml": 4,
+          "max_abs_probability_error": 1.0967254638671875e-05,
+          "max_abs_logit_error": 0.039325714111328125
+        },
+        {
+          "row": 134,
+          "label": 5,
+          "upstream": 5,
+          "coreml": 5,
+          "max_abs_probability_error": 6.258487701416016e-05,
+          "max_abs_logit_error": 0.033056676387786865
+        },
+        {
+          "row": 135,
+          "label": 6,
+          "upstream": 6,
+          "coreml": 6,
+          "max_abs_probability_error": 1.2993812561035156e-05,
+          "max_abs_logit_error": 0.02561807632446289
+        },
+        {
+          "row": 136,
+          "label": 7,
+          "upstream": 7,
+          "coreml": 7,
+          "max_abs_probability_error": 6.198883056640625e-06,
+          "max_abs_logit_error": 0.03448677062988281
+        },
+        {
+          "row": 137,
+          "label": 8,
+          "upstream": 8,
+          "coreml": 8,
+          "max_abs_probability_error": 0.00233614444732666,
+          "max_abs_logit_error": 0.04230833053588867
+        },
+        {
+          "row": 138,
+          "label": 9,
+          "upstream": 9,
+          "coreml": 9,
+          "max_abs_probability_error": 0.00022917985916137695,
+          "max_abs_logit_error": 0.05339241027832031
+        },
+        {
+          "row": 139,
+          "label": 10,
+          "upstream": 10,
+          "coreml": 10,
+          "max_abs_probability_error": 4.851818084716797e-05,
+          "max_abs_logit_error": 0.032756805419921875
+        },
+        {
+          "row": 140,
+          "label": 11,
+          "upstream": 11,
+          "coreml": 11,
+          "max_abs_probability_error": 3.0994415283203125e-06,
+          "max_abs_logit_error": 0.029809951782226562
+        },
+        {
+          "row": 141,
+          "label": 12,
+          "upstream": 12,
+          "coreml": 12,
+          "max_abs_probability_error": 9.34600830078125e-05,
+          "max_abs_logit_error": 0.026587963104248047
+        },
+        {
+          "row": 142,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.9073486328125e-06,
+          "max_abs_logit_error": 0.03671073913574219
+        },
+        {
+          "row": 143,
+          "label": 16,
+          "upstream": 16,
+          "coreml": 16,
+          "max_abs_probability_error": 7.152557373046875e-07,
+          "max_abs_logit_error": 0.05281543731689453
+        },
+        {
+          "row": 144,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.1920928955078125e-06,
+          "max_abs_logit_error": 0.04511451721191406
+        },
+        {
+          "row": 145,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 3.6954879760742188e-06,
+          "max_abs_logit_error": 0.07230281829833984
+        },
+        {
+          "row": 146,
+          "label": 17,
+          "upstream": 17,
+          "coreml": 17,
+          "max_abs_probability_error": 1.1920928955078125e-07,
+          "max_abs_logit_error": 0.03114461898803711
+        },
+        {
+          "row": 147,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 8.429776876539563e-09,
+          "max_abs_logit_error": 0.020772457122802734
+        },
+        {
+          "row": 148,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 9.215698426601193e-09,
+          "max_abs_logit_error": 0.03475606441497803
+        },
+        {
+          "row": 149,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 3.84208304060607e-10,
+          "max_abs_logit_error": 0.020298004150390625
+        },
+        {
+          "row": 150,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 7.947568818333917e-12,
+          "max_abs_logit_error": 0.04283332824707031
+        },
+        {
+          "row": 151,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 5.291168614363073e-10,
+          "max_abs_logit_error": 0.028450965881347656
+        },
+        {
+          "row": 152,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 7.924417788629512e-10,
+          "max_abs_logit_error": 0.02250051498413086
+        },
+        {
+          "row": 153,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.837478791344438e-08,
+          "max_abs_logit_error": 0.06335592269897461
+        },
+        {
+          "row": 154,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 3.3051329034750054e-11,
+          "max_abs_logit_error": 0.02735137939453125
+        },
+        {
+          "row": 155,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.650468511860126e-10,
+          "max_abs_logit_error": 0.015047073364257812
+        },
+        {
+          "row": 156,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 6.181078218703284e-12,
+          "max_abs_logit_error": 0.04868888854980469
+        },
+        {
+          "row": 157,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.6455586893115992e-10,
+          "max_abs_logit_error": 0.026497364044189453
+        },
+        {
+          "row": 158,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 6.109900363426846e-10,
+          "max_abs_logit_error": 0.022236347198486328
+        },
+        {
+          "row": 159,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 6.693580800742893e-09,
+          "max_abs_logit_error": 0.020044326782226562
+        },
+        {
+          "row": 160,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.3718365998727222e-08,
+          "max_abs_logit_error": 0.05533123016357422
+        },
+        {
+          "row": 161,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 8.344650268554688e-07,
+          "max_abs_logit_error": 0.0247042179107666
+        },
+        {
+          "row": 162,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.042114572413766e-08,
+          "max_abs_logit_error": 0.028873443603515625
+        },
+        {
+          "row": 163,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.9131388140358752e-10,
+          "max_abs_logit_error": 0.015494346618652344
+        },
+        {
+          "row": 164,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.0254025184508464e-08,
+          "max_abs_logit_error": 0.02497100830078125
+        },
+        {
+          "row": 165,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 5.15840287151903e-11,
+          "max_abs_logit_error": 0.02679443359375
+        },
+        {
+          "row": 166,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 2.6797479790729994e-09,
+          "max_abs_logit_error": 0.028203964233398438
+        },
+        {
+          "row": 167,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 7.54509399403247e-10,
+          "max_abs_logit_error": 0.03371429443359375
+        },
+        {
+          "row": 168,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 8.344650268554688e-07,
+          "max_abs_logit_error": 0.028280355036258698
+        },
+        {
+          "row": 169,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 2.4169340140378637e-11,
+          "max_abs_logit_error": 0.038245439529418945
+        },
+        {
+          "row": 170,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 3.0329603412093675e-11,
+          "max_abs_logit_error": 0.030043363571166992
+        },
+        {
+          "row": 171,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.7881393432617188e-06,
+          "max_abs_logit_error": 0.02963542938232422
+        },
+        {
+          "row": 172,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 4.66337235494052e-09,
+          "max_abs_logit_error": 0.03510093688964844
+        },
+        {
+          "row": 173,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.2885730260592254e-09,
+          "max_abs_logit_error": 0.04009199142456055
+        },
+        {
+          "row": 174,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.1279766010119374e-09,
+          "max_abs_logit_error": 0.031810760498046875
+        },
+        {
+          "row": 175,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.9073486328125e-06,
+          "max_abs_logit_error": 0.03671073913574219
+        },
+        {
+          "row": 176,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.0728836059570312e-06,
+          "max_abs_logit_error": 0.05123615264892578
+        },
+        {
+          "row": 177,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.2799782567185503e-08,
+          "max_abs_logit_error": 0.03157186508178711
+        },
+        {
+          "row": 178,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 3.6954879760742188e-06,
+          "max_abs_logit_error": 0.07230281829833984
+        },
+        {
+          "row": 179,
+          "label": 17,
+          "upstream": 17,
+          "coreml": 17,
+          "max_abs_probability_error": 1.1920928955078125e-07,
+          "max_abs_logit_error": 0.03114461898803711
+        },
+        {
+          "row": 180,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 8.429776876539563e-09,
+          "max_abs_logit_error": 0.020772457122802734
+        },
+        {
+          "row": 181,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 9.215698426601193e-09,
+          "max_abs_logit_error": 0.03475606441497803
+        },
+        {
+          "row": 182,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 3.84208304060607e-10,
+          "max_abs_logit_error": 0.020298004150390625
+        },
+        {
+          "row": 183,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 7.947568818333917e-12,
+          "max_abs_logit_error": 0.04283332824707031
+        },
+        {
+          "row": 184,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 5.291168614363073e-10,
+          "max_abs_logit_error": 0.028450965881347656
+        },
+        {
+          "row": 185,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 7.924417788629512e-10,
+          "max_abs_logit_error": 0.02250051498413086
+        },
+        {
+          "row": 186,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.837478791344438e-08,
+          "max_abs_logit_error": 0.06335592269897461
+        },
+        {
+          "row": 187,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 3.3051329034750054e-11,
+          "max_abs_logit_error": 0.02735137939453125
+        },
+        {
+          "row": 188,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.650468511860126e-10,
+          "max_abs_logit_error": 0.015047073364257812
+        },
+        {
+          "row": 189,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 6.181078218703284e-12,
+          "max_abs_logit_error": 0.04868888854980469
+        },
+        {
+          "row": 190,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.6455586893115992e-10,
+          "max_abs_logit_error": 0.026497364044189453
+        },
+        {
+          "row": 191,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 6.109900363426846e-10,
+          "max_abs_logit_error": 0.022236347198486328
+        },
+        {
+          "row": 192,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 6.693580800742893e-09,
+          "max_abs_logit_error": 0.020044326782226562
+        },
+        {
+          "row": 193,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.3718365998727222e-08,
+          "max_abs_logit_error": 0.05533123016357422
+        },
+        {
+          "row": 194,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 8.344650268554688e-07,
+          "max_abs_logit_error": 0.0247042179107666
+        },
+        {
+          "row": 195,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.042114572413766e-08,
+          "max_abs_logit_error": 0.028873443603515625
+        }
+      ]
+    }
+  },
+  "passed": true,
+  "code_sha256": {
+    "assets.py": "f7a5bc931075237e00c73bf21fee3575be35531f934f23b5e22627d50a79f92f",
+    "convert-coreml.py": "e1e9cc20277515fc750940e97d75f9db7c15f0443e23c26fe65815cef4e23bb7",
+    "export_model.py": "803b909577fe7c38968c5438b27fff7c07bb095a7c341c1b0430d91256649113",
+    "preprocessing.py": "18c74ede43f95f91cd638c7e9631923a353e3a24d3b522ddb8b18b480f9c2846",
+    "verify.py": "31425c5f0081b07725ce94c611664151ed02678f25ca93774497cfca8f46ceeb"
+  }
+}