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

.gitattributes

@@ -33,3 +33,6 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+demo/browser-demo.gif filter=lfs diff=lfs merge=lfs -text
+demo/browser-demo.mp4 filter=lfs diff=lfs merge=lfs -text
+demo/browser-preview.png filter=lfs diff=lfs merge=lfs -text

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 +1,446 @@
+---
+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 |
+| `synthetic-test.lock.json` | Full published synthetic test revision, count, size, and SHA-256 |
+| `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(),
+)
+raw_probabilities = model.predict(inputs)["probabilities"][0, :len(options)]
+print(options[int(raw_probabilities.argmax())], raw_probabilities)
+```
+
+Download the model artifacts:
+
+```bash
+hf download FluidInference/cua-s1-forms-coreml --revision main --local-dir ./cua-coreml
+cd cua-coreml
+```
+
+## Swift usage
+
+The [FluidAudio integration](https://github.com/FluidInference/FluidAudio/tree/main)
+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)
+```
+
+`decision.probabilities` uses stable Swift softmax; `decision.rawProbabilities`
+preserves the package output shown in the Python example. See the
+[full runtime validation](#swift-probability-fix).
+
+`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 was performed. The separate full synthetic test below evaluates
+the unchanged artifacts and exposes numerical failures outside this demo.
+
+## Full published synthetic test
+
+Evaluated the complete published synthetic [`test.jsonl`](https://huggingface.co/datasets/cua-ai/cua-s1-forms/blob/8273f34778b99ac2e12d9f6e7d57dad99ae20845/test.jsonl):
+**24,370 decisions across 1,040 episode seeds**, with no excluded rows or truncated
+inputs. The dataset revision is `8273f34778b99ac2e12d9f6e7d57dad99ae20845`; its
+SHA-256 is `d63a7e0db195d4d20154a40b2f8dd09ce3bb65487a158c638da5c609d4475e7c`.
+The upstream [model card](https://huggingface.co/cua-ai/cua-s1-forms/blob/f54adbf447f4ca6ec259f529ee3f2e3e09f8cc71/README.md) reports 99.95% on an approximately 15,000-row
+synthetic test. Our result rounds to that accuracy, but the released file contains
+24,370 rows; this does not reconstruct the card's unspecified smaller manifest.
+
+| Model / backend | Correct decisions | Top-1 accuracy | Median call | p95 call |
+| --- | ---: | ---: | ---: | ---: |
+| Upstream PyTorch / CPU | 24,359 / 24,370 | 99.9549% | 1.787 ms | 3.104 ms |
+| Original FP16 Core ML / CPU + ANE | 24,359 / 24,370 | 99.9549% | 1.003 ms | 1.133 ms |
+| ANE-gather FP16 Core ML / CPU + ANE | 24,359 / 24,370 | 99.9549% | 1.052 ms | 1.177 ms |
+
+Both Core ML exports select the same option as PyTorch on **all 24,370 rows**.
+All three share the same 11 errors: choosing `fill` when the label is `skip`.
+The original Cua evaluator counts these as wrong/unsafe actions; this offline
+benchmark executes no actions. All 9,802 fill, 816 check, and 1,040 click labels
+are correct; skip accuracy is 12,701/12,712. Higher ANE placement is approximately
+**4.8% slower** by median here, so the original remains the default.
+
+**Strict numerical conversion parity fails for both exports.** Each has 11 rows
+above the original 0.005 absolute probability-error limit, with a maximum error
+of **0.0204874**. One further row (zero-based index 19270) has a live-probability
+sum of **0.99893665**, which failed the original Swift manager's 0.001
+normalization guard despite a correct argmax. The [Swift probability fix](#swift-probability-fix)
+now handles that output. These raw conversion reports retain the original scores,
+failed gates, tolerances, and model weights; the runtime fix does not establish
+raw numerical parity. The earlier 196-row demo passed its numerical gates.
+
+Measured September 19, 2026 on **Apple M5 Pro, 24 GB, macOS 27.0 (26A428)**,
+Python 3.11.11, PyTorch 2.7.0, coremltools 9.0. Batch size 1, three warmup rows per
+model, one timed pass over the whole split; both Core ML models remain loaded
+and alternate AB/BA order by row. PyTorch uses two CPU threads and one inter-op
+thread, with the Transformer fast path disabled. Timers cover PyTorch
+forward + softmax or synchronous Core ML prediction, excluding encoding,
+validation, loading, UI, and network. These compare deployment backends, not
+algorithms on equal hardware, and differ from the separate Swift timings at the end of this card.
+
+Upstream describes this synthetic split as disjoint from training/validation by
+form signature; those signatures were not independently re-audited here. No
+training, validation inference, test-based tuning, or hosted Jev/API comparison
+was performed. This measures supplied-option classification, not unseen real-world
+GUI completion or document extraction.
+
+[Full report](reports/synthetic-test.json), [complete compressed per-row trace](reports/synthetic-test-decisions.jsonl.gz),
+and [test manifest](synthetic-test.lock.json) retain the exact protocol, hashes,
+paired decisions, action metrics, and all failed numerical checks. Reproduce with
+`benchmark-synthetic.py --require-parity` in the [Mobius toolkit](https://github.com/FluidInference/mobius/tree/main/models/computer-use/cua-s1-forms/coreml#full-published-synthetic-test).
+That command exits 1 for the recorded numerical failures; it does not normalize
+scores or weaken the original gates. Treat these artifacts as under review
+until the numerical failures are resolved and validated.
+
+## ANE profile
+
+A separate September 19 profile uses the same portable-package hashes on the M5
+Pro, with real demo rows 0, 68, and 130 (27, 21, and 19 options). Each policy
+runs two warmup passes and ten timed passes, totaling 30 timed predictions. All
+120 timed predictions select the correct labels.
+
+| Policy | CPU ops | GPU ops | ANE ops | Warm p50 | Warm p95 |
+| --- | ---: | ---: | ---: | ---: | ---: |
+| `CPU_ONLY` | 173 | 0 | 0 | 1.527 ms | 1.602 ms |
+| `CPU_AND_GPU` | 0 | 173 | 0 | 0.929 ms | 2.380 ms |
+| `CPU_AND_NE` | 24 | 0 | 149 | 0.929 ms | 0.973 ms |
+| `ALL` | 0 | 173 | 0 | 0.912 ms | 1.229 ms |
+
+`CPU_AND_NE` assigns 86.1% of operations to ANE; `ALL` chooses the GPU on this
+Mac. CPU fallbacks cover integer/mask preparation and embedding gathers. These
+are public `MLComputePlan` preferred-device assignments, not measurements of
+utilization, energy, or time spent on each device. No Instruments runtime trace
+was captured.
+
+ANE model loading took 566.8 ms, followed by a 1.65 ms first prediction, with
+system caches retained. These are not first-install cold-start numbers. Warm
+timing includes Python model-call overhead and excludes encoding, Swift/UI work,
+and animation. This three-row timing manifest differs from the full conversion
+parity run above; no weights or graph were changed.
+
+See [reports/ane-profile.json](reports/ane-profile.json) for all operation
+assignments, individual timings, hashes, and the protocol;
+[reports/ane-fallback.json](reports/ane-fallback.json) records rejection reasons.
+Reproduce with `uv run --frozen python profile-coreml.py` in the
+[Mobius conversion directory](https://github.com/FluidInference/mobius/tree/main/models/computer-use/cua-s1-forms/coreml).
+
+## Optional higher-ANE variant
+
+The `ane-gather/` directory contains an alternative portable package and compiled
+bundle with the **same int32 inputs and float32 outputs** and all trained weights.
+The variant uses shared float16 mask inputs and unsigned 16-bit embedding indices
+to eliminate negative-index correction and place the gathers on ANE. Valid byte
+IDs 0–256 remain exact. It is **1,509,491 bytes** as a portable package.
+
+On this M5 Pro, the scheduler plan is **162 ANE operations and 3 CPU input casts
+(98.2% ANE)** for both `CPU_AND_NE` and `ALL`. The default model has 149 ANE and
+24 CPU operations (86.1%) under `CPU_AND_NE`. Counts are not runtime or energy
+shares, and host byte encoding still runs outside the model.
+
+The optional variant passes **196/196 decisions** against upstream on `ALL` and
+`CPU_AND_NE`, with maximum probability error **0.002336** under the unchanged
+0.005 tolerance. **28 Python regression tests** pass, including all byte-ID
+boundaries, full option capacity, truncation, and reordered choices. The Swift
+manager independently passes all 196 reference decisions, compiled-cache loading,
+and concurrent/reordered requests; the native demo passes its three-form checks.
+
+A matched same-process ABBA comparison uses three real inputs and 60 timed calls
+per model after warmup:
+
+| Artifact | CPU ops | ANE ops | Warm p50 | Warm p95 |
+| --- | ---: | ---: | ---: | ---: |
+| Root/default | 24 | 149 | 0.915 ms | 0.968 ms |
+| `ane-gather/` | 3 | 162 | 0.970 ms | 0.988 ms |
+
+Higher ANE placement is about **6% slower** in this local comparison, so the
+root/default artifact remains unchanged. No energy or CPU-time saving is claimed.
+The original input names, dtypes, shapes, and byte encoding still apply. Load
+`ane-gather/cua_s1_forms_fp16_options32.mlpackage` with the existing Python or
+Swift APIs, or pass its local path to the Swift demo's `--model` argument.
+
+Reports: [parity](reports/ane-gather-verification.json),
+[Swift validation](reports/swift-ane-validation.json),
+[compute plans](reports/ane-gather-profile.json),
+[fallbacks](reports/ane-gather-fallback.json), and
+[matched comparison](reports/ane-comparison.json). Reproduce with
+`uv run --frozen python convert-coreml.py --optimization ane-gather --output-dir build/ane-gather`
+in the [Mobius conversion directory](https://github.com/FluidInference/mobius/tree/main/models/computer-use/cua-s1-forms/coreml#optional-higher-ane-variant).
+
+## INT8 weight trial
+
+A matched run over all **24,370 synthetic decisions** on the same M5 Pro:
+
+| Export | Package size | Accuracy | Median | p95 |
+| --- | ---: | ---: | ---: | ---: |
+| Original FP16 | 1.51 MB | 99.9549% | 0.990 ms | 1.102 ms |
+| INT8 weights, FP16 compute | 0.81 MB | 99.9549% | 0.990 ms | 1.104 ms |
+
+**46.2% smaller**, with every selected option unchanged and effectively identical
+latency. Numerical parity still fails: 64 rows exceed the 0.005 probability-error
+limit (maximum 0.067738), versus 11 for FP16. No INT8 probability-sum violations
+were observed. The original remains the default.
+
+The experimental `int8-weights/cua_s1_forms_int8_options32.mlpackage` uses INT8
+weights with FP16 computation. Compile it locally or load it with the existing
+Swift `load(from:)` API. Inputs and outputs match the original model. This
+variant was measured on M5 Pro; iPhone behavior has not been measured.
+
+[Full report](reports/int8-synthetic-test.json) ·
+[Complete per-row trace](reports/int8-synthetic-test-decisions.jsonl.gz) ·
+[Quantization manifest](int8-weights/conversion.json) ·
+[Reproduction](https://github.com/FluidInference/mobius/tree/main/models/computer-use/cua-s1-forms/coreml#int8-weight-trial)
+
+The matched benchmark uses CPU+ANE, batch 1, three warmups/model, and one complete
+pass with alternating Core ML order; encoding/loading/UI are excluded. All models
+get the same 24,359 decisions correct. Public compute plans assign 149 ANE and
+32 CPU operations, with 19 unassigned constant-dequantization operations. This
+compression does not establish an INT8 activation path or a speed/energy gain.
+No calibration, training, or test-based tuning was performed. Numerical parity
+remains failed; the default FP16 artifacts are unchanged.
+
+## INT4 weight trial
+
+A matched run over all **24,370 synthetic decisions** on M5 Pro, CPU+ANE:
+
+| Export | Package size | Accuracy | Median | p95 |
+| --- | ---: | ---: | ---: | ---: |
+| FP16 control, iOS 18 target | 1.51 MB | 99.9549% | 0.982 ms | 1.081 ms |
+| INT4 weights, FP16 compute | 0.45 MB | 99.9302% | 0.982 ms | 1.082 ms |
+
+**70.1% smaller**, with essentially unchanged latency. INT4 makes **17 errors
+versus 11** for FP16: 14 choices change, introducing 10 errors and correcting four.
+Numerical parity fails: 356 rows exceed the 0.005 probability-error limit
+(maximum 0.754359); no INT4 probability-sum violations were observed.
+
+Packed INT4 requires **iOS 18/macOS 15**. Both exports use the same decomposed
+attention graph and FP16 computation. Batch-1 timing excludes encoding/loading/UI.
+INT8 preserves all choices at 0.81 MB; the original FP16 remains the default.
+
+The experimental INT4 package is
+`int4-weights/cua_s1_forms_int4_options32.mlpackage`. Its matched FP16 control is
+`int4-source-fp16/cua_s1_forms_fp16_options32.mlpackage`. Nineteen weight tensors
+use per-channel symmetric INT4 with a 2,048-element threshold; no calibration or
+retraining. This is weight compression with FP16 computation. The graph is
+retargeted from the verified original to preserve decomposed attention.
+
+[Full report](reports/int4-synthetic-test.json) ·
+[Complete per-row trace](reports/int4-synthetic-test-decisions.jsonl.gz) ·
+[Quantization manifest](int4-weights/conversion.json) ·
+[FP16 control manifest](int4-source-fp16/conversion.json) ·
+[Reproduction](https://github.com/FluidInference/mobius/tree/main/models/computer-use/cua-s1-forms/coreml#int4-weight-trial)
+
+The demo retains 196/196 choices but fails numerical parity. The full benchmark
+above exposes the accuracy loss. Public compute-plan placement is 149 ANE / 32 CPU
+operations, plus 19 unassigned constant dequantizations; no speed/energy claim.
+
+## 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/main/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).
+
+## Swift probability fix
+
+The Swift manager computes a stable softmax from live logits using Double arithmetic
+and returns Float `probabilities`. The model's original softmax output remains
+available as `rawProbabilities`, including FP16 rounding errors.
+
+All **73,110 real Swift API calls completed** on the pinned synthetic split
+(Apple M5 Pro, 24 GB, macOS 27.0; Swift 6.2.3, release build, CPU+ANE):
+
+| Variant | Calls completed | Correct decisions | Changed choices after fix | Probability-sum failures after fix |
+| --- | ---: | ---: | ---: | ---: |
+| FP16 | 24,370 | 24,359 (99.9549%) | 0 | 0 |
+| INT8 | 24,370 | 24,359 (99.9549%) | 0 | 0 |
+| INT4 | 24,370 | 24,353 (99.9302%) | 0 | 0 |
+
+The recorded FP16 sum failure is fixed. Stable probabilities agree with an
+independent float64 softmax within **0.000000030**. Raw conversion-parity failures
+and INT4's accuracy loss remain. This validation changes no model artifacts and
+makes no new latency claim. Earlier complete Swift timings predate this fix.
+
+[Runtime report](reports/swift-runtime/report.json) · [Full runtime trace](reports/swift-runtime/decisions.jsonl.gz)
+The report pins the dataset, packages, saved reference traces, Swift sources,
+and validation harness by SHA-256.
+
+Reproduce with the [Mobius Swift validation harness](https://github.com/FluidInference/mobius/tree/main/models/computer-use/cua-s1-forms/coreml#swift-probability-fix).
+
+## Live browser proof and expanded Swift benchmark
+
+[![Actual CUA-driven browser forms](demo/browser-demo.gif)](demo/browser-demo.mp4)
+
+The [native Swift browser demo](https://github.com/FluidInference/FluidAudio/tree/7f9eb92b0af8594c4e048a9e57f697340aacfa67/Examples/CuaS1FormsDemo)
+loads both variants into independent WKWebViews. It reads actual DOM labels,
+roles and state, asks the model for a choice, applies compatible fill/check
+actions, dispatches events, and independently verifies the resulting DOM.
+Source values are user-entered or supplied by the original public examples.
+HTML contains controls, not source values or expected choices. The
+[recording](https://github.com/FluidInference/FluidAudio/blob/7f9eb92b0af8594c4e048a9e57f697340aacfa67/Examples/CuaS1FormsDemo/browser-demo.mp4)
+shows patient, job and insurance forms: **100/100 original decisions** across
+both models, with event-count, stale-observation and explicit-click checks.
+Full actual contexts/candidates/actions are in [browser-validation.json](reports/browser-validation.json).
+This is bounded local browser automation, not arbitrary desktop control or PDF extraction.
+
+The expanded **release Swift** comparison uses all 50 initial controls, both
+models resident, one warmup pass per model and ABBA with two full passes per
+block (200 timed calls/model). All 400 choices match upstream labels. On the
+M5 Pro / 24 GB / macOS 27.0 (26A428), original median/p95 is **0.912/0.933 ms**;
+ANE gather is **0.961/0.984 ms**, about 5.4% slower by median. This timer includes
+Swift encoding + Core ML + output decoding and excludes browser/rendering/animation.
+See [swift-variant-comparison.json](reports/swift-variant-comparison.json) for
+raw samples, exact model hashes, per-form statistics, and load/first-call costs.
+The earlier compute-plan counts still apply to these unchanged artifacts;
+no utilization, energy saving or held-out accuracy claim is made.
+
+Reproduce from FluidAudio commit `7f9eb92b0af8594c4e048a9e57f697340aacfa67` (the example is
+retained in history and is not part of the current library PR):
+
+```bash
+git worktree add --detach /tmp/cua-s1-browser-repro 7f9eb92b0af8594c4e048a9e57f697340aacfa67
+cd /tmp/cua-s1-browser-repro
+Examples/CuaS1FormsDemo/run.sh --browser
+swift run --package-path Examples/CuaS1FormsDemo -c release CuaS1FormsDemo \
+  --benchmark --report /absolute/path/to/variant-comparison.json \
+  --hardware "Describe the measured Mac"
+```
+
+Both packages are fetched by pinned revisions and verified hashes, or supplied
+with `--model /path/to/original.mlpackage --ane-model /path/to/ane-gather.mlpackage`.
+The [demo README](https://github.com/FluidInference/FluidAudio/tree/7f9eb92b0af8594c4e048a9e57f697340aacfa67/Examples/CuaS1FormsDemo#matched-swift-benchmark)
+also documents real-browser recording and its separate validation trace.

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.

ane-gather/conversion.json

@@ -0,0 +1,36 @@
+{
+  "model": "cua_s1_forms_fp16_options32.mlpackage",
+  "precision": "float16",
+  "optimization": "ane-gather",
+  "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": 0.6771400420111604,
+  "python": "3.11.11",
+  "torch": "2.7.0",
+  "coremltools": "9.0",
+  "package_files": {
+    "Data/com.apple.CoreML/model.mlmodel": "de18e313c3b625e35d008ed8b6b24108edc6df7bf2fae9533af03519eca11b63",
+    "Data/com.apple.CoreML/weights/weight.bin": "4da9259f798e44f5a1b50769ee1916fd3747c4d723dd9997b516c7fe238c7895",
+    "Manifest.json": "38f812a04eb2322080634ba788c61df362336168466ca67e5549058d346ac793"
+  }
+}

ane-gather/cua_s1_forms_fp16_options32.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:526dccb87bdc036df1e8add0b50dd6540db534808fc2f8caa97cd897b4fc1fa3
+size 243

ane-gather/cua_s1_forms_fp16_options32.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0f2a8ed9fe8abcc39c7ee0575e2562d589d459683b92708fa9788b533be3038b
+size 876

ane-gather/cua_s1_forms_fp16_options32.mlmodelc/model.mil

@@ -0,0 +1,404 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3600.16.1"}, {"coremlc-version", "3600.25.2"}, {"coremltools-component-milinternal", ""}, {"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<string, []> mask_options_to_fp16_dtype_0 = const()[name = tensor<string, []>("mask_options_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, []> var_29_promoted_to_fp16 = const()[name = tensor<string, []>("op_29_promoted_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 32]> option_mask_to_fp16 = cast(dtype = mask_options_to_fp16_dtype_0, x = option_mask)[name = tensor<string, []>("cast_68")];
+            tensor<bool, [1, 32]> option_mask_cast_fp16 = not_equal(x = option_mask_to_fp16, y = var_29_promoted_to_fp16)[name = tensor<string, []>("option_mask_cast_fp16")];
+            tensor<string, []> mask_context_ids_to_fp16_dtype_0 = const()[name = tensor<string, []>("mask_context_ids_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, []> var_31_promoted_to_fp16 = const()[name = tensor<string, []>("op_31_promoted_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 224]> context_ids_to_fp16 = cast(dtype = mask_context_ids_to_fp16_dtype_0, x = context_ids)[name = tensor<string, []>("cast_67")];
+            tensor<bool, [1, 224]> context_mask_cast_fp16 = not_equal(x = context_ids_to_fp16, y = var_31_promoted_to_fp16)[name = tensor<string, []>("context_mask_cast_fp16")];
+            tensor<int32, [2]> var_42_begin_0 = const()[name = tensor<string, []>("op_42_begin_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [2]> var_42_end_0 = const()[name = tensor<string, []>("op_42_end_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<bool, [2]> var_42_end_mask_0 = const()[name = tensor<string, []>("op_42_end_mask_0"), val = tensor<bool, [2]>([true, false])];
+            tensor<fp16, [1, 1]> var_42_cast_fp16 = slice_by_index(begin = var_42_begin_0, end = var_42_end_0, end_mask = var_42_end_mask_0, x = context_ids_to_fp16)[name = tensor<string, []>("op_42_cast_fp16")];
+            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_42_cast_fp16, value = fill_like_0_value_0_to_fp16)[name = tensor<string, []>("fill_like_0_cast_fp16")];
+            tensor<int32, [2]> var_58_begin_0 = const()[name = tensor<string, []>("op_58_begin_0"), val = tensor<int32, [2]>([0, 1])];
+            tensor<int32, [2]> var_58_end_0 = const()[name = tensor<string, []>("op_58_end_0"), val = tensor<int32, [2]>([1, 224])];
+            tensor<bool, [2]> var_58_end_mask_0 = const()[name = tensor<string, []>("op_58_end_mask_0"), val = tensor<bool, [2]>([true, true])];
+            tensor<fp16, [1, 223]> var_58_cast_fp16 = slice_by_index(begin = var_58_begin_0, end = var_58_end_0, end_mask = var_58_end_mask_0, x = context_ids_to_fp16)[name = tensor<string, []>("op_58_cast_fp16")];
+            tensor<int32, []> var_60 = const()[name = tensor<string, []>("op_60"), 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<fp16, [1, 224]> safe_context_ids_cast_fp16 = concat(axis = var_60, interleave = safe_context_ids_interleave_0, values = (fill_like_0_cast_fp16, var_58_cast_fp16))[name = tensor<string, []>("safe_context_ids_cast_fp16")];
+            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, []> cast_0_dtype_0 = const()[name = tensor<string, []>("cast_0_dtype_0"), val = tensor<string, []>("uint16")];
+            tensor<uint16, [1, 224]> cast_0 = cast(dtype = cast_0_dtype_0, x = context_ids_to_fp16)[name = tensor<string, []>("cast_0")];
+            tensor<int32, []> gather_0_axis_0 = const()[name = tensor<string, []>("gather_0_axis_0"), val = tensor<int32, []>(0)];
+            tensor<int32, []> gather_0_batch_dims_0 = const()[name = tensor<string, []>("gather_0_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> gather_0_validate_indices_0 = const()[name = tensor<string, []>("gather_0_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 224, 128]> gather_0 = gather(axis = gather_0_axis_0, batch_dims = gather_0_batch_dims_0, indices = cast_0, validate_indices = gather_0_validate_indices_0, x = model_embedding_weight_to_fp16)[name = tensor<string, []>("gather_0")];
+            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 = gather_0, y = model_position_weight_to_fp16)[name = tensor<string, []>("src_1_cast_fp16")];
+            tensor<fp16, []> var_81_promoted_to_fp16 = const()[name = tensor<string, []>("op_81_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_81_promoted_to_fp16)[name = tensor<string, []>("mask_1_cast_fp16")];
+            tensor<fp16, []> var_97_to_fp16 = const()[name = tensor<string, []>("op_97_to_fp16"), val = tensor<fp16, []>(-inf)];
+            tensor<fp16, [1, 224]> var_105_to_fp16 = const()[name = tensor<string, []>("op_105_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_97_to_fp16, b = var_105_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_84_to_fp16 = const()[name = tensor<string, []>("op_84_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_84_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_139_cast_fp16 = reshape(shape = concat_0, x = linear_0_cast_fp16)[name = tensor<string, []>("op_139_cast_fp16")];
+            tensor<int32, [1]> var_140_axes_0 = const()[name = tensor<string, []>("op_140_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 224, 1, 3, 128]> var_140_cast_fp16 = expand_dims(axes = var_140_axes_0, x = var_139_cast_fp16)[name = tensor<string, []>("op_140_cast_fp16")];
+            tensor<int32, [5]> var_141_perm_0 = const()[name = tensor<string, []>("op_141_perm_0"), val = tensor<int32, [5]>([-2, 1, 2, 0, 4])];
+            tensor<int32, [1]> var_142_axes_0 = const()[name = tensor<string, []>("op_142_axes_0"), val = tensor<int32, [1]>([-2])];
+            tensor<fp16, [3, 224, 1, 1, 128]> var_141_cast_fp16 = transpose(perm = var_141_perm_0, x = var_140_cast_fp16)[name = tensor<string, []>("transpose_23")];
+            tensor<fp16, [3, 224, 1, 128]> var_142_cast_fp16 = squeeze(axes = var_142_axes_0, x = var_141_cast_fp16)[name = tensor<string, []>("op_142_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_142_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_142_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_142_cast_fp16)[name = tensor<string, []>("v_1_cast_fp16")];
+            tensor<int32, [3]> var_150 = const()[name = tensor<string, []>("op_150"), val = tensor<int32, [3]>([224, 4, 32])];
+            tensor<fp16, [224, 4, 32]> var_151_cast_fp16 = reshape(shape = var_150, x = q_1_cast_fp16)[name = tensor<string, []>("op_151_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_157 = const()[name = tensor<string, []>("op_157"), val = tensor<int32, [3]>([224, 4, 32])];
+            tensor<fp16, [224, 4, 32]> var_158_cast_fp16 = reshape(shape = var_157, x = k_1_cast_fp16)[name = tensor<string, []>("op_158_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_164 = const()[name = tensor<string, []>("op_164"), val = tensor<int32, [3]>([224, 4, 32])];
+            tensor<fp16, [224, 4, 32]> var_165_cast_fp16 = reshape(shape = var_164, x = v_1_cast_fp16)[name = tensor<string, []>("op_165_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_168 = const()[name = tensor<string, []>("op_168"), val = tensor<int32, [4]>([1, 1, 1, 224])];
+            tensor<fp16, [1, 1, 1, 224]> var_169_cast_fp16 = reshape(shape = var_168, x = key_padding_mask_1_cast_fp16)[name = tensor<string, []>("op_169_cast_fp16")];
+            tensor<int32, [4]> var_171_reps_0 = const()[name = tensor<string, []>("op_171_reps_0"), val = tensor<int32, [4]>([1, 4, 1, 1])];
+            tensor<fp16, [1, 4, 1, 224]> var_171_cast_fp16 = tile(reps = var_171_reps_0, x = var_169_cast_fp16)[name = tensor<string, []>("op_171_cast_fp16")];
+            tensor<int32, [4]> var_179 = const()[name = tensor<string, []>("op_179"), 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_151_cast_fp16)[name = tensor<string, []>("transpose_22")];
+            tensor<fp16, [1, 4, 224, 32]> q_5_cast_fp16 = reshape(shape = var_179, x = q_3_cast_fp16)[name = tensor<string, []>("q_5_cast_fp16")];
+            tensor<int32, [4]> var_181 = const()[name = tensor<string, []>("op_181"), 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_158_cast_fp16)[name = tensor<string, []>("transpose_21")];
+            tensor<fp16, [1, 4, 224, 32]> k_5_cast_fp16 = reshape(shape = var_181, x = k_3_cast_fp16)[name = tensor<string, []>("k_5_cast_fp16")];
+            tensor<int32, [4]> var_183 = const()[name = tensor<string, []>("op_183"), 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_165_cast_fp16)[name = tensor<string, []>("transpose_20")];
+            tensor<fp16, [1, 4, 224, 32]> v_5_cast_fp16 = reshape(shape = var_183, 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_171_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_186 = const()[name = tensor<string, []>("op_186"), val = tensor<int32, [4]>([2, 0, 1, 3])];
+            tensor<int32, [2]> var_191 = const()[name = tensor<string, []>("op_191"), val = tensor<int32, [2]>([224, 128])];
+            tensor<fp16, [224, 1, 4, 32]> var_187_cast_fp16 = transpose(perm = var_186, x = attn_output_1_cast_fp16)[name = tensor<string, []>("transpose_19")];
+            tensor<fp16, [224, 128]> attn_output_3_cast_fp16 = reshape(shape = var_191, x = var_187_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_195 = const()[name = tensor<string, []>("op_195"), val = tensor<int32, [3]>([224, 1, 128])];
+            tensor<fp16, [224, 1, 128]> attn_output_7_cast_fp16 = reshape(shape = var_195, 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_84_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_84_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_246_cast_fp16 = reshape(shape = concat_2, x = linear_4_cast_fp16)[name = tensor<string, []>("op_246_cast_fp16")];
+            tensor<int32, [1]> var_247_axes_0 = const()[name = tensor<string, []>("op_247_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 224, 1, 3, 128]> var_247_cast_fp16 = expand_dims(axes = var_247_axes_0, x = var_246_cast_fp16)[name = tensor<string, []>("op_247_cast_fp16")];
+            tensor<int32, [5]> var_248_perm_0 = const()[name = tensor<string, []>("op_248_perm_0"), val = tensor<int32, [5]>([-2, 1, 2, 0, 4])];
+            tensor<int32, [1]> var_249_axes_0 = const()[name = tensor<string, []>("op_249_axes_0"), val = tensor<int32, [1]>([-2])];
+            tensor<fp16, [3, 224, 1, 1, 128]> var_248_cast_fp16 = transpose(perm = var_248_perm_0, x = var_247_cast_fp16)[name = tensor<string, []>("transpose_16")];
+            tensor<fp16, [3, 224, 1, 128]> var_249_cast_fp16 = squeeze(axes = var_249_axes_0, x = var_248_cast_fp16)[name = tensor<string, []>("op_249_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_249_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_249_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_249_cast_fp16)[name = tensor<string, []>("v_7_cast_fp16")];
+            tensor<int32, [3]> var_257 = const()[name = tensor<string, []>("op_257"), val = tensor<int32, [3]>([224, 4, 32])];
+            tensor<fp16, [224, 4, 32]> var_258_cast_fp16 = reshape(shape = var_257, x = q_7_cast_fp16)[name = tensor<string, []>("op_258_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_264 = const()[name = tensor<string, []>("op_264"), val = tensor<int32, [3]>([224, 4, 32])];
+            tensor<fp16, [224, 4, 32]> var_265_cast_fp16 = reshape(shape = var_264, x = k_7_cast_fp16)[name = tensor<string, []>("op_265_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_271 = const()[name = tensor<string, []>("op_271"), val = tensor<int32, [3]>([224, 4, 32])];
+            tensor<fp16, [224, 4, 32]> var_272_cast_fp16 = reshape(shape = var_271, x = v_7_cast_fp16)[name = tensor<string, []>("op_272_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_286 = const()[name = tensor<string, []>("op_286"), 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_258_cast_fp16)[name = tensor<string, []>("transpose_15")];
+            tensor<fp16, [1, 4, 224, 32]> q_11_cast_fp16 = reshape(shape = var_286, x = q_9_cast_fp16)[name = tensor<string, []>("q_11_cast_fp16")];
+            tensor<int32, [4]> var_288 = const()[name = tensor<string, []>("op_288"), 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_265_cast_fp16)[name = tensor<string, []>("transpose_14")];
+            tensor<fp16, [1, 4, 224, 32]> k_11_cast_fp16 = reshape(shape = var_288, x = k_9_cast_fp16)[name = tensor<string, []>("k_11_cast_fp16")];
+            tensor<int32, [4]> var_290 = const()[name = tensor<string, []>("op_290"), 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_272_cast_fp16)[name = tensor<string, []>("transpose_13")];
+            tensor<fp16, [1, 4, 224, 32]> v_11_cast_fp16 = reshape(shape = var_290, 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_171_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_293 = const()[name = tensor<string, []>("op_293"), val = tensor<int32, [4]>([2, 0, 1, 3])];
+            tensor<int32, [2]> var_298 = const()[name = tensor<string, []>("op_298"), val = tensor<int32, [2]>([224, 128])];
+            tensor<fp16, [224, 1, 4, 32]> var_294_cast_fp16 = transpose(perm = var_293, x = attn_output_9_cast_fp16)[name = tensor<string, []>("transpose_12")];
+            tensor<fp16, [224, 128]> attn_output_11_cast_fp16 = reshape(shape = var_298, x = var_294_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_302 = const()[name = tensor<string, []>("op_302"), val = tensor<int32, [3]>([224, 1, 128])];
+            tensor<fp16, [224, 1, 128]> attn_output_15_cast_fp16 = reshape(shape = var_302, 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_84_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_340 = const()[name = tensor<string, []>("op_340"), val = tensor<int32, [2]>([32, 96])];
+            tensor<string, []> mask_option_ids_to_fp16_dtype_0 = const()[name = tensor<string, []>("mask_option_ids_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1, 32, 96]> option_ids_to_fp16 = cast(dtype = mask_option_ids_to_fp16_dtype_0, x = option_ids)[name = tensor<string, []>("cast_63")];
+            tensor<fp16, [32, 96]> mask_flat_ids_cast_fp16 = reshape(shape = var_340, x = option_ids_to_fp16)[name = tensor<string, []>("mask_flat_ids_cast_fp16")];
+            tensor<fp16, []> var_342_promoted_to_fp16 = const()[name = tensor<string, []>("op_342_promoted_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<bool, [32, 96]> flat_mask_cast_fp16 = not_equal(x = mask_flat_ids_cast_fp16, y = var_342_promoted_to_fp16)[name = tensor<string, []>("flat_mask_cast_fp16")];
+            tensor<int32, [2]> var_353_begin_0 = const()[name = tensor<string, []>("op_353_begin_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [2]> var_353_end_0 = const()[name = tensor<string, []>("op_353_end_0"), val = tensor<int32, [2]>([32, 1])];
+            tensor<bool, [2]> var_353_end_mask_0 = const()[name = tensor<string, []>("op_353_end_mask_0"), val = tensor<bool, [2]>([true, false])];
+            tensor<fp16, [32, 1]> var_353_cast_fp16 = slice_by_index(begin = var_353_begin_0, end = var_353_end_0, end_mask = var_353_end_mask_0, x = mask_flat_ids_cast_fp16)[name = tensor<string, []>("op_353_cast_fp16")];
+            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_353_cast_fp16, value = fill_like_1_value_0_to_fp16)[name = tensor<string, []>("fill_like_1_cast_fp16")];
+            tensor<int32, [2]> var_369_begin_0 = const()[name = tensor<string, []>("op_369_begin_0"), val = tensor<int32, [2]>([0, 1])];
+            tensor<int32, [2]> var_369_end_0 = const()[name = tensor<string, []>("op_369_end_0"), val = tensor<int32, [2]>([32, 96])];
+            tensor<bool, [2]> var_369_end_mask_0 = const()[name = tensor<string, []>("op_369_end_mask_0"), val = tensor<bool, [2]>([true, true])];
+            tensor<fp16, [32, 95]> var_369_cast_fp16 = slice_by_index(begin = var_369_begin_0, end = var_369_end_0, end_mask = var_369_end_mask_0, x = mask_flat_ids_cast_fp16)[name = tensor<string, []>("op_369_cast_fp16")];
+            tensor<int32, []> var_371 = const()[name = tensor<string, []>("op_371"), val = tensor<int32, []>(1)];
+            tensor<bool, []> safe_ids_interleave_0 = const()[name = tensor<string, []>("safe_ids_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [32, 96]> safe_ids_cast_fp16 = concat(axis = var_371, interleave = safe_ids_interleave_0, values = (fill_like_1_cast_fp16, var_369_cast_fp16))[name = tensor<string, []>("safe_ids_cast_fp16")];
+            tensor<int32, [2]> reshape_0_shape_0 = const()[name = tensor<string, []>("reshape_0_shape_0"), val = tensor<int32, [2]>([32, 96])];
+            tensor<fp16, [32, 96]> reshape_0 = reshape(shape = reshape_0_shape_0, x = option_ids_to_fp16)[name = tensor<string, []>("reshape_0")];
+            tensor<string, []> cast_1_dtype_0 = const()[name = tensor<string, []>("cast_1_dtype_0"), val = tensor<string, []>("uint16")];
+            tensor<uint16, [32, 96]> cast_1 = cast(dtype = cast_1_dtype_0, x = reshape_0)[name = tensor<string, []>("cast_1")];
+            tensor<int32, []> gather_1_axis_0 = const()[name = tensor<string, []>("gather_1_axis_0"), val = tensor<int32, []>(0)];
+            tensor<int32, []> gather_1_batch_dims_0 = const()[name = tensor<string, []>("gather_1_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> gather_1_validate_indices_0 = const()[name = tensor<string, []>("gather_1_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [32, 96, 128]> gather_1 = gather(axis = gather_1_axis_0, batch_dims = gather_1_batch_dims_0, indices = cast_1, validate_indices = gather_1_validate_indices_0, x = model_embedding_weight_to_fp16)[name = tensor<string, []>("gather_1")];
+            tensor<fp16, [96, 128]> var_389_to_fp16 = const()[name = tensor<string, []>("op_389_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 = gather_1, y = var_389_to_fp16)[name = tensor<string, []>("src_cast_fp16")];
+            tensor<fp16, []> var_392_promoted_to_fp16 = const()[name = tensor<string, []>("op_392_promoted_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<bool, [32, 96]> mask_cast_fp16 = equal(x = safe_ids_cast_fp16, y = var_392_promoted_to_fp16)[name = tensor<string, []>("mask_cast_fp16")];
+            tensor<fp16, []> var_408_to_fp16 = const()[name = tensor<string, []>("op_408_to_fp16"), val = tensor<fp16, []>(-inf)];
+            tensor<fp16, [32, 96]> var_414_to_fp16 = const()[name = tensor<string, []>("op_414_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_408_to_fp16, b = var_414_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_395_to_fp16 = const()[name = tensor<string, []>("op_395_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_395_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_448_cast_fp16 = reshape(shape = concat_4, x = linear_8_cast_fp16)[name = tensor<string, []>("op_448_cast_fp16")];
+            tensor<int32, [1]> var_449_axes_0 = const()[name = tensor<string, []>("op_449_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 96, 32, 3, 128]> var_449_cast_fp16 = expand_dims(axes = var_449_axes_0, x = var_448_cast_fp16)[name = tensor<string, []>("op_449_cast_fp16")];
+            tensor<int32, [5]> var_450_perm_0 = const()[name = tensor<string, []>("op_450_perm_0"), val = tensor<int32, [5]>([-2, 1, 2, 0, 4])];
+            tensor<int32, [1]> var_451_axes_0 = const()[name = tensor<string, []>("op_451_axes_0"), val = tensor<int32, [1]>([-2])];
+            tensor<fp16, [3, 96, 32, 1, 128]> var_450_cast_fp16 = transpose(perm = var_450_perm_0, x = var_449_cast_fp16)[name = tensor<string, []>("transpose_9")];
+            tensor<fp16, [3, 96, 32, 128]> var_451_cast_fp16 = squeeze(axes = var_451_axes_0, x = var_450_cast_fp16)[name = tensor<string, []>("op_451_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_451_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_451_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_451_cast_fp16)[name = tensor<string, []>("v_13_cast_fp16")];
+            tensor<int32, [3]> var_459 = const()[name = tensor<string, []>("op_459"), val = tensor<int32, [3]>([96, 128, 32])];
+            tensor<fp16, [96, 128, 32]> var_460_cast_fp16 = reshape(shape = var_459, x = q_13_cast_fp16)[name = tensor<string, []>("op_460_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_466 = const()[name = tensor<string, []>("op_466"), val = tensor<int32, [3]>([96, 128, 32])];
+            tensor<fp16, [96, 128, 32]> var_467_cast_fp16 = reshape(shape = var_466, x = k_13_cast_fp16)[name = tensor<string, []>("op_467_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_473 = const()[name = tensor<string, []>("op_473"), val = tensor<int32, [3]>([96, 128, 32])];
+            tensor<fp16, [96, 128, 32]> var_474_cast_fp16 = reshape(shape = var_473, x = v_13_cast_fp16)[name = tensor<string, []>("op_474_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_477 = const()[name = tensor<string, []>("op_477"), val = tensor<int32, [4]>([32, 1, 1, 96])];
+            tensor<fp16, [32, 1, 1, 96]> var_478_cast_fp16 = reshape(shape = var_477, x = key_padding_mask_7_cast_fp16)[name = tensor<string, []>("op_478_cast_fp16")];
+            tensor<int32, [4]> var_480_reps_0 = const()[name = tensor<string, []>("op_480_reps_0"), val = tensor<int32, [4]>([1, 4, 1, 1])];
+            tensor<fp16, [32, 4, 1, 96]> var_480_cast_fp16 = tile(reps = var_480_reps_0, x = var_478_cast_fp16)[name = tensor<string, []>("op_480_cast_fp16")];
+            tensor<int32, [4]> var_488 = const()[name = tensor<string, []>("op_488"), 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_460_cast_fp16)[name = tensor<string, []>("transpose_8")];
+            tensor<fp16, [32, 4, 96, 32]> q_cast_fp16 = reshape(shape = var_488, x = q_15_cast_fp16)[name = tensor<string, []>("q_cast_fp16")];
+            tensor<int32, [4]> var_490 = const()[name = tensor<string, []>("op_490"), 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_467_cast_fp16)[name = tensor<string, []>("transpose_7")];
+            tensor<fp16, [32, 4, 96, 32]> k_cast_fp16 = reshape(shape = var_490, x = k_15_cast_fp16)[name = tensor<string, []>("k_cast_fp16")];
+            tensor<int32, [4]> var_492 = const()[name = tensor<string, []>("op_492"), 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_474_cast_fp16)[name = tensor<string, []>("transpose_6")];
+            tensor<fp16, [32, 4, 96, 32]> v_cast_fp16 = reshape(shape = var_492, 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_480_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_495 = const()[name = tensor<string, []>("op_495"), val = tensor<int32, [4]>([2, 0, 1, 3])];
+            tensor<int32, [2]> var_500 = const()[name = tensor<string, []>("op_500"), val = tensor<int32, [2]>([3072, 128])];
+            tensor<fp16, [96, 32, 4, 32]> var_496_cast_fp16 = transpose(perm = var_495, x = attn_output_17_cast_fp16)[name = tensor<string, []>("transpose_5")];
+            tensor<fp16, [3072, 128]> attn_output_19_cast_fp16 = reshape(shape = var_500, x = var_496_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_504 = const()[name = tensor<string, []>("op_504"), val = tensor<int32, [3]>([96, 32, 128])];
+            tensor<fp16, [96, 32, 128]> attn_output_cast_fp16 = reshape(shape = var_504, 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_395_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_524_axes_0 = const()[name = tensor<string, []>("op_524_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, [32, 96, 1]> var_524 = expand_dims(axes = var_524_axes_0, x = flat_mask_cast_fp16)[name = tensor<string, []>("op_524")];
+            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_524_to_fp16 = cast(dtype = weights_to_fp16_dtype_0, x = var_524)[name = tensor<string, []>("cast_59")];
+            tensor<fp16, [32, 96, 128]> var_530_cast_fp16 = mul(x = hidden_cast_fp16, y = var_524_to_fp16)[name = tensor<string, []>("op_530_cast_fp16")];
+            tensor<int32, [1]> var_535_axes_0 = const()[name = tensor<string, []>("op_535_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, []> var_535_keep_dims_0 = const()[name = tensor<string, []>("op_535_keep_dims_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [32, 128]> var_535_cast_fp16 = reduce_sum(axes = var_535_axes_0, keep_dims = var_535_keep_dims_0, x = var_530_cast_fp16)[name = tensor<string, []>("op_535_cast_fp16")];
+            tensor<int32, [1]> var_540_axes_0 = const()[name = tensor<string, []>("op_540_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, []> var_540_keep_dims_0 = const()[name = tensor<string, []>("op_540_keep_dims_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [32, 1]> var_540_cast_fp16 = reduce_sum(axes = var_540_axes_0, keep_dims = var_540_keep_dims_0, x = var_524_to_fp16)[name = tensor<string, []>("op_540_cast_fp16")];
+            tensor<fp16, []> var_541_to_fp16 = const()[name = tensor<string, []>("op_541_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [32, 1]> var_542_cast_fp16 = maximum(x = var_540_cast_fp16, y = var_541_to_fp16)[name = tensor<string, []>("op_542_cast_fp16")];
+            tensor<fp16, [32, 128]> pooled_cast_fp16 = real_div(x = var_535_cast_fp16, y = var_542_cast_fp16)[name = tensor<string, []>("pooled_cast_fp16")];
+            tensor<int32, [3]> var_545 = const()[name = tensor<string, []>("op_545"), val = tensor<int32, [3]>([1, 32, -1])];
+            tensor<fp16, [1, 32, 128]> options_cast_fp16 = reshape(shape = var_545, x = pooled_cast_fp16)[name = tensor<string, []>("options_cast_fp16")];
+            tensor<int32, []> var_548 = const()[name = tensor<string, []>("op_548"), 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_556_to_fp16 = const()[name = tensor<string, []>("op_556_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_556_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_556_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_587_axes_0 = const()[name = tensor<string, []>("op_587_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, 224]> var_587 = expand_dims(axes = var_587_axes_0, x = context_mask_cast_fp16)[name = tensor<string, []>("op_587")];
+            tensor<bool, [1, 1, 224]> var_589 = logical_not(x = var_587)[name = tensor<string, []>("op_589")];
+            tensor<fp16, []> var_549_to_fp16 = const()[name = tensor<string, []>("op_549_to_fp16"), val = tensor<fp16, []>(-inf)];
+            tensor<fp16, [1, 32, 224]> scores_cast_fp16 = select(a = var_549_to_fp16, b = _inversed_scores_1_cast_fp16, cond = var_589)[name = tensor<string, []>("scores_cast_fp16")];
+            tensor<fp16, [1, 32, 224]> var_591_cast_fp16 = softmax(axis = var_548, x = scores_cast_fp16)[name = tensor<string, []>("op_591_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_591_cast_fp16, y = linear_14_cast_fp16)[name = tensor<string, []>("matmul_4_cast_fp16")];
+            tensor<fp16, [1, 32, 128]> var_594_cast_fp16 = mul(x = linear_12_cast_fp16, y = matmul_4_cast_fp16)[name = tensor<string, []>("op_594_cast_fp16")];
+            tensor<int32, [1]> var_596_axes_0 = const()[name = tensor<string, []>("op_596_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, []> var_596_keep_dims_0 = const()[name = tensor<string, []>("op_596_keep_dims_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 32]> var_596_cast_fp16 = reduce_sum(axes = var_596_axes_0, keep_dims = var_596_keep_dims_0, x = var_594_cast_fp16)[name = tensor<string, []>("op_596_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_596_cast_fp16, y = _inversed_logits_1_y_0_to_fp16)[name = tensor<string, []>("_inversed_logits_1_cast_fp16")];
+            tensor<bool, [1, 32]> var_599 = logical_not(x = option_mask_cast_fp16)[name = tensor<string, []>("op_599")];
+            tensor<fp16, [1, 32]> logits_3_cast_fp16 = select(a = var_549_to_fp16, b = _inversed_logits_1_cast_fp16, cond = var_599)[name = tensor<string, []>("logits_3_cast_fp16")];
+            tensor<fp16, []> var_607_value_0_to_fp16 = const()[name = tensor<string, []>("op_607_value_0_to_fp16"), val = tensor<fp16, []>(-0x1.388p+13)];
+            tensor<fp16, [1, 32]> var_607_cast_fp16 = fill_like(ref_tensor = logits_3_cast_fp16, value = var_607_value_0_to_fp16)[name = tensor<string, []>("op_607_cast_fp16")];
+            tensor<fp16, [1, 32]> logits_cast_fp16 = select(a = logits_3_cast_fp16, b = var_607_cast_fp16, cond = option_mask_cast_fp16)[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_609 = const()[name = tensor<string, []>("op_609"), val = tensor<int32, []>(-1)];
+            tensor<fp16, [1, 32]> var_611_cast_fp16 = softmax(axis = var_609, x = logits_cast_fp16)[name = tensor<string, []>("op_611_cast_fp16")];
+            tensor<string, []> var_611_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_611_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<fp32, [1, 32]> probabilities = cast(dtype = var_611_cast_fp16_to_fp32_dtype_0, x = var_611_cast_fp16)[name = tensor<string, []>("cast_57")];
+            tensor<fp32, [1, 32]> logits = cast(dtype = logits_cast_fp16_to_fp32_dtype_0, x = logits_cast_fp16)[name = tensor<string, []>("cast_58")];
+        } -> (logits, probabilities);
+}

ane-gather/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

ane-gather/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:de18e313c3b625e35d008ed8b6b24108edc6df7bf2fae9533af03519eca11b63
+size 62154

ane-gather/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

ane-gather/cua_s1_forms_fp16_options32.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "1CB24175-461C-4E33-BE2C-F580A319AEC8": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "869A4EBA-1FCF-41FC-8DD9-70ECEF96023D": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "869A4EBA-1FCF-41FC-8DD9-70ECEF96023D"
+}

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 +1,66 @@
+{
+  "LICENSE": "c0779290c1d4783169aa3dbfb55feb505e563ef8a004bbf55298ceffcfbda8d9",
+  "NOTICES.md": "027c72741eaa695e60d7b6cebd3666372c81d443a96b673b72a897cf432ddfd0",
+  "README.md": "377a7149ce6de7064b3a7d26de00940dd92f1ce7f7935296a735435942cae618",
+  "UPSTREAM-THIRD-PARTY-NOTICES.md": "4091e69b45c8cc97e30a066fbbd56148dbef66ab716432c048d2333c9c464213",
+  "ane-gather/conversion.json": "4d569a4020f1e9711da3f7fff0d11acef1812c4652414cf64728c77bc2bae4c4",
+  "ane-gather/cua_s1_forms_fp16_options32.mlmodelc/analytics/coremldata.bin": "526dccb87bdc036df1e8add0b50dd6540db534808fc2f8caa97cd897b4fc1fa3",
+  "ane-gather/cua_s1_forms_fp16_options32.mlmodelc/coremldata.bin": "0f2a8ed9fe8abcc39c7ee0575e2562d589d459683b92708fa9788b533be3038b",
+  "ane-gather/cua_s1_forms_fp16_options32.mlmodelc/model.mil": "578b5ef1e50c98d0577459a3a4c020f4fd83d7e465213d1263ddf1a021fef077",
+  "ane-gather/cua_s1_forms_fp16_options32.mlmodelc/weights/weight.bin": "4da9259f798e44f5a1b50769ee1916fd3747c4d723dd9997b516c7fe238c7895",
+  "ane-gather/cua_s1_forms_fp16_options32.mlpackage/Data/com.apple.CoreML/model.mlmodel": "de18e313c3b625e35d008ed8b6b24108edc6df7bf2fae9533af03519eca11b63",
+  "ane-gather/cua_s1_forms_fp16_options32.mlpackage/Data/com.apple.CoreML/weights/weight.bin": "4da9259f798e44f5a1b50769ee1916fd3747c4d723dd9997b516c7fe238c7895",
+  "ane-gather/cua_s1_forms_fp16_options32.mlpackage/Manifest.json": "38f812a04eb2322080634ba788c61df362336168466ca67e5549058d346ac793",
+  "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",
+  "demo/browser-demo.gif": "6cf597927e7bcb5617719453eecb86c7c06ab9a91fcaf46de272b42abe12d4fc",
+  "demo/browser-demo.mp4": "5fbb3925c4824d15e6c24c504db91a96d318f4e751a20a070fdedf64e587d1a0",
+  "demo/browser-preview.png": "e624ec2a797d0b3b050e66055d6a9fb09a1f6a9ace67b8eb071ce5f2f9e282b1",
+  "int4-source-fp16/conversion.json": "01105dc8bef7a30fc3430e150caa2a143dc0ba23a295b6280de3b8ff3b40af55",
+  "int4-source-fp16/cua_s1_forms_fp16_options32.mlpackage/Data/com.apple.CoreML/model.mlmodel": "6ab2df320df7081480b76876df0f4bfd2c841592d93f354c568ba504f7269627",
+  "int4-source-fp16/cua_s1_forms_fp16_options32.mlpackage/Data/com.apple.CoreML/weights/weight.bin": "4da9259f798e44f5a1b50769ee1916fd3747c4d723dd9997b516c7fe238c7895",
+  "int4-source-fp16/cua_s1_forms_fp16_options32.mlpackage/Manifest.json": "e5b1c7dc66b2014b3985c180a75548a709788c6e314d9d2ae14e4e6d441c3c68",
+  "int4-weights/conversion.json": "4675febbdef9cc15d46bf8aed17d7c7aaa0e91b3168a525c31b81ad2cfdb6875",
+  "int4-weights/cua_s1_forms_int4_options32.mlpackage/Data/com.apple.CoreML/model.mlmodel": "cadad6bfa6875383a7ee3de68d0d22871148aa21cf117e7fe44c2c971287c4e5",
+  "int4-weights/cua_s1_forms_int4_options32.mlpackage/Data/com.apple.CoreML/weights/weight.bin": "0f766fcb9dbd612ed4f6c7a40117a5993a4789b3ae94afba0df37193c831f63c",
+  "int4-weights/cua_s1_forms_int4_options32.mlpackage/Manifest.json": "fb91c8b5ec3168111567de663bb3140d2790c6e7c48f32e0df3f5135af0673af",
+  "int8-weights/conversion.json": "b8a51ceeb65fe8020b2a186fae2ebf586443956111acfbac1a3f208a062c1c80",
+  "int8-weights/cua_s1_forms_int8_options32.mlpackage/Data/com.apple.CoreML/model.mlmodel": "a7d9747308b8b754a46f3e70c8465b821f90e12376b3a71dc909f7364ae73c3a",
+  "int8-weights/cua_s1_forms_int8_options32.mlpackage/Data/com.apple.CoreML/weights/weight.bin": "f600c4434b7f7e3e789de8f1a4aa54a55f84df51fa9b788e18f5c97b68e4840f",
+  "int8-weights/cua_s1_forms_int8_options32.mlpackage/Manifest.json": "2206ec01787a85c12f1376fa5a832c5ac912f06172c6ff06b26ae8415470d246",
+  "preprocessing.py": "18c74ede43f95f91cd638c7e9631923a353e3a24d3b522ddb8b18b480f9c2846",
+  "reports/ane-comparison.json": "14b29f87704c6e8bd5f9bd866ea825be3972fcea6cf3c4760e38dc8774f937b6",
+  "reports/ane-fallback.json": "2f52b0e11f7cfbea6a6be6942de8816d8fed0630ac384a28551d405648fcb4e0",
+  "reports/ane-gather-fallback.json": "1e9bf8a336ae6d63ff2f07e32fed21b5e5ee4491a8a217b15deb3e1e3ac32988",
+  "reports/ane-gather-profile.json": "023c16760ca124176e399f1c692741f5220bc4cc391cd59bffffa291e2c332a6",
+  "reports/ane-gather-verification.json": "fa51fe0537dcd9cc8d1d75fc0be4747d887aa2e3af199ec08edd8f9400ce3318",
+  "reports/ane-profile.json": "8f654717397a294acc793861676a1480c160bb3f7ab5be6ff67ebb6a4473b832",
+  "reports/browser-validation.json": "299641c7b9f6cc2464cfcd984b4ef4445385e4710bc15abae7cee8ba262331b6",
+  "reports/int4-demo-verification.json": "7465aecddc5c6806a0c1882db3a587eb7ea0f8de7978f60bcbc7e39ffe2cfa32",
+  "reports/int4-fallback.json": "7e14ab1bb6b8c2ede589773af3d39aea1a7349d199a27d597f1cd47a40217b07",
+  "reports/int4-profile.json": "71f6b3ae5c9340a77936e368529201acef0d1e3b987b7436b1d35a829b6e350e",
+  "reports/int4-source-demo-verification.json": "f0c026f462dde1a8f86851d55ec35ab33ab8e68e5af3e4f1eb4b9c52f3dd08b1",
+  "reports/int4-synthetic-test-decisions.jsonl.gz": "6922913041f95ac8872958b92312cb5b869b9ce88116c4684b36ea7753284674",
+  "reports/int4-synthetic-test.json": "9d6a691aeb2ddc352ab7c12327472c92814991c201b67c7da8db39a76c972b32",
+  "reports/int8-demo-verification.json": "5602c8c254c0ae4c696e771291b25d1196ccc0d63f48eaaccdde0245803a5368",
+  "reports/int8-fallback.json": "9bb86a529a8e7e353efa12449a5ea930817dc7b1afdacec6ae24f3c318bc5835",
+  "reports/int8-profile.json": "0b962528dc7e5bb1d53abeb4b549f93b93f549a1e1935c3d77ef2606e83bf6a9",
+  "reports/int8-synthetic-test-decisions.jsonl.gz": "69720b9e11430ecff7ea0718f14bf16792de6f39f114faba250cfe855e4e4156",
+  "reports/int8-synthetic-test.json": "952ebe281d0aef0c05667d3c0dc5f9f175be709639a542028940059393078b5d",
+  "reports/swift-ane-validation.json": "ed323e6d6619fb7698d3546599d6450b745bc8cc6100c02e13f7ad4d59859a3a",
+  "reports/swift-runtime/decisions.jsonl.gz": "5d7ae5df86489b38e8f97167ab58c1d6bf5fb74c171a57b00d818b05d6c772b5",
+  "reports/swift-runtime/report.json": "81c30c4943550356da71ea03b1ff87d6d9a66ec3bbc17c20de5ff802a6a5f817",
+  "reports/swift-variant-comparison.json": "593053d7011a1c6d6881b130c3408099d0fcc09c714c4064d4f853f0dc63b651",
+  "reports/synthetic-test-decisions.jsonl.gz": "b6f6edc9de34fb37b10847a1f9568b530ce2b8e1bd241f51bf2e637f489d6554",
+  "reports/synthetic-test.json": "2ae85cfeb270b4f31be4f1d9082f4c73b3b9b084d6f1ddb661712a87d46b0f59",
+  "reports/upstream-metrics.json": "7d6207e430504dc0baec96d8180cb03f8aadc5a6e016b53c4cf03a80baf169ab",
+  "reports/verification.json": "9a5d22ec4b447f0710c12e03c69d7ed802158bb94e9a3c6b059731730ced9f30",
+  "synthetic-test.lock.json": "e582b951ecc07809a0b06a2bc0790827b90a86ce57c70d5b6177ef04f0b4f1ca"
+}

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"
+}

demo/browser-demo.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5fbb3925c4824d15e6c24c504db91a96d318f4e751a20a070fdedf64e587d1a0
+size 1108747

int4-source-fp16/conversion.json

@@ -0,0 +1,73 @@
+{
+  "model": "cua_s1_forms_fp16_options32.mlpackage",
+  "precision": "float16",
+  "minimum_target": "iOS18/macOS15",
+  "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": "6ab2df320df7081480b76876df0f4bfd2c841592d93f354c568ba504f7269627",
+    "Data/com.apple.CoreML/weights/weight.bin": "4da9259f798e44f5a1b50769ee1916fd3747c4d723dd9997b516c7fe238c7895",
+    "Manifest.json": "e5b1c7dc66b2014b3985c180a75548a709788c6e314d9d2ae14e4e6d441c3c68"
+  },
+  "target_upgrade": {
+    "source_conversion_sha256": "ac94239751ceb7e19602eb7276940ff53446b1776d4900405af9be0382fe4a94",
+    "source_package_files": {
+      "Data/com.apple.CoreML/model.mlmodel": "70485fc18cbb21785df833cbddddc0b5b59acb00d22394b76e55307e2c135dd0",
+      "Data/com.apple.CoreML/weights/weight.bin": "4da9259f798e44f5a1b50769ee1916fd3747c4d723dd9997b516c7fe238c7895",
+      "Manifest.json": "2bc0f5f62337b27fb6b0ecde248f1e3dc269e1ba4b65516aaeede2a60e293dcc"
+    },
+    "method": "Load original MIL at specification 9; empty pass pipeline; retain decomposed attention",
+    "operation_counts": {
+      "add": 13,
+      "cast": 11,
+      "concat": 2,
+      "const": 239,
+      "equal": 2,
+      "expand_dims": 5,
+      "fill_like": 3,
+      "gather": 2,
+      "greater_equal": 2,
+      "layer_norm": 8,
+      "linear": 15,
+      "logical_not": 2,
+      "matmul": 8,
+      "maximum": 1,
+      "mul": 7,
+      "not_equal": 3,
+      "real_div": 1,
+      "reduce_sum": 3,
+      "relu": 3,
+      "reshape": 31,
+      "select": 7,
+      "slice_by_index": 13,
+      "softmax": 5,
+      "squeeze": 3,
+      "tile": 2,
+      "transpose": 21
+    },
+    "script_sha256": "33db3295450428d90c1de5a804ab47bf3507c686fbbdef67cd276ce116758b5f"
+  }
+}

int4-source-fp16/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 sha256:6ab2df320df7081480b76876df0f4bfd2c841592d93f354c568ba504f7269627
+size 63743

int4-source-fp16/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

int4-source-fp16/cua_s1_forms_fp16_options32.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "15C1AFE1-37B2-4975-949D-B817E3C00088": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        },
+        "B997D799-579D-437C-9F80-DE24FF3EFB11": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        }
+    },
+    "rootModelIdentifier": "15C1AFE1-37B2-4975-949D-B817E3C00088"
+}

int4-weights/conversion.json

@@ -0,0 +1,124 @@
+{
+  "model": "cua_s1_forms_int4_options32.mlpackage",
+  "precision": "int4_weights_float16_compute",
+  "minimum_target": "iOS18/macOS15",
+  "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": "cadad6bfa6875383a7ee3de68d0d22871148aa21cf117e7fe44c2c971287c4e5",
+    "Data/com.apple.CoreML/weights/weight.bin": "0f766fcb9dbd612ed4f6c7a40117a5993a4789b3ae94afba0df37193c831f63c",
+    "Manifest.json": "fb91c8b5ec3168111567de663bb3140d2790c6e7c48f32e0df3f5135af0673af"
+  },
+  "target_upgrade": {
+    "source_conversion_sha256": "ac94239751ceb7e19602eb7276940ff53446b1776d4900405af9be0382fe4a94",
+    "source_package_files": {
+      "Data/com.apple.CoreML/model.mlmodel": "70485fc18cbb21785df833cbddddc0b5b59acb00d22394b76e55307e2c135dd0",
+      "Data/com.apple.CoreML/weights/weight.bin": "4da9259f798e44f5a1b50769ee1916fd3747c4d723dd9997b516c7fe238c7895",
+      "Manifest.json": "2bc0f5f62337b27fb6b0ecde248f1e3dc269e1ba4b65516aaeede2a60e293dcc"
+    },
+    "method": "Load original MIL at specification 9; empty pass pipeline; retain decomposed attention",
+    "operation_counts": {
+      "add": 13,
+      "cast": 11,
+      "concat": 2,
+      "const": 239,
+      "equal": 2,
+      "expand_dims": 5,
+      "fill_like": 3,
+      "gather": 2,
+      "greater_equal": 2,
+      "layer_norm": 8,
+      "linear": 15,
+      "logical_not": 2,
+      "matmul": 8,
+      "maximum": 1,
+      "mul": 7,
+      "not_equal": 3,
+      "real_div": 1,
+      "reduce_sum": 3,
+      "relu": 3,
+      "reshape": 31,
+      "select": 7,
+      "slice_by_index": 13,
+      "softmax": 5,
+      "squeeze": 3,
+      "tile": 2,
+      "transpose": 21
+    },
+    "script_sha256": "33db3295450428d90c1de5a804ab47bf3507c686fbbdef67cd276ce116758b5f"
+  },
+  "quantization": {
+    "name": "int4-weights",
+    "settings": {
+      "mode": "linear_symmetric",
+      "dtype": "int4",
+      "granularity": "per_channel",
+      "weight_threshold": 2048
+    },
+    "activations": "float16",
+    "calibration": "None; weight-only, data-free quantization",
+    "source_conversion_sha256": "01105dc8bef7a30fc3430e150caa2a143dc0ba23a295b6280de3b8ff3b40af55",
+    "source_package_files": {
+      "Data/com.apple.CoreML/model.mlmodel": "6ab2df320df7081480b76876df0f4bfd2c841592d93f354c568ba504f7269627",
+      "Data/com.apple.CoreML/weights/weight.bin": "4da9259f798e44f5a1b50769ee1916fd3747c4d723dd9997b516c7fe238c7895",
+      "Manifest.json": "e5b1c7dc66b2014b3985c180a75548a709788c6e314d9d2ae14e4e6d441c3c68"
+    },
+    "source_package_bytes": 1511080,
+    "package_bytes": 451595,
+    "compression_seconds": 0.2996572910051327,
+    "packed_int4_tensors": 19,
+    "operation_counts": {
+      "add": 15,
+      "cast": 13,
+      "concat": 2,
+      "const": 226,
+      "constexpr_blockwise_shift_scale": 19,
+      "equal": 2,
+      "expand_dims": 5,
+      "fill_like": 3,
+      "gather": 2,
+      "greater_equal": 4,
+      "layer_norm": 8,
+      "linear": 15,
+      "logical_not": 2,
+      "matmul": 8,
+      "maximum": 1,
+      "mul": 7,
+      "not_equal": 3,
+      "real_div": 1,
+      "reduce_sum": 3,
+      "relu": 3,
+      "reshape": 31,
+      "select": 9,
+      "slice_by_index": 13,
+      "softmax": 5,
+      "squeeze": 3,
+      "tile": 2,
+      "transpose": 21
+    },
+    "script_sha256": "a91636e539d8934253b5e7628ca706346fbee538fa1aa9e3eaf794f440d49dcc"
+  }
+}

int4-weights/cua_s1_forms_int4_options32.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cadad6bfa6875383a7ee3de68d0d22871148aa21cf117e7fe44c2c971287c4e5
+size 67938

int4-weights/cua_s1_forms_int4_options32.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0f766fcb9dbd612ed4f6c7a40117a5993a4789b3ae94afba0df37193c831f63c
+size 383040

int4-weights/cua_s1_forms_int4_options32.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "D6D01355-3106-40A1-999B-B56148C73691": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "F0E58006-D8D0-4891-AD75-BB9E48868A99": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "F0E58006-D8D0-4891-AD75-BB9E48868A99"
+}

int8-weights/conversion.json

@@ -0,0 +1,85 @@
+{
+  "model": "cua_s1_forms_int8_options32.mlpackage",
+  "precision": "int8_weights_float16_compute",
+  "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": "a7d9747308b8b754a46f3e70c8465b821f90e12376b3a71dc909f7364ae73c3a",
+    "Data/com.apple.CoreML/weights/weight.bin": "f600c4434b7f7e3e789de8f1a4aa54a55f84df51fa9b788e18f5c97b68e4840f",
+    "Manifest.json": "2206ec01787a85c12f1376fa5a832c5ac912f06172c6ff06b26ae8415470d246"
+  },
+  "quantization": {
+    "name": "int8-weights",
+    "settings": {
+      "mode": "linear_symmetric",
+      "dtype": "int8",
+      "granularity": "per_channel",
+      "weight_threshold": 2048
+    },
+    "activations": "float16",
+    "calibration": "None; weight-only, data-free quantization",
+    "source_conversion_sha256": "ac94239751ceb7e19602eb7276940ff53446b1776d4900405af9be0382fe4a94",
+    "source_package_files": {
+      "Data/com.apple.CoreML/model.mlmodel": "70485fc18cbb21785df833cbddddc0b5b59acb00d22394b76e55307e2c135dd0",
+      "Data/com.apple.CoreML/weights/weight.bin": "4da9259f798e44f5a1b50769ee1916fd3747c4d723dd9997b516c7fe238c7895",
+      "Manifest.json": "2bc0f5f62337b27fb6b0ecde248f1e3dc269e1ba4b65516aaeede2a60e293dcc"
+    },
+    "source_package_bytes": 1511163,
+    "package_bytes": 813422,
+    "compression_seconds": 0.3180606249952689,
+    "operation_counts": {
+      "add": 15,
+      "cast": 13,
+      "concat": 2,
+      "const": 226,
+      "constexpr_affine_dequantize": 19,
+      "equal": 2,
+      "expand_dims": 5,
+      "fill_like": 3,
+      "gather": 2,
+      "greater_equal": 4,
+      "layer_norm": 8,
+      "linear": 15,
+      "logical_not": 2,
+      "matmul": 8,
+      "maximum": 1,
+      "mul": 7,
+      "not_equal": 3,
+      "real_div": 1,
+      "reduce_sum": 3,
+      "relu": 3,
+      "reshape": 31,
+      "select": 9,
+      "slice_by_index": 13,
+      "softmax": 5,
+      "squeeze": 3,
+      "tile": 2,
+      "transpose": 21
+    },
+    "script_sha256": "22071babfcb02ef77225203a479d9bb14b9fa35462ba38bcc4bf07d29c125a97"
+  }
+}

int8-weights/cua_s1_forms_int8_options32.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a7d9747308b8b754a46f3e70c8465b821f90e12376b3a71dc909f7364ae73c3a
+size 69573

int8-weights/cua_s1_forms_int8_options32.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f600c4434b7f7e3e789de8f1a4aa54a55f84df51fa9b788e18f5c97b68e4840f
+size 743232

int8-weights/cua_s1_forms_int8_options32.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "6D5496F1-54BB-4AAA-8375-1DFA1FFE6768": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        },
+        "B0C97583-069E-42B8-897B-30A05F77B14A": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        }
+    },
+    "rootModelIdentifier": "6D5496F1-54BB-4AAA-8375-1DFA1FFE6768"
+}

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-comparison.json

@@ -0,0 +1,784 @@
+{
+  "created_utc": "2026-09-19T20:07:16.396741+00:00",
+  "purpose": "Matched warm model-call comparison, not a power or utilization measurement",
+  "environment": {
+    "chip": "Apple M5 Pro",
+    "macos": "27.0",
+    "os_build": "26A428",
+    "python": "3.11.11",
+    "coremltools": "9.0"
+  },
+  "protocol": {
+    "compute_units": "CPU_AND_NE",
+    "rows": [
+      0,
+      68,
+      130
+    ],
+    "schedule": [
+      "baseline",
+      "candidate",
+      "candidate",
+      "baseline"
+    ],
+    "warmup_passes_per_block": 2,
+    "timed_passes_per_block": 10,
+    "timing_scope": "Synchronous Python predict with pre-encoded inputs; validation outside timer",
+    "model_loading": "Both models resident in one process; loads excluded and system caches retained"
+  },
+  "dataset_revision": "8273f34778b99ac2e12d9f6e7d57dad99ae20845",
+  "dataset_sha256": "4f43b442e79ba2e2ce731e27e9b8e340c2b5dfcaffc92d8ff564c34f115ff1ca",
+  "package_files": {
+    "baseline": {
+      "Data/com.apple.CoreML/model.mlmodel": "70485fc18cbb21785df833cbddddc0b5b59acb00d22394b76e55307e2c135dd0",
+      "Data/com.apple.CoreML/weights/weight.bin": "4da9259f798e44f5a1b50769ee1916fd3747c4d723dd9997b516c7fe238c7895",
+      "Manifest.json": "2bc0f5f62337b27fb6b0ecde248f1e3dc269e1ba4b65516aaeede2a60e293dcc"
+    },
+    "candidate": {
+      "Data/com.apple.CoreML/model.mlmodel": "de18e313c3b625e35d008ed8b6b24108edc6df7bf2fae9533af03519eca11b63",
+      "Data/com.apple.CoreML/weights/weight.bin": "4da9259f798e44f5a1b50769ee1916fd3747c4d723dd9997b516c7fe238c7895",
+      "Manifest.json": "38f812a04eb2322080634ba788c61df362336168466ca67e5549058d346ac793"
+    }
+  },
+  "summary": {
+    "baseline": {
+      "count": 60,
+      "median_ms": 0.9150624828180298,
+      "p95_ms": 0.9682062489446251,
+      "min_ms": 0.88170898379758,
+      "max_ms": 1.0281660070177168
+    },
+    "candidate": {
+      "count": 60,
+      "median_ms": 0.9703329997137189,
+      "p95_ms": 0.9877315780613571,
+      "min_ms": 0.9434169915039092,
+      "max_ms": 1.0195420181844383
+    }
+  },
+  "candidate_over_baseline_median": 1.0604008118937165,
+  "correct_timed_predictions": 120,
+  "calls": [
+    {
+      "model": "baseline",
+      "block": 0,
+      "row": 0,
+      "milliseconds": 0.919707992579788
+    },
+    {
+      "model": "baseline",
+      "block": 0,
+      "row": 68,
+      "milliseconds": 0.9307919826824218
+    },
+    {
+      "model": "baseline",
+      "block": 0,
+      "row": 130,
+      "milliseconds": 0.9576249867677689
+    },
+    {
+      "model": "baseline",
+      "block": 0,
+      "row": 0,
+      "milliseconds": 0.9234579920303077
+    },
+    {
+      "model": "baseline",
+      "block": 0,
+      "row": 68,
+      "milliseconds": 0.9247499983757734
+    },
+    {
+      "model": "baseline",
+      "block": 0,
+      "row": 130,
+      "milliseconds": 0.9314999915659428
+    },
+    {
+      "model": "baseline",
+      "block": 0,
+      "row": 0,
+      "milliseconds": 0.9133330022450536
+    },
+    {
+      "model": "baseline",
+      "block": 0,
+      "row": 68,
+      "milliseconds": 0.9190000128000975
+    },
+    {
+      "model": "baseline",
+      "block": 0,
+      "row": 130,
+      "milliseconds": 0.9215839963871986
+    },
+    {
+      "model": "baseline",
+      "block": 0,
+      "row": 0,
+      "milliseconds": 0.9041670127771795
+    },
+    {
+      "model": "baseline",
+      "block": 0,
+      "row": 68,
+      "milliseconds": 0.9001670114230365
+    },
+    {
+      "model": "baseline",
+      "block": 0,
+      "row": 130,
+      "milliseconds": 0.9362080018036067
+    },
+    {
+      "model": "baseline",
+      "block": 0,
+      "row": 0,
+      "milliseconds": 1.0092500015161932
+    },
+    {
+      "model": "baseline",
+      "block": 0,
+      "row": 68,
+      "milliseconds": 0.9958750160876662
+    },
+    {
+      "model": "baseline",
+      "block": 0,
+      "row": 130,
+      "milliseconds": 1.0281660070177168
+    },
+    {
+      "model": "baseline",
+      "block": 0,
+      "row": 0,
+      "milliseconds": 0.96674999804236
+    },
+    {
+      "model": "baseline",
+      "block": 0,
+      "row": 68,
+      "milliseconds": 0.9511670214124024
+    },
+    {
+      "model": "baseline",
+      "block": 0,
+      "row": 130,
+      "milliseconds": 0.9152919810730964
+    },
+    {
+      "model": "baseline",
+      "block": 0,
+      "row": 0,
+      "milliseconds": 0.9129999962169677
+    },
+    {
+      "model": "baseline",
+      "block": 0,
+      "row": 68,
+      "milliseconds": 0.9089170198421925
+    },
+    {
+      "model": "baseline",
+      "block": 0,
+      "row": 130,
+      "milliseconds": 0.9163749928120524
+    },
+    {
+      "model": "baseline",
+      "block": 0,
+      "row": 0,
+      "milliseconds": 0.9026250045280904
+    },
+    {
+      "model": "baseline",
+      "block": 0,
+      "row": 68,
+      "milliseconds": 0.9094580018427223
+    },
+    {
+      "model": "baseline",
+      "block": 0,
+      "row": 130,
+      "milliseconds": 0.9331250039394945
+    },
+    {
+      "model": "baseline",
+      "block": 0,
+      "row": 0,
+      "milliseconds": 0.9327089937869459
+    },
+    {
+      "model": "baseline",
+      "block": 0,
+      "row": 68,
+      "milliseconds": 0.9123749914579093
+    },
+    {
+      "model": "baseline",
+      "block": 0,
+      "row": 130,
+      "milliseconds": 0.9423340088687837
+    },
+    {
+      "model": "baseline",
+      "block": 0,
+      "row": 0,
+      "milliseconds": 0.9080420131795108
+    },
+    {
+      "model": "baseline",
+      "block": 0,
+      "row": 68,
+      "milliseconds": 0.9152499842457473
+    },
+    {
+      "model": "baseline",
+      "block": 0,
+      "row": 130,
+      "milliseconds": 0.9090829989872873
+    },
+    {
+      "model": "candidate",
+      "block": 1,
+      "row": 0,
+      "milliseconds": 0.9569999820087105
+    },
+    {
+      "model": "candidate",
+      "block": 1,
+      "row": 68,
+      "milliseconds": 0.9605420054867864
+    },
+    {
+      "model": "candidate",
+      "block": 1,
+      "row": 130,
+      "milliseconds": 0.971125002251938
+    },
+    {
+      "model": "candidate",
+      "block": 1,
+      "row": 0,
+      "milliseconds": 0.9833329822868109
+    },
+    {
+      "model": "candidate",
+      "block": 1,
+      "row": 68,
+      "milliseconds": 0.9731249883770943
+    },
+    {
+      "model": "candidate",
+      "block": 1,
+      "row": 130,
+      "milliseconds": 0.9744580020196736
+    },
+    {
+      "model": "candidate",
+      "block": 1,
+      "row": 0,
+      "milliseconds": 0.9531670075375587
+    },
+    {
+      "model": "candidate",
+      "block": 1,
+      "row": 68,
+      "milliseconds": 0.960625009611249
+    },
+    {
+      "model": "candidate",
+      "block": 1,
+      "row": 130,
+      "milliseconds": 0.9434169915039092
+    },
+    {
+      "model": "candidate",
+      "block": 1,
+      "row": 0,
+      "milliseconds": 0.965375016676262
+    },
+    {
+      "model": "candidate",
+      "block": 1,
+      "row": 68,
+      "milliseconds": 0.9999169851653278
+    },
+    {
+      "model": "candidate",
+      "block": 1,
+      "row": 130,
+      "milliseconds": 0.9775409998837858
+    },
+    {
+      "model": "candidate",
+      "block": 1,
+      "row": 0,
+      "milliseconds": 0.983749981969595
+    },
+    {
+      "model": "candidate",
+      "block": 1,
+      "row": 68,
+      "milliseconds": 0.9759170061443001
+    },
+    {
+      "model": "candidate",
+      "block": 1,
+      "row": 130,
+      "milliseconds": 0.979083008132875
+    },
+    {
+      "model": "candidate",
+      "block": 1,
+      "row": 0,
+      "milliseconds": 0.9849590132944286
+    },
+    {
+      "model": "candidate",
+      "block": 1,
+      "row": 68,
+      "milliseconds": 0.9607080137357116
+    },
+    {
+      "model": "candidate",
+      "block": 1,
+      "row": 130,
+      "milliseconds": 0.9490000084042549
+    },
+    {
+      "model": "candidate",
+      "block": 1,
+      "row": 0,
+      "milliseconds": 0.9506659989710897
+    },
+    {
+      "model": "candidate",
+      "block": 1,
+      "row": 68,
+      "milliseconds": 0.9708750003483146
+    },
+    {
+      "model": "candidate",
+      "block": 1,
+      "row": 130,
+      "milliseconds": 0.9733340120874345
+    },
+    {
+      "model": "candidate",
+      "block": 1,
+      "row": 0,
+      "milliseconds": 0.9872919763438404
+    },
+    {
+      "model": "candidate",
+      "block": 1,
+      "row": 68,
+      "milliseconds": 0.970749999396503
+    },
+    {
+      "model": "candidate",
+      "block": 1,
+      "row": 130,
+      "milliseconds": 0.9687079873401672
+    },
+    {
+      "model": "candidate",
+      "block": 1,
+      "row": 0,
+      "milliseconds": 0.977249990683049
+    },
+    {
+      "model": "candidate",
+      "block": 1,
+      "row": 68,
+      "milliseconds": 0.9785419970285147
+    },
+    {
+      "model": "candidate",
+      "block": 1,
+      "row": 130,
+      "milliseconds": 0.9744160051923245
+    },
+    {
+      "model": "candidate",
+      "block": 1,
+      "row": 0,
+      "milliseconds": 0.9864169987849891
+    },
+    {
+      "model": "candidate",
+      "block": 1,
+      "row": 68,
+      "milliseconds": 0.982333003776148
+    },
+    {
+      "model": "candidate",
+      "block": 1,
+      "row": 130,
+      "milliseconds": 0.9689999860711396
+    },
+    {
+      "model": "candidate",
+      "block": 2,
+      "row": 0,
+      "milliseconds": 0.9796250087674707
+    },
+    {
+      "model": "candidate",
+      "block": 2,
+      "row": 68,
+      "milliseconds": 0.9837090037763119
+    },
+    {
+      "model": "candidate",
+      "block": 2,
+      "row": 130,
+      "milliseconds": 0.9523750049993396
+    },
+    {
+      "model": "candidate",
+      "block": 2,
+      "row": 0,
+      "milliseconds": 0.9744999988470227
+    },
+    {
+      "model": "candidate",
+      "block": 2,
+      "row": 68,
+      "milliseconds": 0.9580420155543834
+    },
+    {
+      "model": "candidate",
+      "block": 2,
+      "row": 130,
+      "milliseconds": 0.9489580115769058
+    },
+    {
+      "model": "candidate",
+      "block": 2,
+      "row": 0,
+      "milliseconds": 0.9510410018265247
+    },
+    {
+      "model": "candidate",
+      "block": 2,
+      "row": 68,
+      "milliseconds": 0.9449580102227628
+    },
+    {
+      "model": "candidate",
+      "block": 2,
+      "row": 130,
+      "milliseconds": 0.9497080172877759
+    },
+    {
+      "model": "candidate",
+      "block": 2,
+      "row": 0,
+      "milliseconds": 0.9529579838272184
+    },
+    {
+      "model": "candidate",
+      "block": 2,
+      "row": 68,
+      "milliseconds": 0.9495420090388507
+    },
+    {
+      "model": "candidate",
+      "block": 2,
+      "row": 130,
+      "milliseconds": 0.9463340102229267
+    },
+    {
+      "model": "candidate",
+      "block": 2,
+      "row": 0,
+      "milliseconds": 0.9437499975319952
+    },
+    {
+      "model": "candidate",
+      "block": 2,
+      "row": 68,
+      "milliseconds": 0.9775829967111349
+    },
+    {
+      "model": "candidate",
+      "block": 2,
+      "row": 130,
+      "milliseconds": 0.9694590116851032
+    },
+    {
+      "model": "candidate",
+      "block": 2,
+      "row": 0,
+      "milliseconds": 0.996084010694176
+    },
+    {
+      "model": "candidate",
+      "block": 2,
+      "row": 68,
+      "milliseconds": 1.0195420181844383
+    },
+    {
+      "model": "candidate",
+      "block": 2,
+      "row": 130,
+      "milliseconds": 0.9871670044958591
+    },
+    {
+      "model": "candidate",
+      "block": 2,
+      "row": 0,
+      "milliseconds": 0.9699160000309348
+    },
+    {
+      "model": "candidate",
+      "block": 2,
+      "row": 68,
+      "milliseconds": 0.9592499991413206
+    },
+    {
+      "model": "candidate",
+      "block": 2,
+      "row": 130,
+      "milliseconds": 0.9731660247780383
+    },
+    {
+      "model": "candidate",
+      "block": 2,
+      "row": 0,
+      "milliseconds": 0.9729999874252826
+    },
+    {
+      "model": "candidate",
+      "block": 2,
+      "row": 68,
+      "milliseconds": 0.9556250006426126
+    },
+    {
+      "model": "candidate",
+      "block": 2,
+      "row": 130,
+      "milliseconds": 0.9530829847790301
+    },
+    {
+      "model": "candidate",
+      "block": 2,
+      "row": 0,
+      "milliseconds": 0.9495830163359642
+    },
+    {
+      "model": "candidate",
+      "block": 2,
+      "row": 68,
+      "milliseconds": 0.9697090135887265
+    },
+    {
+      "model": "candidate",
+      "block": 2,
+      "row": 130,
+      "milliseconds": 0.9829580085352063
+    },
+    {
+      "model": "candidate",
+      "block": 2,
+      "row": 0,
+      "milliseconds": 0.9604159859009087
+    },
+    {
+      "model": "candidate",
+      "block": 2,
+      "row": 68,
+      "milliseconds": 0.9743330010678619
+    },
+    {
+      "model": "candidate",
+      "block": 2,
+      "row": 130,
+      "milliseconds": 0.9570419788360596
+    },
+    {
+      "model": "baseline",
+      "block": 3,
+      "row": 0,
+      "milliseconds": 0.927665998460725
+    },
+    {
+      "model": "baseline",
+      "block": 3,
+      "row": 68,
+      "milliseconds": 0.9297500073444098
+    },
+    {
+      "model": "baseline",
+      "block": 3,
+      "row": 130,
+      "milliseconds": 0.9148749813903123
+    },
+    {
+      "model": "baseline",
+      "block": 3,
+      "row": 0,
+      "milliseconds": 0.8912909834180027
+    },
+    {
+      "model": "baseline",
+      "block": 3,
+      "row": 68,
+      "milliseconds": 0.88170898379758
+    },
+    {
+      "model": "baseline",
+      "block": 3,
+      "row": 130,
+      "milliseconds": 0.9040829900186509
+    },
+    {
+      "model": "baseline",
+      "block": 3,
+      "row": 0,
+      "milliseconds": 0.8952920034062117
+    },
+    {
+      "model": "baseline",
+      "block": 3,
+      "row": 68,
+      "milliseconds": 0.9061660093721002
+    },
+    {
+      "model": "baseline",
+      "block": 3,
+      "row": 130,
+      "milliseconds": 0.9120829927269369
+    },
+    {
+      "model": "baseline",
+      "block": 3,
+      "row": 0,
+      "milliseconds": 0.9074159897863865
+    },
+    {
+      "model": "baseline",
+      "block": 3,
+      "row": 68,
+      "milliseconds": 0.9095839923247695
+    },
+    {
+      "model": "baseline",
+      "block": 3,
+      "row": 130,
+      "milliseconds": 0.883542001247406
+    },
+    {
+      "model": "baseline",
+      "block": 3,
+      "row": 0,
+      "milliseconds": 0.8904590213205665
+    },
+    {
+      "model": "baseline",
+      "block": 3,
+      "row": 68,
+      "milliseconds": 0.8850830199662596
+    },
+    {
+      "model": "baseline",
+      "block": 3,
+      "row": 130,
+      "milliseconds": 0.8904580026865005
+    },
+    {
+      "model": "baseline",
+      "block": 3,
+      "row": 0,
+      "milliseconds": 0.9044160251505673
+    },
+    {
+      "model": "baseline",
+      "block": 3,
+      "row": 68,
+      "milliseconds": 0.9244170214515179
+    },
+    {
+      "model": "baseline",
+      "block": 3,
+      "row": 130,
+      "milliseconds": 0.9261659870389849
+    },
+    {
+      "model": "baseline",
+      "block": 3,
+      "row": 0,
+      "milliseconds": 0.9326669969595969
+    },
+    {
+      "model": "baseline",
+      "block": 3,
+      "row": 68,
+      "milliseconds": 0.9306249849032611
+    },
+    {
+      "model": "baseline",
+      "block": 3,
+      "row": 130,
+      "milliseconds": 0.8886670111678541
+    },
+    {
+      "model": "baseline",
+      "block": 3,
+      "row": 0,
+      "milliseconds": 0.8824579999782145
+    },
+    {
+      "model": "baseline",
+      "block": 3,
+      "row": 68,
+      "milliseconds": 0.8989160123746842
+    },
+    {
+      "model": "baseline",
+      "block": 3,
+      "row": 130,
+      "milliseconds": 0.8977499965112656
+    },
+    {
+      "model": "baseline",
+      "block": 3,
+      "row": 0,
+      "milliseconds": 0.8965420129243284
+    },
+    {
+      "model": "baseline",
+      "block": 3,
+      "row": 68,
+      "milliseconds": 0.9054580004885793
+    },
+    {
+      "model": "baseline",
+      "block": 3,
+      "row": 130,
+      "milliseconds": 0.93029101844877
+    },
+    {
+      "model": "baseline",
+      "block": 3,
+      "row": 0,
+      "milliseconds": 0.9197920153383166
+    },
+    {
+      "model": "baseline",
+      "block": 3,
+      "row": 68,
+      "milliseconds": 0.9319579985458404
+    },
+    {
+      "model": "baseline",
+      "block": 3,
+      "row": 130,
+      "milliseconds": 0.9407080069649965
+    }
+  ],
+  "script_sha256": "4633df7ad5f90d6aed37d5ef4f7a2d2231fdadd272ef1aa3057c237b2f937df8"
+}

reports/ane-fallback.json

@@ -0,0 +1,84 @@
+{
+  "hardware": {
+    "device": "arm64",
+    "chip": "Apple M5 Pro",
+    "ram": "24GB",
+    "os_version": "macOS 27.0",
+    "timestamp": "2026-09-19T19:41:20.560977+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/ane-gather-fallback.json

@@ -0,0 +1,38 @@
+{
+  "hardware": {
+    "device": "arm64",
+    "chip": "Apple M5 Pro",
+    "ram": "24GB",
+    "os_version": "macOS 27.0",
+    "timestamp": "2026-09-19T20:08:52.988799+00:00"
+  },
+  "models": [
+    {
+      "model_path": "build/ane-gather/cua_s1_forms_fp16_options32.mlmodelc",
+      "model_name": "cua_s1_forms_fp16_options32",
+      "fallback": {
+        "compute_units": "cpu_and_neural_engine",
+        "total_ops": 165,
+        "ane_ops": 162,
+        "gpu_ops": 0,
+        "cpu_ops": 3,
+        "ane_percent": 98.2,
+        "reasons": [
+          {
+            "reason": "Unsupported tensor data type: int32",
+            "count": 3,
+            "estimated_cpu_runtime_ms": 0.003,
+            "op_types": {
+              "ios17.cast": 3
+            },
+            "ops": [
+              "option_mask_to_fp16",
+              "context_ids_to_fp16",
+              "option_ids_to_fp16"
+            ]
+          }
+        ]
+      }
+    }
+  ]
+}

reports/ane-gather-verification.json

@@ -0,0 +1,3314 @@
+{
+  "purpose": "Local conversion parity on the upstream demo; not a generalization or live GUI benchmark",
+  "created_utc": "2026-09-19T20:05:25.654377+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",
+    "optimization": "ane-gather",
+    "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": 0.6771400420111604,
+    "python": "3.11.11",
+    "torch": "2.7.0",
+    "coremltools": "9.0",
+    "package_files": {
+      "Data/com.apple.CoreML/model.mlmodel": "de18e313c3b625e35d008ed8b6b24108edc6df7bf2fae9533af03519eca11b63",
+      "Data/com.apple.CoreML/weights/weight.bin": "4da9259f798e44f5a1b50769ee1916fd3747c4d723dd9997b516c7fe238c7895",
+      "Manifest.json": "38f812a04eb2322080634ba788c61df362336168466ca67e5549058d346ac793"
+    }
+  },
+  "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": 2.46754199906718,
+      "p95_ms": 3.106593489064835,
+      "min_ms": 1.7850839940365404,
+      "max_ms": 3.8927909918129444
+    }
+  },
+  "export_adapter_fp32": {
+    "argmax_agreement": 196,
+    "max_abs_probability_error": 1.1324882507324219e-06
+  },
+  "host_preprocessing_latency_ms": {
+    "count": 196,
+    "median_ms": 0.03879200085066259,
+    "p95_ms": 0.050938004278577864,
+    "min_ms": 0.026541994884610176,
+    "max_ms": 0.06945800851099193
+  },
+  "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": 0.00233614444732666,
+      "mean_row_max_abs_probability_error": 2.934475840582507e-05,
+      "max_abs_logit_error": 0.07230281829833984,
+      "load_seconds": 0.6175256659917068,
+      "warm_inference_latency_ms": {
+        "count": 196,
+        "median_ms": 0.9540419996483251,
+        "p95_ms": 1.0018329994636588,
+        "min_ms": 0.9285419946536422,
+        "max_ms": 1.5723749820608646
+      },
+      "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
+        }
+      ]
+    },
+    "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": 0.578353041986702,
+      "warm_inference_latency_ms": {
+        "count": 196,
+        "median_ms": 0.9623960067983717,
+        "p95_ms": 0.9909170039463788,
+        "min_ms": 0.9270830196328461,
+        "max_ms": 1.0197919909842312
+      },
+      "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": {
+    "ane_gather.py": "c3379a3797041dcfed82eaaea3990771ec91f1dff7ec146b025988ad7a8c9aae",
+    "assets.py": "f7a5bc931075237e00c73bf21fee3575be35531f934f23b5e22627d50a79f92f",
+    "convert-coreml.py": "46fe0194df924033126286e601214b6789f5d667aae38c153cf4dbc1b0e629b0",
+    "export-reference.py": "ff9eab4d6e4e72ffaf074d82610f41fbea1e802da7c0e987ce562a89eb561b99",
+    "export_model.py": "c61da5d678c77030d1e9a78359626063ea31da46fb6434e49517fdeaa7800949",
+    "preprocessing.py": "18c74ede43f95f91cd638c7e9631923a353e3a24d3b522ddb8b18b480f9c2846",
+    "profile-coreml.py": "d5e1386f54fc6c482647120e593e7ffbbe4bbb277c2cc6d4949154660fe239fd",
+    "score-report.py": "c78e45caa86fb59f9a56ef18cc6a35020ea6e281a6a23d23618f56da85a0d95e",
+    "verify.py": "bc20ec9df47708e9b48e411076abe4eab7c18313f67e11334de2d9e3ecb3efa7"
+  }
+}

reports/int4-demo-verification.json

@@ -0,0 +1,1799 @@
+{
+  "purpose": "Local conversion parity on the upstream demo; not a generalization or live GUI benchmark",
+  "created_utc": "2026-09-19T22:25:44.244422+00:00",
+  "model_revision": "f54adbf447f4ca6ec259f529ee3f2e3e09f8cc71",
+  "source_revision": "83f142c4290a0f7d9ed545ae8532858c6e4f8145",
+  "dataset_revision": "8273f34778b99ac2e12d9f6e7d57dad99ae20845",
+  "dataset_file": "demo.jsonl",
+  "dataset_sha256": "4f43b442e79ba2e2ce731e27e9b8e340c2b5dfcaffc92d8ff564c34f115ff1ca",
+  "conversion": {
+    "model": "cua_s1_forms_int4_options32.mlpackage",
+    "precision": "int4_weights_float16_compute",
+    "minimum_target": "iOS18/macOS15",
+    "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": "cadad6bfa6875383a7ee3de68d0d22871148aa21cf117e7fe44c2c971287c4e5",
+      "Data/com.apple.CoreML/weights/weight.bin": "0f766fcb9dbd612ed4f6c7a40117a5993a4789b3ae94afba0df37193c831f63c",
+      "Manifest.json": "fb91c8b5ec3168111567de663bb3140d2790c6e7c48f32e0df3f5135af0673af"
+    },
+    "target_upgrade": {
+      "source_conversion_sha256": "ac94239751ceb7e19602eb7276940ff53446b1776d4900405af9be0382fe4a94",
+      "source_package_files": {
+        "Data/com.apple.CoreML/model.mlmodel": "70485fc18cbb21785df833cbddddc0b5b59acb00d22394b76e55307e2c135dd0",
+        "Data/com.apple.CoreML/weights/weight.bin": "4da9259f798e44f5a1b50769ee1916fd3747c4d723dd9997b516c7fe238c7895",
+        "Manifest.json": "2bc0f5f62337b27fb6b0ecde248f1e3dc269e1ba4b65516aaeede2a60e293dcc"
+      },
+      "method": "Load original MIL at specification 9; empty pass pipeline; retain decomposed attention",
+      "operation_counts": {
+        "add": 13,
+        "cast": 11,
+        "concat": 2,
+        "const": 239,
+        "equal": 2,
+        "expand_dims": 5,
+        "fill_like": 3,
+        "gather": 2,
+        "greater_equal": 2,
+        "layer_norm": 8,
+        "linear": 15,
+        "logical_not": 2,
+        "matmul": 8,
+        "maximum": 1,
+        "mul": 7,
+        "not_equal": 3,
+        "real_div": 1,
+        "reduce_sum": 3,
+        "relu": 3,
+        "reshape": 31,
+        "select": 7,
+        "slice_by_index": 13,
+        "softmax": 5,
+        "squeeze": 3,
+        "tile": 2,
+        "transpose": 21
+      },
+      "script_sha256": "33db3295450428d90c1de5a804ab47bf3507c686fbbdef67cd276ce116758b5f"
+    },
+    "quantization": {
+      "name": "int4-weights",
+      "settings": {
+        "mode": "linear_symmetric",
+        "dtype": "int4",
+        "granularity": "per_channel",
+        "weight_threshold": 2048
+      },
+      "activations": "float16",
+      "calibration": "None; weight-only, data-free quantization",
+      "source_conversion_sha256": "01105dc8bef7a30fc3430e150caa2a143dc0ba23a295b6280de3b8ff3b40af55",
+      "source_package_files": {
+        "Data/com.apple.CoreML/model.mlmodel": "6ab2df320df7081480b76876df0f4bfd2c841592d93f354c568ba504f7269627",
+        "Data/com.apple.CoreML/weights/weight.bin": "4da9259f798e44f5a1b50769ee1916fd3747c4d723dd9997b516c7fe238c7895",
+        "Manifest.json": "e5b1c7dc66b2014b3985c180a75548a709788c6e314d9d2ae14e4e6d441c3c68"
+      },
+      "source_package_bytes": 1511080,
+      "package_bytes": 451595,
+      "compression_seconds": 0.2996572910051327,
+      "packed_int4_tensors": 19,
+      "operation_counts": {
+        "add": 15,
+        "cast": 13,
+        "concat": 2,
+        "const": 226,
+        "constexpr_blockwise_shift_scale": 19,
+        "equal": 2,
+        "expand_dims": 5,
+        "fill_like": 3,
+        "gather": 2,
+        "greater_equal": 4,
+        "layer_norm": 8,
+        "linear": 15,
+        "logical_not": 2,
+        "matmul": 8,
+        "maximum": 1,
+        "mul": 7,
+        "not_equal": 3,
+        "real_div": 1,
+        "reduce_sum": 3,
+        "relu": 3,
+        "reshape": 31,
+        "select": 9,
+        "slice_by_index": 13,
+        "softmax": 5,
+        "squeeze": 3,
+        "tile": 2,
+        "transpose": 21
+      },
+      "script_sha256": "a91636e539d8934253b5e7628ca706346fbee538fa1aa9e3eaf794f440d49dcc"
+    }
+  },
+  "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": 2.3931670002639294,
+      "p95_ms": 3.033687506103888,
+      "min_ms": 1.7705829814076424,
+      "max_ms": 3.819000005023554
+    }
+  },
+  "export_adapter_fp32": {
+    "argmax_agreement": 196,
+    "max_abs_probability_error": 1.1324882507324219e-06
+  },
+  "host_preprocessing_latency_ms": {
+    "count": 196,
+    "median_ms": 0.02962502185255289,
+    "p95_ms": 0.038823003706056625,
+    "min_ms": 0.024125009076669812,
+    "max_ms": 0.06787499296478927
+  },
+  "thresholds": {
+    "argmax_agreement": 1.0,
+    "max_abs_probability_error": 0.005,
+    "export_adapter_max_abs_probability_error": 0.0001,
+    "allow_accuracy_loss": false
+  },
+  "backends": {
+    "CPU_AND_NE": {
+      "passed": false,
+      "correct": 196,
+      "accuracy": 1.0,
+      "argmax_agreement": 196,
+      "max_abs_probability_error": 0.20076976716518402,
+      "mean_row_max_abs_probability_error": 0.0011279110022456198,
+      "max_abs_logit_error": 6.7208967208862305,
+      "load_seconds": 0.6122147499991115,
+      "warm_inference_latency_ms": {
+        "count": 196,
+        "median_ms": 0.9312290058005601,
+        "p95_ms": 1.0691460047382861,
+        "min_ms": 0.8852920145727694,
+        "max_ms": 1.2105409987270832
+      },
+      "reversed_option_order": {
+        "rows": 6,
+        "max_abs_probability_error": 6.0313243011478335e-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": 6.0313243011478335e-05,
+          "max_abs_logit_error": 2.864961624145508
+        },
+        {
+          "row": 1,
+          "label": 23,
+          "upstream": 23,
+          "coreml": 23,
+          "max_abs_probability_error": 3.814697265625e-05,
+          "max_abs_logit_error": 3.1371307373046875
+        },
+        {
+          "row": 2,
+          "label": 2,
+          "upstream": 2,
+          "coreml": 2,
+          "max_abs_probability_error": 1.1920928955078125e-07,
+          "max_abs_logit_error": 2.247527599334717
+        },
+        {
+          "row": 3,
+          "label": 5,
+          "upstream": 5,
+          "coreml": 5,
+          "max_abs_probability_error": 0.00022571229783352464,
+          "max_abs_logit_error": 4.824214935302734
+        },
+        {
+          "row": 4,
+          "label": 4,
+          "upstream": 4,
+          "coreml": 4,
+          "max_abs_probability_error": 3.3578715374460444e-05,
+          "max_abs_logit_error": 3.618511438369751
+        },
+        {
+          "row": 5,
+          "label": 6,
+          "upstream": 6,
+          "coreml": 6,
+          "max_abs_probability_error": 6.828465848229825e-05,
+          "max_abs_logit_error": 3.7253708839416504
+        },
+        {
+          "row": 6,
+          "label": 7,
+          "upstream": 7,
+          "coreml": 7,
+          "max_abs_probability_error": 2.3365020751953125e-05,
+          "max_abs_logit_error": 2.4195241928100586
+        },
+        {
+          "row": 7,
+          "label": 8,
+          "upstream": 8,
+          "coreml": 8,
+          "max_abs_probability_error": 0.004736423492431641,
+          "max_abs_logit_error": 3.612703323364258
+        },
+        {
+          "row": 8,
+          "label": 9,
+          "upstream": 9,
+          "coreml": 9,
+          "max_abs_probability_error": 1.1920928955078125e-06,
+          "max_abs_logit_error": 3.931535482406616
+        },
+        {
+          "row": 9,
+          "label": 10,
+          "upstream": 10,
+          "coreml": 10,
+          "max_abs_probability_error": 2.1457672119140625e-05,
+          "max_abs_logit_error": 3.5902247428894043
+        },
+        {
+          "row": 10,
+          "label": 11,
+          "upstream": 11,
+          "coreml": 11,
+          "max_abs_probability_error": 9.810033816393116e-07,
+          "max_abs_logit_error": 4.762610912322998
+        },
+        {
+          "row": 11,
+          "label": 13,
+          "upstream": 13,
+          "coreml": 13,
+          "max_abs_probability_error": 4.136032657697797e-06,
+          "max_abs_logit_error": 3.376967430114746
+        },
+        {
+          "row": 12,
+          "label": 15,
+          "upstream": 15,
+          "coreml": 15,
+          "max_abs_probability_error": 3.272867388659506e-06,
+          "max_abs_logit_error": 2.41459059715271
+        },
+        {
+          "row": 13,
+          "label": 16,
+          "upstream": 16,
+          "coreml": 16,
+          "max_abs_probability_error": 0.006299853324890137,
+          "max_abs_logit_error": 2.80385684967041
+        },
+        {
+          "row": 14,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 7.480397471226752e-05,
+          "max_abs_logit_error": 2.1806278228759766
+        },
+        {
+          "row": 15,
+          "label": 24,
+          "upstream": 24,
+          "coreml": 24,
+          "max_abs_probability_error": 1.6689300537109375e-05,
+          "max_abs_logit_error": 1.947547435760498
+        },
+        {
+          "row": 16,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 4.5203719878372794e-08,
+          "max_abs_logit_error": 3.743783950805664
+        },
+        {
+          "row": 17,
+          "label": 25,
+          "upstream": 25,
+          "coreml": 25,
+          "max_abs_probability_error": 5.519996193470433e-06,
+          "max_abs_logit_error": 3.9063854217529297
+        },
+        {
+          "row": 18,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 8.430758313693332e-09,
+          "max_abs_logit_error": 1.3515470027923584
+        },
+        {
+          "row": 19,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 3.0266471551243512e-09,
+          "max_abs_logit_error": 3.501469850540161
+        },
+        {
+          "row": 20,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 2.4500966588902884e-08,
+          "max_abs_logit_error": 3.2016849517822266
+        },
+        {
+          "row": 21,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 1.647379281599637e-11,
+          "max_abs_logit_error": 2.692106246948242
+        },
+        {
+          "row": 22,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 7.152557373046875e-07,
+          "max_abs_logit_error": 2.046893835067749
+        },
+        {
+          "row": 23,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 5.960464477539062e-07,
+          "max_abs_logit_error": 4.133575439453125
+        },
+        {
+          "row": 24,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 2.5055844551924444e-11,
+          "max_abs_logit_error": 2.4618735313415527
+        },
+        {
+          "row": 25,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 6.995705792434137e-09,
+          "max_abs_logit_error": 3.3376502990722656
+        },
+        {
+          "row": 26,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 6.470584068551943e-10,
+          "max_abs_logit_error": 2.2433438301086426
+        },
+        {
+          "row": 27,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 6.973024935241767e-11,
+          "max_abs_logit_error": 2.810356616973877
+        },
+        {
+          "row": 28,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 5.994321927715873e-09,
+          "max_abs_logit_error": 3.7122511863708496
+        },
+        {
+          "row": 29,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 2.4426039316183257e-11,
+          "max_abs_logit_error": 2.806028366088867
+        },
+        {
+          "row": 30,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 2.498983349141781e-07,
+          "max_abs_logit_error": 2.376333236694336
+        },
+        {
+          "row": 31,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 7.78299757975276e-10,
+          "max_abs_logit_error": 2.9533097743988037
+        },
+        {
+          "row": 32,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 6.185003176284454e-09,
+          "max_abs_logit_error": 2.1958632469177246
+        },
+        {
+          "row": 33,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 3.4556297823229443e-09,
+          "max_abs_logit_error": 3.4985671043395996
+        },
+        {
+          "row": 34,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 0.0004820823669433594,
+          "max_abs_logit_error": 3.9588379859924316
+        },
+        {
+          "row": 35,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 1.462750809366753e-09,
+          "max_abs_logit_error": 2.490422248840332
+        },
+        {
+          "row": 36,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 2.7510341227277024e-10,
+          "max_abs_logit_error": 1.6093449592590332
+        },
+        {
+          "row": 37,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 3.74394488744656e-08,
+          "max_abs_logit_error": 1.8185267448425293
+        },
+        {
+          "row": 38,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 9.552543478452691e-11,
+          "max_abs_logit_error": 2.2058591842651367
+        },
+        {
+          "row": 39,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 4.04716482549361e-10,
+          "max_abs_logit_error": 2.799337863922119
+        },
+        {
+          "row": 40,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 0.0019283411093056202,
+          "max_abs_logit_error": 3.8344945907592773
+        },
+        {
+          "row": 41,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 5.345841103121529e-08,
+          "max_abs_logit_error": 2.8801822662353516
+        },
+        {
+          "row": 42,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 1.8819494851385343e-09,
+          "max_abs_logit_error": 1.621840476989746
+        },
+        {
+          "row": 43,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 7.152557373046875e-07,
+          "max_abs_logit_error": 2.8215017318725586
+        },
+        {
+          "row": 44,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 1.780525865635596e-10,
+          "max_abs_logit_error": 1.614362359046936
+        },
+        {
+          "row": 45,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 3.414331317674879e-11,
+          "max_abs_logit_error": 1.6064977645874023
+        },
+        {
+          "row": 46,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 5.599257724142603e-10,
+          "max_abs_logit_error": 1.4032363891601562
+        },
+        {
+          "row": 47,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 5.212314607705437e-11,
+          "max_abs_logit_error": 1.6732616424560547
+        },
+        {
+          "row": 48,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 7.480397471226752e-05,
+          "max_abs_logit_error": 2.1806278228759766
+        },
+        {
+          "row": 49,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 1.932842080831776e-10,
+          "max_abs_logit_error": 1.8353347778320312
+        },
+        {
+          "row": 50,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 4.5203719878372794e-08,
+          "max_abs_logit_error": 3.743783950805664
+        },
+        {
+          "row": 51,
+          "label": 25,
+          "upstream": 25,
+          "coreml": 25,
+          "max_abs_probability_error": 5.519996193470433e-06,
+          "max_abs_logit_error": 3.9063854217529297
+        },
+        {
+          "row": 52,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 8.430758313693332e-09,
+          "max_abs_logit_error": 1.3515470027923584
+        },
+        {
+          "row": 53,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 3.0266471551243512e-09,
+          "max_abs_logit_error": 3.501469850540161
+        },
+        {
+          "row": 54,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 2.4500966588902884e-08,
+          "max_abs_logit_error": 3.2016849517822266
+        },
+        {
+          "row": 55,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 1.647379281599637e-11,
+          "max_abs_logit_error": 2.692106246948242
+        },
+        {
+          "row": 56,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 7.152557373046875e-07,
+          "max_abs_logit_error": 2.046893835067749
+        },
+        {
+          "row": 57,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 5.960464477539062e-07,
+          "max_abs_logit_error": 4.133575439453125
+        },
+        {
+          "row": 58,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 2.5055844551924444e-11,
+          "max_abs_logit_error": 2.4618735313415527
+        },
+        {
+          "row": 59,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 6.995705792434137e-09,
+          "max_abs_logit_error": 3.3376502990722656
+        },
+        {
+          "row": 60,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 6.470584068551943e-10,
+          "max_abs_logit_error": 2.2433438301086426
+        },
+        {
+          "row": 61,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 6.973024935241767e-11,
+          "max_abs_logit_error": 2.810356616973877
+        },
+        {
+          "row": 62,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 5.994321927715873e-09,
+          "max_abs_logit_error": 3.7122511863708496
+        },
+        {
+          "row": 63,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 2.4426039316183257e-11,
+          "max_abs_logit_error": 2.806028366088867
+        },
+        {
+          "row": 64,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 2.498983349141781e-07,
+          "max_abs_logit_error": 2.376333236694336
+        },
+        {
+          "row": 65,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 7.78299757975276e-10,
+          "max_abs_logit_error": 2.9533097743988037
+        },
+        {
+          "row": 66,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 6.185003176284454e-09,
+          "max_abs_logit_error": 2.1958632469177246
+        },
+        {
+          "row": 67,
+          "label": 26,
+          "upstream": 26,
+          "coreml": 26,
+          "max_abs_probability_error": 3.4556297823229443e-09,
+          "max_abs_logit_error": 3.4985671043395996
+        },
+        {
+          "row": 68,
+          "label": 0,
+          "upstream": 0,
+          "coreml": 0,
+          "max_abs_probability_error": 0.00011065257422160357,
+          "max_abs_logit_error": 2.1242733001708984
+        },
+        {
+          "row": 69,
+          "label": 1,
+          "upstream": 1,
+          "coreml": 1,
+          "max_abs_probability_error": 5.5948574299691245e-05,
+          "max_abs_logit_error": 4.156556606292725
+        },
+        {
+          "row": 70,
+          "label": 2,
+          "upstream": 2,
+          "coreml": 2,
+          "max_abs_probability_error": 3.2760276553744916e-06,
+          "max_abs_logit_error": 4.688884735107422
+        },
+        {
+          "row": 71,
+          "label": 3,
+          "upstream": 3,
+          "coreml": 3,
+          "max_abs_probability_error": 5.781622166978195e-05,
+          "max_abs_logit_error": 2.7651166915893555
+        },
+        {
+          "row": 72,
+          "label": 5,
+          "upstream": 5,
+          "coreml": 5,
+          "max_abs_probability_error": 1.0177071089856327e-05,
+          "max_abs_logit_error": 2.8724193572998047
+        },
+        {
+          "row": 73,
+          "label": 6,
+          "upstream": 6,
+          "coreml": 6,
+          "max_abs_probability_error": 2.4447179384878837e-05,
+          "max_abs_logit_error": 2.4996681213378906
+        },
+        {
+          "row": 74,
+          "label": 7,
+          "upstream": 7,
+          "coreml": 7,
+          "max_abs_probability_error": 4.470348358154297e-05,
+          "max_abs_logit_error": 6.7208967208862305
+        },
+        {
+          "row": 75,
+          "label": 8,
+          "upstream": 8,
+          "coreml": 8,
+          "max_abs_probability_error": 3.379677946213633e-05,
+          "max_abs_logit_error": 5.287045478820801
+        },
+        {
+          "row": 76,
+          "label": 9,
+          "upstream": 9,
+          "coreml": 9,
+          "max_abs_probability_error": 4.347872618382098e-06,
+          "max_abs_logit_error": 3.2474308013916016
+        },
+        {
+          "row": 77,
+          "label": 13,
+          "upstream": 13,
+          "coreml": 13,
+          "max_abs_probability_error": 0.0006129741668701172,
+          "max_abs_logit_error": 4.197287559509277
+        },
+        {
+          "row": 78,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 2.0681543446698925e-06,
+          "max_abs_logit_error": 2.6077210903167725
+        },
+        {
+          "row": 79,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 0.0035958904772996902,
+          "max_abs_logit_error": 1.8085737228393555
+        },
+        {
+          "row": 80,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.3343336604521028e-06,
+          "max_abs_logit_error": 1.546478271484375
+        },
+        {
+          "row": 81,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.7881393432617188e-06,
+          "max_abs_logit_error": 1.7384634017944336
+        },
+        {
+          "row": 82,
+          "label": 19,
+          "upstream": 19,
+          "coreml": 19,
+          "max_abs_probability_error": 6.477808369709237e-07,
+          "max_abs_logit_error": 2.3306055068969727
+        },
+        {
+          "row": 83,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 5.0942090545902374e-09,
+          "max_abs_logit_error": 1.5485305786132812
+        },
+        {
+          "row": 84,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 2.384185791015625e-07,
+          "max_abs_logit_error": 1.5241689682006836
+        },
+        {
+          "row": 85,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 3.5405053888659666e-10,
+          "max_abs_logit_error": 2.1916251182556152
+        },
+        {
+          "row": 86,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 5.236582296674896e-07,
+          "max_abs_logit_error": 3.8295583724975586
+        },
+        {
+          "row": 87,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.1920928955078125e-07,
+          "max_abs_logit_error": 3.709770679473877
+        },
+        {
+          "row": 88,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.3188037328859537e-09,
+          "max_abs_logit_error": 3.0971665382385254
+        },
+        {
+          "row": 89,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.9970605169561395e-09,
+          "max_abs_logit_error": 4.972865104675293
+        },
+        {
+          "row": 90,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.7617375336342889e-09,
+          "max_abs_logit_error": 1.8527463674545288
+        },
+        {
+          "row": 91,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.612549260787688e-10,
+          "max_abs_logit_error": 2.7531187534332275
+        },
+        {
+          "row": 92,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 6.506330630512425e-12,
+          "max_abs_logit_error": 3.0166378021240234
+        },
+        {
+          "row": 93,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.5150824594911683e-08,
+          "max_abs_logit_error": 3.1289987564086914
+        },
+        {
+          "row": 94,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.3666674458789885e-09,
+          "max_abs_logit_error": 1.6400773525238037
+        },
+        {
+          "row": 95,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 3.969476136678196e-10,
+          "max_abs_logit_error": 1.7170515060424805
+        },
+        {
+          "row": 96,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 2.657726538846106e-10,
+          "max_abs_logit_error": 2.2562785148620605
+        },
+        {
+          "row": 97,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.9073486328125e-06,
+          "max_abs_logit_error": 1.4498934745788574
+        },
+        {
+          "row": 98,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 2.384185791015625e-07,
+          "max_abs_logit_error": 2.764932632446289
+        },
+        {
+          "row": 99,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 2.352701585550676e-06,
+          "max_abs_logit_error": 3.0918569564819336
+        },
+        {
+          "row": 100,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 9.5367431640625e-07,
+          "max_abs_logit_error": 2.7188186645507812
+        },
+        {
+          "row": 101,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 4.1713853704550274e-08,
+          "max_abs_logit_error": 1.6199064254760742
+        },
+        {
+          "row": 102,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 2.5862225694339713e-09,
+          "max_abs_logit_error": 4.226329803466797
+        },
+        {
+          "row": 103,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 0.00024586915969848633,
+          "max_abs_logit_error": 2.840273380279541
+        },
+        {
+          "row": 104,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 5.841255187988281e-06,
+          "max_abs_logit_error": 3.6580047607421875
+        },
+        {
+          "row": 105,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 2.3792570402747515e-07,
+          "max_abs_logit_error": 6.307399272918701
+        },
+        {
+          "row": 106,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 2.4136979082101107e-09,
+          "max_abs_logit_error": 1.3745036125183105
+        },
+        {
+          "row": 107,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 6.302839361538304e-10,
+          "max_abs_logit_error": 1.8023838996887207
+        },
+        {
+          "row": 108,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.1920928955078125e-07,
+          "max_abs_logit_error": 3.241445541381836
+        },
+        {
+          "row": 109,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 2.0681543446698925e-06,
+          "max_abs_logit_error": 2.6077210903167725
+        },
+        {
+          "row": 110,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 9.5367431640625e-07,
+          "max_abs_logit_error": 2.4543886184692383
+        },
+        {
+          "row": 111,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.811981201171875e-05,
+          "max_abs_logit_error": 1.956678867340088
+        },
+        {
+          "row": 112,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.7881393432617188e-06,
+          "max_abs_logit_error": 1.7384634017944336
+        },
+        {
+          "row": 113,
+          "label": 19,
+          "upstream": 19,
+          "coreml": 19,
+          "max_abs_probability_error": 6.477808369709237e-07,
+          "max_abs_logit_error": 2.3306055068969727
+        },
+        {
+          "row": 114,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 5.0942090545902374e-09,
+          "max_abs_logit_error": 1.5485305786132812
+        },
+        {
+          "row": 115,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 2.384185791015625e-07,
+          "max_abs_logit_error": 1.5241689682006836
+        },
+        {
+          "row": 116,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 3.5405053888659666e-10,
+          "max_abs_logit_error": 2.1916251182556152
+        },
+        {
+          "row": 117,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 5.236582296674896e-07,
+          "max_abs_logit_error": 3.8295583724975586
+        },
+        {
+          "row": 118,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.1920928955078125e-07,
+          "max_abs_logit_error": 3.709770679473877
+        },
+        {
+          "row": 119,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.3188037328859537e-09,
+          "max_abs_logit_error": 3.0971665382385254
+        },
+        {
+          "row": 120,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.9970605169561395e-09,
+          "max_abs_logit_error": 4.972865104675293
+        },
+        {
+          "row": 121,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.7617375336342889e-09,
+          "max_abs_logit_error": 1.8527463674545288
+        },
+        {
+          "row": 122,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.612549260787688e-10,
+          "max_abs_logit_error": 2.7531187534332275
+        },
+        {
+          "row": 123,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 6.506330630512425e-12,
+          "max_abs_logit_error": 3.0166378021240234
+        },
+        {
+          "row": 124,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.5150824594911683e-08,
+          "max_abs_logit_error": 3.1289987564086914
+        },
+        {
+          "row": 125,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.3666674458789885e-09,
+          "max_abs_logit_error": 1.6400773525238037
+        },
+        {
+          "row": 126,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 3.969476136678196e-10,
+          "max_abs_logit_error": 1.7170515060424805
+        },
+        {
+          "row": 127,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 2.657726538846106e-10,
+          "max_abs_logit_error": 2.2562785148620605
+        },
+        {
+          "row": 128,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 1.9073486328125e-06,
+          "max_abs_logit_error": 1.4498934745788574
+        },
+        {
+          "row": 129,
+          "label": 20,
+          "upstream": 20,
+          "coreml": 20,
+          "max_abs_probability_error": 2.384185791015625e-07,
+          "max_abs_logit_error": 2.764932632446289
+        },
+        {
+          "row": 130,
+          "label": 1,
+          "upstream": 1,
+          "coreml": 1,
+          "max_abs_probability_error": 4.3291383917676285e-05,
+          "max_abs_logit_error": 3.559276580810547
+        },
+        {
+          "row": 131,
+          "label": 2,
+          "upstream": 2,
+          "coreml": 2,
+          "max_abs_probability_error": 8.344650268554688e-07,
+          "max_abs_logit_error": 3.2550487518310547
+        },
+        {
+          "row": 132,
+          "label": 3,
+          "upstream": 3,
+          "coreml": 3,
+          "max_abs_probability_error": 2.7894973754882812e-05,
+          "max_abs_logit_error": 3.068411111831665
+        },
+        {
+          "row": 133,
+          "label": 4,
+          "upstream": 4,
+          "coreml": 4,
+          "max_abs_probability_error": 0.0002166023332392797,
+          "max_abs_logit_error": 3.101825714111328
+        },
+        {
+          "row": 134,
+          "label": 5,
+          "upstream": 5,
+          "coreml": 5,
+          "max_abs_probability_error": 6.258487701416016e-05,
+          "max_abs_logit_error": 2.6334457397460938
+        },
+        {
+          "row": 135,
+          "label": 6,
+          "upstream": 6,
+          "coreml": 6,
+          "max_abs_probability_error": 1.2993812561035156e-05,
+          "max_abs_logit_error": 3.064577102661133
+        },
+        {
+          "row": 136,
+          "label": 7,
+          "upstream": 7,
+          "coreml": 7,
+          "max_abs_probability_error": 3.9360515074804425e-05,
+          "max_abs_logit_error": 3.498318672180176
+        },
+        {
+          "row": 137,
+          "label": 8,
+          "upstream": 8,
+          "coreml": 8,
+          "max_abs_probability_error": 0.20076976716518402,
+          "max_abs_logit_error": 4.9444427490234375
+        },
+        {
+          "row": 138,
+          "label": 9,
+          "upstream": 9,
+          "coreml": 9,
+          "max_abs_probability_error": 0.000747382640838623,
+          "max_abs_logit_error": 2.80314302444458
+        },
+        {
+          "row": 139,
+          "label": 10,
+          "upstream": 10,
+          "coreml": 10,
+          "max_abs_probability_error": 5.860753663000651e-05,
+          "max_abs_logit_error": 4.217677116394043
+        },
+        {
+          "row": 140,
+          "label": 11,
+          "upstream": 11,
+          "coreml": 11,
+          "max_abs_probability_error": 1.1063334568461869e-05,
+          "max_abs_logit_error": 3.9767203330993652
+        },
+        {
+          "row": 141,
+          "label": 12,
+          "upstream": 12,
+          "coreml": 12,
+          "max_abs_probability_error": 9.34600830078125e-05,
+          "max_abs_logit_error": 2.8596105575561523
+        },
+        {
+          "row": 142,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.9073486328125e-06,
+          "max_abs_logit_error": 3.8807570934295654
+        },
+        {
+          "row": 143,
+          "label": 16,
+          "upstream": 16,
+          "coreml": 16,
+          "max_abs_probability_error": 1.6191354461625451e-06,
+          "max_abs_logit_error": 2.416942596435547
+        },
+        {
+          "row": 144,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 2.6463640097063035e-06,
+          "max_abs_logit_error": 2.728714942932129
+        },
+        {
+          "row": 145,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 3.6954879760742188e-06,
+          "max_abs_logit_error": 3.8880889415740967
+        },
+        {
+          "row": 146,
+          "label": 17,
+          "upstream": 17,
+          "coreml": 17,
+          "max_abs_probability_error": 5.8316572904004715e-06,
+          "max_abs_logit_error": 4.496463298797607
+        },
+        {
+          "row": 147,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 8.429776876539563e-09,
+          "max_abs_logit_error": 2.1004066467285156
+        },
+        {
+          "row": 148,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 9.215698426601193e-09,
+          "max_abs_logit_error": 1.8194046020507812
+        },
+        {
+          "row": 149,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 3.84208304060607e-10,
+          "max_abs_logit_error": 4.05654764175415
+        },
+        {
+          "row": 150,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 7.947568818333917e-12,
+          "max_abs_logit_error": 2.169252395629883
+        },
+        {
+          "row": 151,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 5.291168614363073e-10,
+          "max_abs_logit_error": 2.1067094802856445
+        },
+        {
+          "row": 152,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 7.924417788629512e-10,
+          "max_abs_logit_error": 3.3871490955352783
+        },
+        {
+          "row": 153,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.0083449808462319e-07,
+          "max_abs_logit_error": 4.667112827301025
+        },
+        {
+          "row": 154,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 3.3051329034750054e-11,
+          "max_abs_logit_error": 2.944230079650879
+        },
+        {
+          "row": 155,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.650468511860126e-10,
+          "max_abs_logit_error": 2.179288864135742
+        },
+        {
+          "row": 156,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 6.181078218703284e-12,
+          "max_abs_logit_error": 2.217021942138672
+        },
+        {
+          "row": 157,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.6455586893115992e-10,
+          "max_abs_logit_error": 2.0631823539733887
+        },
+        {
+          "row": 158,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 6.109900363426846e-10,
+          "max_abs_logit_error": 3.4993996620178223
+        },
+        {
+          "row": 159,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 6.693580800742893e-09,
+          "max_abs_logit_error": 2.16013240814209
+        },
+        {
+          "row": 160,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 4.588627788848498e-08,
+          "max_abs_logit_error": 3.608731269836426
+        },
+        {
+          "row": 161,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 8.344650268554688e-07,
+          "max_abs_logit_error": 3.2034151554107666
+        },
+        {
+          "row": 162,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.042114572413766e-08,
+          "max_abs_logit_error": 3.6694984436035156
+        },
+        {
+          "row": 163,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.9131388140358752e-10,
+          "max_abs_logit_error": 2.9220056533813477
+        },
+        {
+          "row": 164,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 4.935061781452532e-08,
+          "max_abs_logit_error": 1.4873247146606445
+        },
+        {
+          "row": 165,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 5.15840287151903e-11,
+          "max_abs_logit_error": 1.6720411777496338
+        },
+        {
+          "row": 166,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 2.6797479790729994e-09,
+          "max_abs_logit_error": 1.2103171348571777
+        },
+        {
+          "row": 167,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 7.54509399403247e-10,
+          "max_abs_logit_error": 2.835115432739258
+        },
+        {
+          "row": 168,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 8.344650268554688e-07,
+          "max_abs_logit_error": 3.596881866455078
+        },
+        {
+          "row": 169,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 2.4169340140378637e-11,
+          "max_abs_logit_error": 1.5558462142944336
+        },
+        {
+          "row": 170,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 3.0329603412093675e-11,
+          "max_abs_logit_error": 2.1443252563476562
+        },
+        {
+          "row": 171,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.7881393432617188e-06,
+          "max_abs_logit_error": 1.609299659729004
+        },
+        {
+          "row": 172,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 4.66337235494052e-09,
+          "max_abs_logit_error": 2.1106224060058594
+        },
+        {
+          "row": 173,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.2885730260592254e-09,
+          "max_abs_logit_error": 1.5084967613220215
+        },
+        {
+          "row": 174,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.1279766010119374e-09,
+          "max_abs_logit_error": 2.5634703636169434
+        },
+        {
+          "row": 175,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.9073486328125e-06,
+          "max_abs_logit_error": 3.8807570934295654
+        },
+        {
+          "row": 176,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.0728836059570312e-06,
+          "max_abs_logit_error": 2.2860074043273926
+        },
+        {
+          "row": 177,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 6.665992202670168e-08,
+          "max_abs_logit_error": 2.732346534729004
+        },
+        {
+          "row": 178,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 3.6954879760742188e-06,
+          "max_abs_logit_error": 3.8880889415740967
+        },
+        {
+          "row": 179,
+          "label": 17,
+          "upstream": 17,
+          "coreml": 17,
+          "max_abs_probability_error": 5.8316572904004715e-06,
+          "max_abs_logit_error": 4.496463298797607
+        },
+        {
+          "row": 180,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 8.429776876539563e-09,
+          "max_abs_logit_error": 2.1004066467285156
+        },
+        {
+          "row": 181,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 9.215698426601193e-09,
+          "max_abs_logit_error": 1.8194046020507812
+        },
+        {
+          "row": 182,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 3.84208304060607e-10,
+          "max_abs_logit_error": 4.05654764175415
+        },
+        {
+          "row": 183,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 7.947568818333917e-12,
+          "max_abs_logit_error": 2.169252395629883
+        },
+        {
+          "row": 184,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 5.291168614363073e-10,
+          "max_abs_logit_error": 2.1067094802856445
+        },
+        {
+          "row": 185,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 7.924417788629512e-10,
+          "max_abs_logit_error": 3.3871490955352783
+        },
+        {
+          "row": 186,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.0083449808462319e-07,
+          "max_abs_logit_error": 4.667112827301025
+        },
+        {
+          "row": 187,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 3.3051329034750054e-11,
+          "max_abs_logit_error": 2.944230079650879
+        },
+        {
+          "row": 188,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.650468511860126e-10,
+          "max_abs_logit_error": 2.179288864135742
+        },
+        {
+          "row": 189,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 6.181078218703284e-12,
+          "max_abs_logit_error": 2.217021942138672
+        },
+        {
+          "row": 190,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.6455586893115992e-10,
+          "max_abs_logit_error": 2.0631823539733887
+        },
+        {
+          "row": 191,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 6.109900363426846e-10,
+          "max_abs_logit_error": 3.4993996620178223
+        },
+        {
+          "row": 192,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 6.693580800742893e-09,
+          "max_abs_logit_error": 2.16013240814209
+        },
+        {
+          "row": 193,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 4.588627788848498e-08,
+          "max_abs_logit_error": 3.608731269836426
+        },
+        {
+          "row": 194,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 8.344650268554688e-07,
+          "max_abs_logit_error": 3.2034151554107666
+        },
+        {
+          "row": 195,
+          "label": 18,
+          "upstream": 18,
+          "coreml": 18,
+          "max_abs_probability_error": 1.042114572413766e-08,
+          "max_abs_logit_error": 3.6694984436035156
+        }
+      ]
+    }
+  },
+  "passed": false,
+  "code_sha256": {
+    "ane_gather.py": "c3379a3797041dcfed82eaaea3990771ec91f1dff7ec146b025988ad7a8c9aae",
+    "assets.py": "f7a5bc931075237e00c73bf21fee3575be35531f934f23b5e22627d50a79f92f",
+    "benchmark-synthetic.py": "ba8255fd3f63d87ef340c26c4c251800f2e20878d80858ad1e6557cf59f249a2",
+    "compare-ane.py": "4633df7ad5f90d6aed37d5ef4f7a2d2231fdadd272ef1aa3057c237b2f937df8",
+    "convert-coreml.py": "46fe0194df924033126286e601214b6789f5d667aae38c153cf4dbc1b0e629b0",
+    "export-reference.py": "ff9eab4d6e4e72ffaf074d82610f41fbea1e802da7c0e987ce562a89eb561b99",
+    "export_model.py": "c61da5d678c77030d1e9a78359626063ea31da46fb6434e49517fdeaa7800949",
+    "prepare-int4-source.py": "33db3295450428d90c1de5a804ab47bf3507c686fbbdef67cd276ce116758b5f",
+    "preprocessing.py": "18c74ede43f95f91cd638c7e9631923a353e3a24d3b522ddb8b18b480f9c2846",
+    "profile-coreml.py": "d5e1386f54fc6c482647120e593e7ffbbe4bbb277c2cc6d4949154660fe239fd",
+    "quantize-int4.py": "a91636e539d8934253b5e7628ca706346fbee538fa1aa9e3eaf794f440d49dcc",
+    "quantize-int8.py": "22071babfcb02ef77225203a479d9bb14b9fa35462ba38bcc4bf07d29c125a97",
+    "score-report.py": "c78e45caa86fb59f9a56ef18cc6a35020ea6e281a6a23d23618f56da85a0d95e",
+    "synthetic_test.py": "e2b68725236e6827eb422480c8cc8aa189b2490c1594650d8f2928771d1bc626",
+    "verify.py": "bc20ec9df47708e9b48e411076abe4eab7c18313f67e11334de2d9e3ecb3efa7"
+  }
+}

reports/int4-fallback.json

@@ -0,0 +1,112 @@
+{
+  "hardware": {
+    "device": "arm64",
+    "chip": "Apple M5 Pro",
+    "ram": "24GB",
+    "os_version": "macOS 27.0",
+    "timestamp": "2026-09-19T22:28:18.483796+00:00"
+  },
+  "models": [
+    {
+      "model_path": "/Users/hanweng/Documents/FluidAudio/.mobius/cua-s1-prs/mobius/models/computer-use/cua-s1-forms/coreml/build/int4-weights/cua_s1_forms_int4_options32.mlmodelc",
+      "model_name": "cua_s1_forms_int4_options32",
+      "fallback": {
+        "compute_units": "cpu_and_neural_engine",
+        "total_ops": 200,
+        "ane_ops": 149,
+        "gpu_ops": 0,
+        "cpu_ops": 51,
+        "ane_percent": 74.5,
+        "reasons": [
+          {
+            "reason": "Unsupported tensor data type: int32",
+            "count": 25,
+            "estimated_cpu_runtime_ms": 0.0375,
+            "op_types": {
+              "ios18.cast": 10,
+              "ios18.slice_by_index": 4,
+              "ios18.add": 4,
+              "select": 4,
+              "ios18.gather": 2,
+              "ios18.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",
+              "cast_0",
+              "add_0",
+              "select_0_1",
+              "op_59_cast_fp16_cast_uint16_cast_uint16",
+              "flat_ids",
+              "var_333",
+              "var_349",
+              "var_349_to_fp16",
+              "flat_ids_to_int16",
+              "cast_55",
+              "add_4",
+              "select_1",
+              "select_1_to_int16",
+              "cast_1",
+              "add_1",
+              "select_1_1",
+              "op_365_cast_fp16_cast_uint16_cast_uint16"
+            ]
+          },
+          {
+            "reason": "ANE not available for this op",
+            "count": 19,
+            "estimated_cpu_runtime_ms": 0,
+            "op_types": {
+              "ios18.constexpr_blockwise_shift_scale": 19
+            },
+            "ops": [
+              "model_embedding_weight_to_fp16_quantized",
+              "model_position_weight_to_fp16_quantized",
+              "model_encoder_layers_0_self_attn_in_proj_weight_to_fp16_quantized",
+              "model_encoder_layers_0_self_attn_out_proj_weight_to_fp16_quantized",
+              "model_encoder_layers_0_linear1_weight_to_fp16_quantized",
+              "model_encoder_layers_0_linear2_weight_to_fp16_quantized",
+              "model_encoder_layers_1_self_attn_in_proj_weight_to_fp16_quantized",
+              "model_encoder_layers_1_self_attn_out_proj_weight_to_fp16_quantized",
+              "model_encoder_layers_1_linear1_weight_to_fp16_quantized",
+              "model_encoder_layers_1_linear2_weight_to_fp16_quantized",
+              "op_369_to_fp16_quantized",
+              "op_394_to_fp16_quantized",
+              "model_option_encoder_layers_0_self_attn_in_proj_weight_to_fp16_quantized",
+              "model_option_encoder_layers_0_self_attn_out_proj_weight_to_fp16_quantized",
+              "model_option_encoder_layers_0_linear1_weight_to_fp16_quantized",
+              "model_option_encoder_layers_0_linear2_weight_to_fp16_quantized",
+              "model_head_query_weight_to_fp16_quantized",
+              "model_head_key_weight_to_fp16_quantized",
+              "model_head_value_weight_to_fp16_quantized"
+            ]
+          },
+          {
+            "reason": "Unable to resolve operation input \"y\".",
+            "count": 7,
+            "estimated_cpu_runtime_ms": 0.007,
+            "op_types": {
+              "ios18.greater_equal": 4,
+              "ios18.not_equal": 3
+            },
+            "ops": [
+              "option_mask_1",
+              "context_mask",
+              "greater_equal_0",
+              "greater_equal_0_1",
+              "flat_mask",
+              "greater_equal_1",
+              "greater_equal_1_1"
+            ]
+          }
+        ]
+      }
+    }
+  ]
+}

reports/int4-profile.json

@@ -0,0 +1,1105 @@
+{
+  "created_utc": "2026-09-19T22:27:49.746625+00:00",
+  "purpose": "Bounded real-input latency and scheduler placement; not a general accuracy benchmark",
+  "model_revision": "f54adbf447f4ca6ec259f529ee3f2e3e09f8cc71",
+  "package_files": {
+    "Data/com.apple.CoreML/model.mlmodel": "cadad6bfa6875383a7ee3de68d0d22871148aa21cf117e7fe44c2c971287c4e5",
+    "Data/com.apple.CoreML/weights/weight.bin": "0f766fcb9dbd612ed4f6c7a40117a5993a4789b3ae94afba0df37193c831f63c",
+    "Manifest.json": "fb91c8b5ec3168111567de663bb3140d2790c6e7c48f32e0df3f5135af0673af"
+  },
+  "dataset_revision": "8273f34778b99ac2e12d9f6e7d57dad99ae20845",
+  "dataset_sha256": "4f43b442e79ba2e2ce731e27e9b8e340c2b5dfcaffc92d8ff564c34f115ff1ca",
+  "environment": {
+    "chip": "Apple M5 Pro",
+    "memory_bytes": 25769803776,
+    "macos": "27.0",
+    "os_build": "26A428",
+    "python": "3.11.11",
+    "coremltools": "9.0"
+  },
+  "protocol": {
+    "rows": [
+      0,
+      68,
+      130
+    ],
+    "pages": [
+      "patient-registration",
+      "job-application",
+      "auto-claim"
+    ],
+    "option_counts": [
+      27,
+      21,
+      19
+    ],
+    "first_call_row": 0,
+    "warmup_passes": 2,
+    "timed_passes": 10,
+    "backend_order": [
+      "CPU_AND_NE"
+    ],
+    "timing_scope": "Synchronous Python MLModel.predict; pre-encoded inputs; output checks outside timer",
+    "load_scope": "MLModel construction; system caches retained, not first-install cold start",
+    "first_call_scope": "First predict after this load; may use existing system/ANE caches",
+    "plan_order": "Compute plan loaded after timing each backend"
+  },
+  "backends": {
+    "CPU_AND_NE": {
+      "load_ms": 628.8241670117714,
+      "first_prediction_ms": 2.4197499733418226,
+      "warm_prediction": {
+        "count": 30,
+        "median_ms": 0.9370829939143732,
+        "p95_ms": 0.969368452206254,
+        "min_ms": 0.8955409866757691,
+        "max_ms": 0.985999999102205
+      },
+      "timed_predictions_correct": 30,
+      "calls": [
+        {
+          "row": 0,
+          "milliseconds": 0.9536660218145698
+        },
+        {
+          "row": 68,
+          "milliseconds": 0.9417080145794898
+        },
+        {
+          "row": 130,
+          "milliseconds": 0.9293750044889748
+        },
+        {
+          "row": 0,
+          "milliseconds": 0.9304159902967513
+        },
+        {
+          "row": 68,
+          "milliseconds": 0.9457910200580955
+        },
+        {
+          "row": 130,
+          "milliseconds": 0.9614159935154021
+        },
+        {
+          "row": 0,
+          "milliseconds": 0.9506249916739762
+        },
+        {
+          "row": 68,
+          "milliseconds": 0.9356669906992465
+        },
+        {
+          "row": 130,
+          "milliseconds": 0.9454160172026604
+        },
+        {
+          "row": 0,
+          "milliseconds": 0.9370410116389394
+        },
+        {
+          "row": 68,
+          "milliseconds": 0.9322499972768128
+        },
+        {
+          "row": 130,
+          "milliseconds": 0.975875009316951
+        },
+        {
+          "row": 0,
+          "milliseconds": 0.985999999102205
+        },
+        {
+          "row": 68,
+          "milliseconds": 0.960250006755814
+        },
+        {
+          "row": 130,
+          "milliseconds": 0.9522080072201788
+        },
+        {
+          "row": 0,
+          "milliseconds": 0.9541249892208725
+        },
+        {
+          "row": 68,
+          "milliseconds": 0.9371249761898071
+        },
+        {
+          "row": 130,
+          "milliseconds": 0.9184999798890203
+        },
+        {
+          "row": 0,
+          "milliseconds": 0.9398749971296638
+        },
+        {
+          "row": 68,
+          "milliseconds": 0.9219589992426336
+        },
+        {
+          "row": 130,
+          "milliseconds": 0.9396669920533895
+        },
+        {
+          "row": 0,
+          "milliseconds": 0.9407500037923455
+        },
+        {
+          "row": 68,
+          "milliseconds": 0.9344589780084789
+        },
+        {
+          "row": 130,
+          "milliseconds": 0.931665999814868
+        },
+        {
+          "row": 0,
+          "milliseconds": 0.9232910233549774
+        },
+        {
+          "row": 68,
+          "milliseconds": 0.9075840061996132
+        },
+        {
+          "row": 130,
+          "milliseconds": 0.8955409866757691
+        },
+        {
+          "row": 0,
+          "milliseconds": 0.9247090201824903
+        },
+        {
+          "row": 68,
+          "milliseconds": 0.9097080037463456
+        },
+        {
+          "row": 130,
+          "milliseconds": 0.9349999891128391
+        }
+      ],
+      "compute_plan": {
+        "kind": "MLComputePlan preferred device; not a runtime utilization or power trace",
+        "assigned_operation_count": 181,
+        "device_counts": {
+          "cpu": 32,
+          "gpu": 0,
+          "ane": 149,
+          "unknown": 0
+        },
+        "unassigned_operation_types": {
+          "const": 226,
+          "ios18.constexpr_blockwise_shift_scale": 19
+        },
+        "operations": [
+          {
+            "name": "option_mask_1",
+            "type": "ios18.not_equal",
+            "device": "cpu"
+          },
+          {
+            "name": "context_mask",
+            "type": "ios18.not_equal",
+            "device": "cpu"
+          },
+          {
+            "name": "var_27",
+            "type": "ios18.slice_by_index",
+            "device": "cpu"
+          },
+          {
+            "name": "fill_like_0_cast_fp16",
+            "type": "ios16.fill_like",
+            "device": "ane"
+          },
+          {
+            "name": "var_43",
+            "type": "ios18.slice_by_index",
+            "device": "cpu"
+          },
+          {
+            "name": "var_43_to_fp16",
+            "type": "ios18.cast",
+            "device": "cpu"
+          },
+          {
+            "name": "safe_context_ids_cast_fp16",
+            "type": "ios18.concat",
+            "device": "ane"
+          },
+          {
+            "name": "context_ids_to_int16",
+            "type": "ios18.cast",
+            "device": "cpu"
+          },
+          {
+            "name": "cast_54",
+            "type": "ios18.cast",
+            "device": "cpu"
+          },
+          {
+            "name": "greater_equal_0",
+            "type": "ios18.greater_equal",
+            "device": "cpu"
+          },
+          {
+            "name": "add_3",
+            "type": "ios18.add",
+            "device": "cpu"
+          },
+          {
+            "name": "select_0",
+            "type": "select",
+            "device": "cpu"
+          },
+          {
+            "name": "select_0_to_int16",
+            "type": "ios18.cast",
+            "device": "cpu"
+          },
+          {
+            "name": "cast_0",
+            "type": "ios18.cast",
+            "device": "cpu"
+          },
+          {
+            "name": "greater_equal_0_1",
+            "type": "ios18.greater_equal",
+            "device": "cpu"
+          },
+          {
+            "name": "add_0",
+            "type": "ios18.add",
+            "device": "cpu"
+          },
+          {
+            "name": "select_0_1",
+            "type": "select",
+            "device": "cpu"
+          },
+          {
+            "name": "op_59_cast_fp16_cast_uint16_cast_uint16",
+            "type": "ios18.gather",
+            "device": "cpu"
+          },
+          {
+            "name": "src_1_cast_fp16",
+            "type": "ios18.add",
+            "device": "ane"
+          },
+          {
+            "name": "mask_1_cast_fp16",
+            "type": "ios18.equal",
+            "device": "ane"
+          },
+          {
+            "name": "key_padding_mask_1_cast_fp16",
+            "type": "select",
+            "device": "ane"
+          },
+          {
+            "name": "query_1_cast_fp16",
+            "type": "ios18.layer_norm",
+            "device": "ane"
+          },
+          {
+            "name": "query_3_cast_fp16",
+            "type": "ios18.transpose",
+            "device": "ane"
+          },
+          {
+            "name": "linear_0_cast_fp16",
+            "type": "ios18.linear",
+            "device": "ane"
+          },
+          {
+            "name": "var_124_cast_fp16",
+            "type": "ios18.reshape",
+            "device": "ane"
+          },
+          {
+            "name": "var_125_cast_fp16",
+            "type": "ios18.expand_dims",
+            "device": "ane"
+          },
+          {
+            "name": "var_126_cast_fp16",
+            "type": "ios18.transpose",
+            "device": "ane"
+          },
+          {
+            "name": "var_127_cast_fp16",
+            "type": "ios18.squeeze",
+            "device": "ane"
+          },
+          {
+            "name": "q_1_cast_fp16",
+            "type": "ios18.slice_by_index",
+            "device": "ane"
+          },
+          {
+            "name": "k_1_cast_fp16",
+            "type": "ios18.slice_by_index",
+            "device": "ane"
+          },
+          {
+            "name": "v_1_cast_fp16",
+            "type": "ios18.slice_by_index",
+            "device": "ane"
+          },
+          {
+            "name": "var_136_cast_fp16",
+            "type": "ios18.reshape",
+            "device": "ane"
+          },
+          {
+            "name": "var_143_cast_fp16",
+            "type": "ios18.reshape",
+            "device": "ane"
+          },
+          {
+            "name": "var_150_cast_fp16",
+            "type": "ios18.reshape",
+            "device": "ane"
+          },
+          {
+            "name": "var_154_cast_fp16",
+            "type": "ios18.reshape",
+            "device": "ane"
+          },
+          {
+            "name": "var_156_cast_fp16",
+            "type": "tile",
+            "device": "ane"
+          },
+          {
+            "name": "q_3_cast_fp16",
+            "type": "ios18.transpose",
+            "device": "ane"
+          },
+          {
+            "name": "q_5_cast_fp16",
+            "type": "ios18.reshape",
+            "device": "ane"
+          },
+          {
+            "name": "k_3_cast_fp16",
+            "type": "ios18.transpose",
+            "device": "ane"
+          },
+          {
+            "name": "k_5_cast_fp16",
+            "type": "ios18.reshape",
+            "device": "ane"
+          },
+          {
+            "name": "v_3_cast_fp16",
+            "type": "ios18.transpose",
+            "device": "ane"
+          },
+          {
+            "name": "v_5_cast_fp16",
+            "type": "ios18.reshape",
+            "device": "ane"
+          },
+          {
+            "name": "mul_1_cast_fp16",
+            "type": "ios18.mul",
+            "device": "ane"
+          },
+          {
+            "name": "matmul_0_cast_fp16",
+            "type": "ios18.matmul",
+            "device": "ane"
+          },
+          {
+            "name": "add_0_cast_fp16",
+            "type": "ios18.add",
+            "device": "ane"
+          },
+          {
+            "name": "softmax_0_cast_fp16",
+            "type": "ios18.softmax",
+            "device": "ane"
+          },
+          {
+            "name": "attn_output_1_cast_fp16",
+            "type": "ios18.matmul",
+            "device": "ane"
+          },
+          {
+            "name": "var_172_cast_fp16",
+            "type": "ios18.transpose",
+            "device": "ane"
+          },
+          {
+            "name": "attn_output_3_cast_fp16",
+            "type": "ios18.reshape",
+            "device": "ane"
+          },
+          {
+            "name": "linear_1_cast_fp16",
+            "type": "ios18.linear",
+            "device": "ane"
+          },
+          {
+            "name": "attn_output_7_cast_fp16",
+            "type": "ios18.reshape",
+            "device": "ane"
+          },
+          {
+            "name": "input_3_cast_fp16",
+            "type": "ios18.transpose",
+            "device": "ane"
+          },
+          {
+            "name": "input_5_cast_fp16",
+            "type": "ios18.add",
+            "device": "ane"
+          },
+          {
+            "name": "input_7_cast_fp16",
+            "type": "ios18.layer_norm",
+            "device": "ane"
+          },
+          {
+            "name": "linear_2_cast_fp16",
+            "type": "ios18.linear",
+            "device": "ane"
+          },
+          {
+            "name": "input_11_cast_fp16",
+            "type": "ios18.relu",
+            "device": "ane"
+          },
+          {
+            "name": "linear_3_cast_fp16",
+            "type": "ios18.linear",
+            "device": "ane"
+          },
+          {
+            "name": "input_17_cast_fp16",
+            "type": "ios18.add",
+            "device": "ane"
+          },
+          {
+            "name": "query_5_cast_fp16",
+            "type": "ios18.layer_norm",
+            "device": "ane"
+          },
+          {
+            "name": "query_7_cast_fp16",
+            "type": "ios18.transpose",
+            "device": "ane"
+          },
+          {
+            "name": "linear_4_cast_fp16",
+            "type": "ios18.linear",
+            "device": "ane"
+          },
+          {
+            "name": "var_231_cast_fp16",
+            "type": "ios18.reshape",
+            "device": "ane"
+          },
+          {
+            "name": "var_232_cast_fp16",
+            "type": "ios18.expand_dims",
+            "device": "ane"
+          },
+          {
+            "name": "var_233_cast_fp16",
+            "type": "ios18.transpose",
+            "device": "ane"
+          },
+          {
+            "name": "var_234_cast_fp16",
+            "type": "ios18.squeeze",
+            "device": "ane"
+          },
+          {
+            "name": "q_7_cast_fp16",
+            "type": "ios18.slice_by_index",
+            "device": "ane"
+          },
+          {
+            "name": "k_7_cast_fp16",
+            "type": "ios18.slice_by_index",
+            "device": "ane"
+          },
+          {
+            "name": "v_7_cast_fp16",
+            "type": "ios18.slice_by_index",
+            "device": "ane"
+          },
+          {
+            "name": "var_243_cast_fp16",
+            "type": "ios18.reshape",
+            "device": "ane"
+          },
+          {
+            "name": "var_250_cast_fp16",
+            "type": "ios18.reshape",
+            "device": "ane"
+          },
+          {
+            "name": "var_257_cast_fp16",
+            "type": "ios18.reshape",
+            "device": "ane"
+          },
+          {
+            "name": "q_9_cast_fp16",
+            "type": "ios18.transpose",
+            "device": "ane"
+          },
+          {
+            "name": "q_11_cast_fp16",
+            "type": "ios18.reshape",
+            "device": "ane"
+          },
+          {
+            "name": "k_9_cast_fp16",
+            "type": "ios18.transpose",
+            "device": "ane"
+          },
+          {
+            "name": "k_11_cast_fp16",
+            "type": "ios18.reshape",
+            "device": "ane"
+          },
+          {
+            "name": "v_9_cast_fp16",
+            "type": "ios18.transpose",
+            "device": "ane"
+          },
+          {
+            "name": "v_11_cast_fp16",
+            "type": "ios18.reshape",
+            "device": "ane"
+          },
+          {
+            "name": "mul_3_cast_fp16",
+            "type": "ios18.mul",
+            "device": "ane"
+          },
+          {
+            "name": "matmul_1_cast_fp16",
+            "type": "ios18.matmul",
+            "device": "ane"
+          },
+          {
+            "name": "add_1_cast_fp16",
+            "type": "ios18.add",
+            "device": "ane"
+          },
+          {
+            "name": "softmax_1_cast_fp16",
+            "type": "ios18.softmax",
+            "device": "ane"
+          },
+          {
+            "name": "attn_output_9_cast_fp16",
+            "type": "ios18.matmul",
+            "device": "ane"
+          },
+          {
+            "name": "var_279_cast_fp16",
+            "type": "ios18.transpose",
+            "device": "ane"
+          },
+          {
+            "name": "attn_output_11_cast_fp16",
+            "type": "ios18.reshape",
+            "device": "ane"
+          },
+          {
+            "name": "linear_5_cast_fp16",
+            "type": "ios18.linear",
+            "device": "ane"
+          },
+          {
+            "name": "attn_output_15_cast_fp16",
+            "type": "ios18.reshape",
+            "device": "ane"
+          },
+          {
+            "name": "input_19_cast_fp16",
+            "type": "ios18.transpose",
+            "device": "ane"
+          },
+          {
+            "name": "input_21_cast_fp16",
+            "type": "ios18.add",
+            "device": "ane"
+          },
+          {
+            "name": "input_23_cast_fp16",
+            "type": "ios18.layer_norm",
+            "device": "ane"
+          },
+          {
+            "name": "linear_6_cast_fp16",
+            "type": "ios18.linear",
+            "device": "ane"
+          },
+          {
+            "name": "input_27_cast_fp16",
+            "type": "ios18.relu",
+            "device": "ane"
+          },
+          {
+            "name": "linear_7_cast_fp16",
+            "type": "ios18.linear",
+            "device": "ane"
+          },
+          {
+            "name": "context_cast_fp16",
+            "type": "ios18.add",
+            "device": "ane"
+          },
+          {
+            "name": "flat_ids",
+            "type": "ios18.reshape",
+            "device": "cpu"
+          },
+          {
+            "name": "flat_mask",
+            "type": "ios18.not_equal",
+            "device": "cpu"
+          },
+          {
+            "name": "var_333",
+            "type": "ios18.slice_by_index",
+            "device": "cpu"
+          },
+          {
+            "name": "fill_like_1_cast_fp16",
+            "type": "ios16.fill_like",
+            "device": "ane"
+          },
+          {
+            "name": "var_349",
+            "type": "ios18.slice_by_index",
+            "device": "cpu"
+          },
+          {
+            "name": "var_349_to_fp16",
+            "type": "ios18.cast",
+            "device": "cpu"
+          },
+          {
+            "name": "safe_ids_cast_fp16",
+            "type": "ios18.concat",
+            "device": "ane"
+          },
+          {
+            "name": "flat_ids_to_int16",
+            "type": "ios18.cast",
+            "device": "cpu"
+          },
+          {
+            "name": "cast_55",
+            "type": "ios18.cast",
+            "device": "cpu"
+          },
+          {
+            "name": "greater_equal_1",
+            "type": "ios18.greater_equal",
+            "device": "cpu"
+          },
+          {
+            "name": "add_4",
+            "type": "ios18.add",
+            "device": "cpu"
+          },
+          {
+            "name": "select_1",
+            "type": "select",
+            "device": "cpu"
+          },
+          {
+            "name": "select_1_to_int16",
+            "type": "ios18.cast",
+            "device": "cpu"
+          },
+          {
+            "name": "cast_1",
+            "type": "ios18.cast",
+            "device": "cpu"
+          },
+          {
+            "name": "greater_equal_1_1",
+            "type": "ios18.greater_equal",
+            "device": "cpu"
+          },
+          {
+            "name": "add_1",
+            "type": "ios18.add",
+            "device": "cpu"
+          },
+          {
+            "name": "select_1_1",
+            "type": "select",
+            "device": "cpu"
+          },
+          {
+            "name": "op_365_cast_fp16_cast_uint16_cast_uint16",
+            "type": "ios18.gather",
+            "device": "cpu"
+          },
+          {
+            "name": "src_cast_fp16",
+            "type": "ios18.add",
+            "device": "ane"
+          },
+          {
+            "name": "mask_cast_fp16",
+            "type": "ios18.equal",
+            "device": "ane"
+          },
+          {
+            "name": "key_padding_mask_7_cast_fp16",
+            "type": "select",
+            "device": "ane"
+          },
+          {
+            "name": "query_9_cast_fp16",
+            "type": "ios18.layer_norm",
+            "device": "ane"
+          },
+          {
+            "name": "query_11_cast_fp16",
+            "type": "ios18.transpose",
+            "device": "ane"
+          },
+          {
+            "name": "linear_8_cast_fp16",
+            "type": "ios18.linear",
+            "device": "ane"
+          },
+          {
+            "name": "var_428_cast_fp16",
+            "type": "ios18.reshape",
+            "device": "ane"
+          },
+          {
+            "name": "var_429_cast_fp16",
+            "type": "ios18.expand_dims",
+            "device": "ane"
+          },
+          {
+            "name": "var_430_cast_fp16",
+            "type": "ios18.transpose",
+            "device": "ane"
+          },
+          {
+            "name": "var_431_cast_fp16",
+            "type": "ios18.squeeze",
+            "device": "ane"
+          },
+          {
+            "name": "q_13_cast_fp16",
+            "type": "ios18.slice_by_index",
+            "device": "ane"
+          },
+          {
+            "name": "k_13_cast_fp16",
+            "type": "ios18.slice_by_index",
+            "device": "ane"
+          },
+          {
+            "name": "v_13_cast_fp16",
+            "type": "ios18.slice_by_index",
+            "device": "ane"
+          },
+          {
+            "name": "var_440_cast_fp16",
+            "type": "ios18.reshape",
+            "device": "ane"
+          },
+          {
+            "name": "var_447_cast_fp16",
+            "type": "ios18.reshape",
+            "device": "ane"
+          },
+          {
+            "name": "var_454_cast_fp16",
+            "type": "ios18.reshape",
+            "device": "ane"
+          },
+          {
+            "name": "var_458_cast_fp16",
+            "type": "ios18.reshape",
+            "device": "ane"
+          },
+          {
+            "name": "var_460_cast_fp16",
+            "type": "tile",
+            "device": "ane"
+          },
+          {
+            "name": "q_15_cast_fp16",
+            "type": "ios18.transpose",
+            "device": "ane"
+          },
+          {
+            "name": "q_cast_fp16",
+            "type": "ios18.reshape",
+            "device": "ane"
+          },
+          {
+            "name": "k_15_cast_fp16",
+            "type": "ios18.transpose",
+            "device": "ane"
+          },
+          {
+            "name": "k_cast_fp16",
+            "type": "ios18.reshape",
+            "device": "ane"
+          },
+          {
+            "name": "v_15_cast_fp16",
+            "type": "ios18.transpose",
+            "device": "ane"
+          },
+          {
+            "name": "v_cast_fp16",
+            "type": "ios18.reshape",
+            "device": "ane"
+          },
+          {
+            "name": "mul_5_cast_fp16",
+            "type": "ios18.mul",
+            "device": "ane"
+          },
+          {
+            "name": "matmul_2_cast_fp16",
+            "type": "ios18.matmul",
+            "device": "ane"
+          },
+          {
+            "name": "add_2_cast_fp16",
+            "type": "ios18.add",
+            "device": "ane"
+          },
+          {
+            "name": "softmax_2_cast_fp16",
+            "type": "ios18.softmax",
+            "device": "ane"
+          },
+          {
+            "name": "attn_output_17_cast_fp16",
+            "type": "ios18.matmul",
+            "device": "ane"
+          },
+          {
+            "name": "var_476_cast_fp16",
+            "type": "ios18.transpose",
+            "device": "ane"
+          },
+          {
+            "name": "attn_output_19_cast_fp16",
+            "type": "ios18.reshape",
+            "device": "ane"
+          },
+          {
+            "name": "linear_9_cast_fp16",
+            "type": "ios18.linear",
+            "device": "ane"
+          },
+          {
+            "name": "attn_output_cast_fp16",
+            "type": "ios18.reshape",
+            "device": "ane"
+          },
+          {
+            "name": "input_35_cast_fp16",
+            "type": "ios18.transpose",
+            "device": "ane"
+          },
+          {
+            "name": "input_37_cast_fp16",
+            "type": "ios18.add",
+            "device": "ane"
+          },
+          {
+            "name": "input_39_cast_fp16",
+            "type": "ios18.layer_norm",
+            "device": "ane"
+          },
+          {
+            "name": "linear_10_cast_fp16",
+            "type": "ios18.linear",
+            "device": "ane"
+          },
+          {
+            "name": "input_43_cast_fp16",
+            "type": "ios18.relu",
+            "device": "ane"
+          },
+          {
+            "name": "linear_11_cast_fp16",
+            "type": "ios18.linear",
+            "device": "ane"
+          },
+          {
+            "name": "hidden_cast_fp16",
+            "type": "ios18.add",
+            "device": "ane"
+          },
+          {
+            "name": "var_504",
+            "type": "ios18.expand_dims",
+            "device": "ane"
+          },
+          {
+            "name": "var_504_to_fp16",
+            "type": "ios18.cast",
+            "device": "ane"
+          },
+          {
+            "name": "var_510_cast_fp16",
+            "type": "ios18.mul",
+            "device": "ane"
+          },
+          {
+            "name": "var_515_cast_fp16",
+            "type": "ios16.reduce_sum",
+            "device": "ane"
+          },
+          {
+            "name": "var_520_cast_fp16",
+            "type": "ios16.reduce_sum",
+            "device": "ane"
+          },
+          {
+            "name": "var_522_cast_fp16",
+            "type": "ios18.maximum",
+            "device": "ane"
+          },
+          {
+            "name": "pooled_cast_fp16",
+            "type": "ios18.real_div",
+            "device": "ane"
+          },
+          {
+            "name": "options_cast_fp16",
+            "type": "ios18.reshape",
+            "device": "ane"
+          },
+          {
+            "name": "input_cast_fp16",
+            "type": "ios18.layer_norm",
+            "device": "ane"
+          },
+          {
+            "name": "input_53_cast_fp16",
+            "type": "ios18.layer_norm",
+            "device": "ane"
+          },
+          {
+            "name": "linear_12_cast_fp16",
+            "type": "ios18.linear",
+            "device": "ane"
+          },
+          {
+            "name": "linear_13_cast_fp16",
+            "type": "ios18.linear",
+            "device": "ane"
+          },
+          {
+            "name": "linear_14_cast_fp16",
+            "type": "ios18.linear",
+            "device": "ane"
+          },
+          {
+            "name": "matmul_3_cast_fp16",
+            "type": "ios18.matmul",
+            "device": "ane"
+          },
+          {
+            "name": "_inversed_scores_1_cast_fp16",
+            "type": "ios18.mul",
+            "device": "ane"
+          },
+          {
+            "name": "var_567",
+            "type": "ios18.expand_dims",
+            "device": "ane"
+          },
+          {
+            "name": "var_569",
+            "type": "ios18.logical_not",
+            "device": "ane"
+          },
+          {
+            "name": "scores_cast_fp16",
+            "type": "select",
+            "device": "ane"
+          },
+          {
+            "name": "var_571_cast_fp16",
+            "type": "ios18.softmax",
+            "device": "ane"
+          },
+          {
+            "name": "matmul_4_cast_fp16",
+            "type": "ios18.matmul",
+            "device": "ane"
+          },
+          {
+            "name": "var_574_cast_fp16",
+            "type": "ios18.mul",
+            "device": "ane"
+          },
+          {
+            "name": "var_576_cast_fp16",
+            "type": "ios16.reduce_sum",
+            "device": "ane"
+          },
+          {
+            "name": "_inversed_logits_1_cast_fp16",
+            "type": "ios18.mul",
+            "device": "ane"
+          },
+          {
+            "name": "var_579",
+            "type": "ios18.logical_not",
+            "device": "ane"
+          },
+          {
+            "name": "logits_3_cast_fp16",
+            "type": "select",
+            "device": "ane"
+          },
+          {
+            "name": "var_587_cast_fp16",
+            "type": "ios16.fill_like",
+            "device": "ane"
+          },
+          {
+            "name": "logits_cast_fp16",
+            "type": "select",
+            "device": "ane"
+          },
+          {
+            "name": "var_591_cast_fp16",
+            "type": "ios18.softmax",
+            "device": "ane"
+          },
+          {
+            "name": "logits",
+            "type": "ios18.cast",
+            "device": "ane"
+          },
+          {
+            "name": "probabilities",
+            "type": "ios18.cast",
+            "device": "ane"
+          }
+        ]
+      }
+    }
+  },
+  "script_sha256": "d5e1386f54fc6c482647120e593e7ffbbe4bbb277c2cc6d4949154660fe239fd"
+}

reports/int4-source-demo-verification.json

@@ -0,0 +1,1748 @@
+{
+  "purpose": "Local conversion parity on the upstream demo; not a generalization or live GUI benchmark",
+  "created_utc": "2026-09-19T22:24:56.262849+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": "iOS18/macOS15",
+    "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": "6ab2df320df7081480b76876df0f4bfd2c841592d93f354c568ba504f7269627",
+      "Data/com.apple.CoreML/weights/weight.bin": "4da9259f798e44f5a1b50769ee1916fd3747c4d723dd9997b516c7fe238c7895",
+      "Manifest.json": "e5b1c7dc66b2014b3985c180a75548a709788c6e314d9d2ae14e4e6d441c3c68"
+    },
+    "target_upgrade": {
+      "source_conversion_sha256": "ac94239751ceb7e19602eb7276940ff53446b1776d4900405af9be0382fe4a94",
+      "source_package_files": {
+        "Data/com.apple.CoreML/model.mlmodel": "70485fc18cbb21785df833cbddddc0b5b59acb00d22394b76e55307e2c135dd0",
+        "Data/com.apple.CoreML/weights/weight.bin": "4da9259f798e44f5a1b50769ee1916fd3747c4d723dd9997b516c7fe238c7895",
+        "Manifest.json": "2bc0f5f62337b27fb6b0ecde248f1e3dc269e1ba4b65516aaeede2a60e293dcc"
+      },
+      "method": "Load original MIL at specification 9; empty pass pipeline; retain decomposed attention",
+      "operation_counts": {
+        "add": 13,
+        "cast": 11,
+        "concat": 2,
+        "const": 239,
+        "equal": 2,
+        "expand_dims": 5,
+        "fill_like": 3,
+        "gather": 2,
+        "greater_equal": 2,
+        "layer_norm": 8,
+        "linear": 15,
+        "logical_not": 2,
+        "matmul": 8,
+        "maximum": 1,
+        "mul": 7,
+        "not_equal": 3,
+        "real_div": 1,
+        "reduce_sum": 3,
+        "relu": 3,
+        "reshape": 31,
+        "select": 7,
+        "slice_by_index": 13,
+        "softmax": 5,
+        "squeeze": 3,
+        "tile": 2,
+        "transpose": 21
+      },
+      "script_sha256": "33db3295450428d90c1de5a804ab47bf3507c686fbbdef67cd276ce116758b5f"
+    }
+  },
+  "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": 2.572771001723595,
+      "p95_ms": 3.2190517595154233,
+      "min_ms": 1.7389580025337636,
+      "max_ms": 4.086540982825682
+    }
+  },
+  "export_adapter_fp32": {
+    "argmax_agreement": 196,
+    "max_abs_probability_error": 1.1324882507324219e-06
+  },
+  "host_preprocessing_latency_ms": {
+    "count": 196,
+    "median_ms": 0.041624996811151505,
+    "p95_ms": 0.052635499741882086,
+    "min_ms": 0.029125018045306206,
+    "max_ms": 0.08079098188318312
+  },
+  "thresholds": {
+    "argmax_agreement": 1.0,
+    "max_abs_probability_error": 0.005,
+    "export_adapter_max_abs_probability_error": 0.0001,
+    "allow_accuracy_loss": false
+  },
+  "backends": {
+    "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": 0.6436757920018863,
+      "warm_inference_latency_ms": {
+        "count": 196,
+        "median_ms": 0.9285830019507557,
+        "p95_ms": 0.9653437373344786,
+        "min_ms": 0.8930829935707152,
+        "max_ms": 1.0020830086432397
+      },
+      "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": {
+    "ane_gather.py": "c3379a3797041dcfed82eaaea3990771ec91f1dff7ec146b025988ad7a8c9aae",
+    "assets.py": "f7a5bc931075237e00c73bf21fee3575be35531f934f23b5e22627d50a79f92f",
+    "benchmark-synthetic.py": "ba8255fd3f63d87ef340c26c4c251800f2e20878d80858ad1e6557cf59f249a2",
+    "compare-ane.py": "4633df7ad5f90d6aed37d5ef4f7a2d2231fdadd272ef1aa3057c237b2f937df8",
+    "convert-coreml.py": "46fe0194df924033126286e601214b6789f5d667aae38c153cf4dbc1b0e629b0",
+    "export-reference.py": "ff9eab4d6e4e72ffaf074d82610f41fbea1e802da7c0e987ce562a89eb561b99",
+    "export_model.py": "c61da5d678c77030d1e9a78359626063ea31da46fb6434e49517fdeaa7800949",
+    "prepare-int4-source.py": "33db3295450428d90c1de5a804ab47bf3507c686fbbdef67cd276ce116758b5f",
+    "preprocessing.py": "18c74ede43f95f91cd638c7e9631923a353e3a24d3b522ddb8b18b480f9c2846",
+    "profile-coreml.py": "d5e1386f54fc6c482647120e593e7ffbbe4bbb277c2cc6d4949154660fe239fd",
+    "quantize-int4.py": "a91636e539d8934253b5e7628ca706346fbee538fa1aa9e3eaf794f440d49dcc",
+    "quantize-int8.py": "22071babfcb02ef77225203a479d9bb14b9fa35462ba38bcc4bf07d29c125a97",
+    "score-report.py": "c78e45caa86fb59f9a56ef18cc6a35020ea6e281a6a23d23618f56da85a0d95e",
+    "synthetic_test.py": "e2b68725236e6827eb422480c8cc8aa189b2490c1594650d8f2928771d1bc626",
+    "verify.py": "bc20ec9df47708e9b48e411076abe4eab7c18313f67e11334de2d9e3ecb3efa7"
+  }
+}