Upload 101 files

README.md

@@ -0,0 +1,212 @@
+---
+license: other
+license_name: yl4579-styletts2
+license_link: https://github.com/yl4579/StyleTTS2#pre-requisites
+language:
+- en
+library_name: coreml
+tags:
+- text-to-speech
+- styletts2
+- coreml
+- apple-silicon
+- libritts
+- on-device
+pipeline_tag: text-to-speech
+inference: false
+---
+
+# StyleTTS2 (LibriTTS) — CoreML
+
+Apple-Silicon-optimized CoreML conversion of [yl4579/StyleTTS2](https://github.com/yl4579/StyleTTS2)
+LibriTTS multi-speaker checkpoint
+([`yl4579/StyleTTS2-LibriTTS` → `Models/LibriTTS/epochs_2nd_00020.pth`](https://huggingface.co/yl4579/StyleTTS2-LibriTTS)).
+
+Four-stage pipeline; per-stage compute-unit placement; selective int8 PTQ on
+the text-and-prosody predictor; fp32 decoder.
+
+> [!IMPORTANT]
+> **These weights carry use restrictions beyond MIT. Read the License
+> section before downloading.** They are not a drop-in permissively-licensed
+> TTS model. If you need permissive terms, use
+> [Kokoro](https://huggingface.co/hexgrad/Kokoro-82M) instead.
+
+## License & use restrictions
+
+The upstream repository code is MIT, but the pre-trained LibriTTS weights
+carry **two non-negotiable restrictions** declared in
+[yl4579/StyleTTS2's README](https://github.com/yl4579/StyleTTS2#pre-requisites):
+
+1. **Synthetic-origin disclosure.** Any deployment that produces audio from
+   these weights must clearly disclose to listeners that the audio is
+   synthetic. No undisclosed synthetic-speech publishing.
+2. **Speaker consent for voice cloning.** Cloning a real person's voice
+   requires their consent. No unauthorized celebrity / public-figure /
+   non-consenting third-party voice cloning.
+
+These restrictions ride with the weights through every redistribution,
+fine-tune, and downstream derivative. Anyone downloading this repo inherits
+them and must propagate them in turn.
+
+If you cannot or will not honor these terms, **do not download these
+weights**.
+
+License-of-record: [github.com/yl4579/StyleTTS2](https://github.com/yl4579/StyleTTS2)
+upstream README at the time of conversion (see *Conversion provenance* below
+for the pinned commit).
+
+## What's in this repo
+
+| Package | Compute unit | Precision | Buckets | Called |
+|---|---|---|---|---|
+| `styletts2_text_predictor_{32,64,128,256,512}.mlpackage` | ANE | fp16 | 5 token-length | 1× per utterance |
+| `styletts2_diffusion_step_512.mlpackage` | CPU+GPU | fp16 | 1 (B=512 only) | ~5× per utterance |
+| `styletts2_f0n_energy.mlpackage` | ANE | fp16 | dynamic | 1× per utterance |
+| `styletts2_decoder_{256,512,1024,2048,4096}.mlpackage` | CPU+GPU | **fp32** | 5 mel-length | 1× per utterance |
+| `constants/text_cleaner_vocab.json` | — | — | — | phoneme→id table |
+| `config.json` | — | — | — | bundle runtime contract (audio/sampler/buckets) |
+
+Total on-disk size: ~1.4 GB per format.
+
+Both source `.mlpackage` (uncompiled, portable across Xcode versions) and
+pre-compiled `.mlmodelc` (Apple Silicon, ready for `MLModel(contentsOf:)`)
+are shipped. The `.mlmodelc` artifacts are under `compiled/`. Pick one:
+
+- **`*.mlpackage`** — load via `MLModel(contentsOf:)`; the OS compiles on
+  first load (~5–20 s cold start the first time, cached afterward).
+- **`compiled/*.mlmodelc`** — already compiled; same loader path skips the
+  on-device compile. Useful for shipping inside an app bundle.
+
+The diffusion sampler loop (ADPM2 + Karras schedule + CFG) and the
+hard-alignment matrix (cumsum-of-durations → one-hot → matmul) live in your
+host application (Swift / Python). Per-step inference is in CoreML; control
+flow is not.
+
+### Why the precision split looks like this
+
+- **text_predictor is fp16.** Selective int8 PTQ was tried and dropped:
+  on Apple Silicon ANE the int8 path saves only ~3 MB per bucket of
+  weight bandwidth, has no exposed int8 GEMM, and dequantizes back to
+  fp16 on load. The savings did not justify the parity risk on small
+  projections.
+- **diffusion_step stays fp16.** It runs 5 times per utterance through an
+  ODE-style sampler; quantization noise compounds through iterations.
+  Same lesson as PocketTTS issue #7.
+- **f0n_energy stays fp16.** ~6 MB. No bandwidth payoff; quantizing
+  small projections injects audible pitch noise.
+- **decoder is fp32, not fp16.** SineGen's harmonic source accumulates
+  phase via `cumsum × 2π × hop=300`, reaching magnitudes ~4000
+  mid-frame. fp16 precision at that magnitude (~4) is much larger than
+  the per-sample increment (~0.05 rad), which scrambles the sine output
+  and produces audibly robotic synthesis. fp32 is required end-to-end.
+
+### Why only one diffusion bucket
+
+Empirically every observed `bert_dur` fits in B=512. The 32/64/128/256
+buckets were dead weight (~192 MB) given the non-linear cost ladder
+(B=32 ≈ 66 ms/step, B=512 ≈ 152 ms/step). Dropping them adds at most
+~430 ms per utterance in the worst short case.
+
+## Performance
+
+- **RTFx:** 4.32× warm on M-series Mac (5-step ADPM2 sampler, all buckets
+  pre-warmed).
+- **Log-mel cosine vs PyTorch fp32:** 0.9687.
+- **ECAPA-TDNN cosine to reference clip:** 0.18 — at the model's
+  architectural ceiling. PyTorch fp32 itself only reaches 0.29 on the
+  same metric. Voice-clone fidelity is bounded by StyleTTS2's
+  architecture, not by this conversion.
+
+## How to use
+
+### Phonemizer
+
+espeak-ng IPA + stress. The 178-token vocabulary in
+`constants/text_cleaner_vocab.json` mirrors `text_utils.TextCleaner` from
+the upstream repo: `[pad] + punctuation + ASCII letters + IPA letters`.
+
+Pad token is `$` at id 0.
+
+### Inference shape
+
+```text
+text → phonemes → token ids
+                     │
+                     ▼
+text_predictor (ANE, int8)
+   │   ├─ d_en (1, T_dur, hidden)
+   │   ├─ s_pred (1, 256)             (sampler init via diffusion)
+   │   └─ duration logits → duration → one-hot alignment matrix (host)
+   │
+   ▼
+diffusion_step × 5  (CPU+GPU, fp16)   (ADPM2 + Karras schedule + CFG)
+   │
+   ▼
+[blend(s, ref_s) + alignment]
+   │
+   ▼
+f0n_energy (ANE, fp16) → F0_curve, N
+   │
+   ▼
+decoder (CPU+GPU, fp32) → 24 kHz waveform
+```
+
+The Swift host owns the sampler loop, alignment construction, and bucket
+routing. A reference Swift integration is in
+[FluidInference/FluidAudio](https://github.com/FluidInference/FluidAudio).
+
+### Bucket routing
+
+Round each variable-length input up to the next bucket. Pad with zeros.
+
+| Input | Axis | Buckets |
+|---|---|---|
+| text_predictor `tokens` | T_tok | 32 / 64 / 128 / 256 / 512 |
+| diffusion_step `embedding` | T_bert | 512 only (pad) |
+| decoder `asr` | T_mel | 256 / 512 / 1024 / 2048 / 4096 |
+
+f0n_energy is shape-flexible.
+
+## Conversion provenance
+
+- **Upstream code:** [yl4579/StyleTTS2](https://github.com/yl4579/StyleTTS2)
+- **Upstream weights:** [yl4579/StyleTTS2-LibriTTS](https://huggingface.co/yl4579/StyleTTS2-LibriTTS),
+  file `Models/LibriTTS/epochs_2nd_00020.pth`
+- **Conversion scripts:** [FluidInference/mobius PR #46](https://github.com/FluidInference/mobius/pull/46)
+  (`models/tts/styletts2/scripts/`)
+- **Quantization:** `coremltools.optimize.coreml.linear_quantize_weights`,
+  `mode=linear_symmetric`, `dtype=int8`, `granularity=per_channel`,
+  `weight_threshold=200_000`
+- **Target:** `coremltools` ≥ 8.0, `minimum_deployment_target=iOS17`
+  (macOS 14+ / iOS 17+)
+
+## Known limitations
+
+- **English (LibriTTS) only.** No multilingual support in this
+  checkpoint.
+- **HiFi-GAN decoder, not iSTFTNet.** LibriTTS upstream uses HiFi-GAN, so
+  no `torch.stft` / complex tensors in the conversion path.
+- **Decoder is fp32, not fp16.** Documented above. The mlpackage size
+  reflects this (≈210 MB per bucket).
+- **Voice-clone fidelity ceiling is architectural.** ECAPA-TDNN cosine
+  to reference clip ≈ 0.18 here, ≈ 0.29 in PyTorch fp32. The same-speaker
+  threshold is ~0.30. This isn't a quantization or conversion artifact;
+  see PR #46 TRIALS.md Phase 5.
+- **No streaming.** Whole utterance only. Add chunked streaming on the
+  host side if you need it.
+
+## Citation & acknowledgments
+
+- Yinghao Aaron Li et al. — StyleTTS2 architecture and LibriTTS
+  checkpoint.
+- LibriTTS authors (CC-BY-4.0 training data).
+- espeak-ng — phonemization frontend.
+
+```bibtex
+@inproceedings{li2023styletts2,
+  title  = {StyleTTS 2: Towards Human-Level Text-to-Speech through Style Diffusion and Adversarial Training with Large Speech Language Models},
+  author = {Li, Yinghao Aaron and Han, Cong and Raghavan, Vinay and Mischler, Gavin and Mesgarani, Nima},
+  booktitle = {NeurIPS},
+  year   = {2023}
+}
+```

compiled/styletts2_decoder_1024.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3c8d4484ec862aee189fe8f7b84fd9e74ad87af7f903b5e3022491f0819b6242
+size 243

compiled/styletts2_decoder_1024.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:176b03f7aeb6643777e10d07e26748bc532eb92756ee547a1756b72bdd7563bf
+size 447

compiled/styletts2_decoder_1024.mlmodelc/metadata.json

@@ -0,0 +1,114 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "StyleTTS2 HiFi-GAN decoder (LibriTTS) @ T_mel=1024. 24 kHz, hop=300.",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 614400)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 614400]",
+        "name" : "waveform",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Float32",
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios17.greater" : 1,
+      "UpsampleNearestNeighbor" : 2,
+      "Ios17.floor" : 1,
+      "Ios17.instanceNorm" : 98,
+      "Ios17.reshape" : 110,
+      "Ios17.transpose" : 8,
+      "Ios17.sin" : 103,
+      "Split" : 106,
+      "Ios17.convTranspose" : 5,
+      "Ios17.expandDims" : 4,
+      "Ios17.add" : 381,
+      "Tile" : 2,
+      "Ios17.sliceByIndex" : 6,
+      "Ios17.squeeze" : 2,
+      "Ios17.conv" : 119,
+      "Ios17.sub" : 2,
+      "Ios17.cast" : 1,
+      "Ios17.tanh" : 2,
+      "Ios17.pow" : 101,
+      "Ios17.linear" : 107,
+      "Ios17.leakyRelu" : 10,
+      "Ios17.concat" : 6,
+      "Ios16.cumsum" : 1,
+      "Ios17.mul" : 331
+    },
+    "computePrecision" : "Mixed (Float32, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 512 × 1024)",
+        "shortDescription" : "",
+        "shape" : "[1, 512, 1024]",
+        "name" : "asr",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 2048)",
+        "shortDescription" : "",
+        "shape" : "[1, 2048]",
+        "name" : "F0_curve",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 2048)",
+        "shortDescription" : "",
+        "shape" : "[1, 2048]",
+        "name" : "N",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 128]",
+        "name" : "s",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.version" : "8.3.0",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript",
+      "com.github.apple.coremltools.source" : "torch==2.11.0"
+    },
+    "generatedClassName" : "styletts2_decoder_1024",
+    "method" : "predict"
+  }
+]

compiled/styletts2_decoder_1024.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ceacbdc62672599460d67312ac7ad25ea1cbf62ec5f251d54392b191f3003853
+size 222055552

compiled/styletts2_decoder_2048.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2f876eefb7dca5fd52ad07e040486cb103e7e926ffcdb23c114ca6c41493738e
+size 243

compiled/styletts2_decoder_2048.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1078e4402384405983e7bafa9f6016d816fcaa846dac7deba8dbe8275a950884
+size 447

compiled/styletts2_decoder_2048.mlmodelc/metadata.json

@@ -0,0 +1,114 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "StyleTTS2 HiFi-GAN decoder (LibriTTS) @ T_mel=2048. 24 kHz, hop=300.",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 1228800)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 1228800]",
+        "name" : "waveform",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Float32",
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios17.greater" : 1,
+      "UpsampleNearestNeighbor" : 2,
+      "Ios17.floor" : 1,
+      "Ios17.instanceNorm" : 98,
+      "Ios17.reshape" : 110,
+      "Ios17.transpose" : 8,
+      "Ios17.sin" : 103,
+      "Split" : 106,
+      "Ios17.convTranspose" : 5,
+      "Ios17.expandDims" : 4,
+      "Ios17.add" : 381,
+      "Tile" : 2,
+      "Ios17.sliceByIndex" : 6,
+      "Ios17.squeeze" : 2,
+      "Ios17.conv" : 119,
+      "Ios17.sub" : 2,
+      "Ios17.cast" : 1,
+      "Ios17.tanh" : 2,
+      "Ios17.pow" : 101,
+      "Ios17.linear" : 107,
+      "Ios17.leakyRelu" : 10,
+      "Ios17.concat" : 6,
+      "Ios16.cumsum" : 1,
+      "Ios17.mul" : 331
+    },
+    "computePrecision" : "Mixed (Float32, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 512 × 2048)",
+        "shortDescription" : "",
+        "shape" : "[1, 512, 2048]",
+        "name" : "asr",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 4096)",
+        "shortDescription" : "",
+        "shape" : "[1, 4096]",
+        "name" : "F0_curve",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 4096)",
+        "shortDescription" : "",
+        "shape" : "[1, 4096]",
+        "name" : "N",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 128]",
+        "name" : "s",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.version" : "8.3.0",
+      "com.github.apple.coremltools.source" : "torch==2.11.0",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "generatedClassName" : "styletts2_decoder_2048",
+    "method" : "predict"
+  }
+]

compiled/styletts2_decoder_2048.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d099eb64c9b697e596e97f0c54077e104da9d690c302576ecb624083377b75eb
+size 226970752

compiled/styletts2_decoder_256.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6c2984630321f205643a9417db8ed42db88d0ec2e038daa0402543a1d01aec02
+size 243

compiled/styletts2_decoder_256.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:822f4f79fcf843f3e6d1674a9a6870f62fca1a4c1daa3f4297d6c52a01c99809
+size 446

compiled/styletts2_decoder_256.mlmodelc/metadata.json

@@ -0,0 +1,114 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "StyleTTS2 HiFi-GAN decoder (LibriTTS) @ T_mel=256. 24 kHz, hop=300.",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 153600)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 153600]",
+        "name" : "waveform",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Float32",
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios17.greater" : 1,
+      "UpsampleNearestNeighbor" : 2,
+      "Ios17.floor" : 1,
+      "Ios17.instanceNorm" : 98,
+      "Ios17.reshape" : 110,
+      "Ios17.transpose" : 8,
+      "Ios17.sin" : 103,
+      "Split" : 106,
+      "Ios17.convTranspose" : 5,
+      "Ios17.expandDims" : 4,
+      "Ios17.add" : 381,
+      "Tile" : 2,
+      "Ios17.sliceByIndex" : 6,
+      "Ios17.squeeze" : 2,
+      "Ios17.conv" : 119,
+      "Ios17.sub" : 2,
+      "Ios17.cast" : 1,
+      "Ios17.tanh" : 2,
+      "Ios17.pow" : 101,
+      "Ios17.linear" : 107,
+      "Ios17.leakyRelu" : 10,
+      "Ios17.concat" : 6,
+      "Ios16.cumsum" : 1,
+      "Ios17.mul" : 331
+    },
+    "computePrecision" : "Mixed (Float32, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 512 × 256)",
+        "shortDescription" : "",
+        "shape" : "[1, 512, 256]",
+        "name" : "asr",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 512)",
+        "shortDescription" : "",
+        "shape" : "[1, 512]",
+        "name" : "F0_curve",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 512)",
+        "shortDescription" : "",
+        "shape" : "[1, 512]",
+        "name" : "N",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 128]",
+        "name" : "s",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.version" : "8.3.0",
+      "com.github.apple.coremltools.source" : "torch==2.11.0",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "generatedClassName" : "styletts2_decoder_256",
+    "method" : "predict"
+  }
+]

compiled/styletts2_decoder_256.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c21c31ce508196d7314e00dfac02b545d09a1961b529cbbfe98c77d0109b003b
+size 218369152

compiled/styletts2_decoder_4096.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d33d6ef2159363ecb4586b70214f62159fc6871fe92ff7060fdd3614811e6abe
+size 243

compiled/styletts2_decoder_4096.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:20c4e04c0d1558b8a249986d2553654a434646c06a2d565302b360d1aebe0a24
+size 448

compiled/styletts2_decoder_4096.mlmodelc/metadata.json

@@ -0,0 +1,114 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "StyleTTS2 HiFi-GAN decoder (LibriTTS) @ T_mel=4096. 24 kHz, hop=300.",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 2457600)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 2457600]",
+        "name" : "waveform",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Float32",
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios17.greater" : 1,
+      "UpsampleNearestNeighbor" : 2,
+      "Ios17.floor" : 1,
+      "Ios17.instanceNorm" : 98,
+      "Ios17.reshape" : 110,
+      "Ios17.transpose" : 8,
+      "Ios17.sin" : 103,
+      "Split" : 106,
+      "Ios17.convTranspose" : 5,
+      "Ios17.expandDims" : 4,
+      "Ios17.add" : 381,
+      "Tile" : 2,
+      "Ios17.sliceByIndex" : 6,
+      "Ios17.squeeze" : 2,
+      "Ios17.conv" : 119,
+      "Ios17.sub" : 2,
+      "Ios17.cast" : 1,
+      "Ios17.tanh" : 2,
+      "Ios17.pow" : 101,
+      "Ios17.linear" : 107,
+      "Ios17.leakyRelu" : 10,
+      "Ios17.concat" : 6,
+      "Ios16.cumsum" : 1,
+      "Ios17.mul" : 331
+    },
+    "computePrecision" : "Mixed (Float32, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 512 × 4096)",
+        "shortDescription" : "",
+        "shape" : "[1, 512, 4096]",
+        "name" : "asr",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 8192)",
+        "shortDescription" : "",
+        "shape" : "[1, 8192]",
+        "name" : "F0_curve",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 8192)",
+        "shortDescription" : "",
+        "shape" : "[1, 8192]",
+        "name" : "N",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 128]",
+        "name" : "s",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.version" : "8.3.0",
+      "com.github.apple.coremltools.source" : "torch==2.11.0",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "generatedClassName" : "styletts2_decoder_4096",
+    "method" : "predict"
+  }
+]

compiled/styletts2_decoder_4096.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5e2825d97b280217d84f597c085dc5c33328c736fb5b5d63f288209ae612b4f7
+size 236801152

compiled/styletts2_decoder_512.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e42c6710d2a1441e67bef87fda6ffb6af6f1feb6b06f1585a71b99086b550a0e
+size 243

compiled/styletts2_decoder_512.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6abfad25a5df29fdcc4e135bcbd614a9480355904e58b24a33afae72c03bcc9a
+size 446

compiled/styletts2_decoder_512.mlmodelc/metadata.json

@@ -0,0 +1,114 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "StyleTTS2 HiFi-GAN decoder (LibriTTS) @ T_mel=512. 24 kHz, hop=300.",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 307200)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 307200]",
+        "name" : "waveform",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Float32",
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios17.greater" : 1,
+      "UpsampleNearestNeighbor" : 2,
+      "Ios17.floor" : 1,
+      "Ios17.instanceNorm" : 98,
+      "Ios17.reshape" : 110,
+      "Ios17.transpose" : 8,
+      "Ios17.sin" : 103,
+      "Split" : 106,
+      "Ios17.convTranspose" : 5,
+      "Ios17.expandDims" : 4,
+      "Ios17.add" : 381,
+      "Tile" : 2,
+      "Ios17.sliceByIndex" : 6,
+      "Ios17.squeeze" : 2,
+      "Ios17.conv" : 119,
+      "Ios17.sub" : 2,
+      "Ios17.cast" : 1,
+      "Ios17.tanh" : 2,
+      "Ios17.pow" : 101,
+      "Ios17.linear" : 107,
+      "Ios17.leakyRelu" : 10,
+      "Ios17.concat" : 6,
+      "Ios16.cumsum" : 1,
+      "Ios17.mul" : 331
+    },
+    "computePrecision" : "Mixed (Float32, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 512 × 512)",
+        "shortDescription" : "",
+        "shape" : "[1, 512, 512]",
+        "name" : "asr",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1024)",
+        "shortDescription" : "",
+        "shape" : "[1, 1024]",
+        "name" : "F0_curve",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1024)",
+        "shortDescription" : "",
+        "shape" : "[1, 1024]",
+        "name" : "N",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 128]",
+        "name" : "s",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.version" : "8.3.0",
+      "com.github.apple.coremltools.source" : "torch==2.11.0",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "generatedClassName" : "styletts2_decoder_512",
+    "method" : "predict"
+  }
+]

compiled/styletts2_decoder_512.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5eaca30dac2dce98bb9341edf9739915b3f3e73ddfaa907ae523c89946dc2964
+size 219597952

compiled/styletts2_diffusion_step_512.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bd74ef01b9d41584bb2f2da6e5957336afff35a91f1772f83e1e8ff757928fea
+size 243

compiled/styletts2_diffusion_step_512.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:686dbb5576a97eb595764e9bcb9e46aa2d560c67502a4ddf1afc683886bd1e53
+size 467

compiled/styletts2_diffusion_step_512.mlmodelc/metadata.json

@@ -0,0 +1,114 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "StyleTTS2 KDiffusion single denoise step (StyleTransformer1d) @ T_tok=512.",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 1 × 256)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 256]",
+        "name" : "denoised",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Float16",
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios17.cos" : 1,
+      "Ios17.reshape" : 20,
+      "Ios17.inverse" : 2,
+      "Ios16.softmax" : 3,
+      "Ios17.matmul" : 6,
+      "Ios17.sin" : 1,
+      "Ios17.transpose" : 27,
+      "Split" : 9,
+      "Ios16.reduceMean" : 1,
+      "Ios17.expandDims" : 2,
+      "Ios17.add" : 23,
+      "Tile" : 2,
+      "Ios16.reduceSum" : 1,
+      "Ios17.layerNorm" : 3,
+      "Ios17.log" : 1,
+      "Ios17.sqrt" : 1,
+      "Ios16.gelu" : 7,
+      "Ios17.cast" : 4,
+      "Ios17.conv" : 1,
+      "Ios17.realDiv" : 1,
+      "Ios17.linear" : 25,
+      "Stack" : 1,
+      "Ios17.concat" : 2,
+      "Ios17.mul" : 19
+    },
+    "computePrecision" : "Mixed (Float16, Float32, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 256)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 256]",
+        "name" : "x_noisy",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "sigma",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 512 × 768)",
+        "shortDescription" : "",
+        "shape" : "[1, 512, 768]",
+        "name" : "embedding",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 256)",
+        "shortDescription" : "",
+        "shape" : "[1, 256]",
+        "name" : "features",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.version" : "8.3.0",
+      "com.github.apple.coremltools.source" : "torch==2.11.0",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "generatedClassName" : "styletts2_diffusion_step_512",
+    "method" : "predict"
+  }
+]

compiled/styletts2_diffusion_step_512.mlmodelc/model.mil

@@ -0,0 +1,351 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.11.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios17>(tensor<fp32, [1, 512, 768]> embedding, tensor<fp32, [1, 256]> features, tensor<fp32, [1]> sigma, tensor<fp32, [1, 1, 256]> x_noisy) {
+            tensor<int32, [3]> var_25 = const()[name = tensor<string, []>("op_25"), val = tensor<int32, [3]>([-1, 1, 1])];
+            tensor<string, []> sigma_to_fp16_dtype_0 = const()[name = tensor<string, []>("sigma_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1]> sigma_to_fp16 = cast(dtype = sigma_to_fp16_dtype_0, x = sigma)[name = tensor<string, []>("cast_27")];
+            tensor<fp16, [1, 1, 1]> s_cast_fp16 = reshape(shape = var_25, x = sigma_to_fp16)[name = tensor<string, []>("s_cast_fp16")];
+            tensor<fp16, [1, 1, 1]> var_27_cast_fp16 = mul(x = s_cast_fp16, y = s_cast_fp16)[name = tensor<string, []>("op_27_cast_fp16")];
+            tensor<fp16, []> var_29_to_fp16 = const()[name = tensor<string, []>("op_29_to_fp16"), val = tensor<fp16, []>(0x1.47cp-5)];
+            tensor<fp16, [1, 1, 1]> var_30_cast_fp16 = add(x = var_27_cast_fp16, y = var_29_to_fp16)[name = tensor<string, []>("op_30_cast_fp16")];
+            tensor<fp32, []> var_31_epsilon_0 = const()[name = tensor<string, []>("op_31_epsilon_0"), val = tensor<fp32, []>(0x1.a36e2ep-14)];
+            tensor<fp16, [1, 1, 1]> var_31_cast_fp16 = inverse(epsilon = var_31_epsilon_0, x = var_30_cast_fp16)[name = tensor<string, []>("op_31_cast_fp16")];
+            tensor<fp16, []> var_32_to_fp16 = const()[name = tensor<string, []>("op_32_to_fp16"), val = tensor<fp16, []>(0x1.47cp-5)];
+            tensor<fp16, [1, 1, 1]> c_skip_cast_fp16 = mul(x = var_31_cast_fp16, y = var_32_to_fp16)[name = tensor<string, []>("c_skip_cast_fp16")];
+            tensor<fp16, []> var_34_to_fp16 = const()[name = tensor<string, []>("op_34_to_fp16"), val = tensor<fp16, []>(0x1.998p-3)];
+            tensor<fp16, [1, 1, 1]> var_35_cast_fp16 = mul(x = s_cast_fp16, y = var_34_to_fp16)[name = tensor<string, []>("op_35_cast_fp16")];
+            tensor<fp16, [1, 1, 1]> var_40_cast_fp16 = sqrt(x = var_30_cast_fp16)[name = tensor<string, []>("op_40_cast_fp16")];
+            tensor<fp16, [1, 1, 1]> c_out_cast_fp16 = real_div(x = var_35_cast_fp16, y = var_40_cast_fp16)[name = tensor<string, []>("c_out_cast_fp16")];
+            tensor<fp32, []> var_47_epsilon_0 = const()[name = tensor<string, []>("op_47_epsilon_0"), val = tensor<fp32, []>(0x1.a36e2ep-14)];
+            tensor<fp16, [1, 1, 1]> var_47_cast_fp16 = inverse(epsilon = var_47_epsilon_0, x = var_40_cast_fp16)[name = tensor<string, []>("op_47_cast_fp16")];
+            tensor<fp32, []> var_50_epsilon_0 = const()[name = tensor<string, []>("op_50_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
+            tensor<fp16, [1]> var_50_cast_fp16 = log(epsilon = var_50_epsilon_0, x = sigma_to_fp16)[name = tensor<string, []>("op_50_cast_fp16")];
+            tensor<fp16, []> var_51_to_fp16 = const()[name = tensor<string, []>("op_51_to_fp16"), val = tensor<fp16, []>(0x1p-2)];
+            tensor<fp16, [1]> x_1_cast_fp16 = mul(x = var_50_cast_fp16, y = var_51_to_fp16)[name = tensor<string, []>("x_1_cast_fp16")];
+            tensor<string, []> x_noisy_to_fp16_dtype_0 = const()[name = tensor<string, []>("x_noisy_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1, 1, 256]> x_noisy_to_fp16 = cast(dtype = x_noisy_to_fp16_dtype_0, x = x_noisy)[name = tensor<string, []>("cast_26")];
+            tensor<fp16, [1, 1, 256]> x_11_cast_fp16 = mul(x = var_47_cast_fp16, y = x_noisy_to_fp16)[name = tensor<string, []>("x_11_cast_fp16")];
+            tensor<int32, []> var_55 = const()[name = tensor<string, []>("op_55"), val = tensor<int32, []>(-1)];
+            tensor<int32, [2]> var_67 = const()[name = tensor<string, []>("op_67"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [1, 1]> x_5_cast_fp16 = reshape(shape = var_67, x = x_1_cast_fp16)[name = tensor<string, []>("x_5_cast_fp16")];
+            tensor<fp16, [1, 128]> var_75_to_fp16 = const()[name = tensor<string, []>("op_75_to_fp16"), val = tensor<fp16, [1, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<fp16, [1, 128]> var_76_cast_fp16 = mul(x = x_5_cast_fp16, y = var_75_to_fp16)[name = tensor<string, []>("op_76_cast_fp16")];
+            tensor<fp16, []> var_77_promoted_to_fp16 = const()[name = tensor<string, []>("op_77_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
+            tensor<fp16, [1, 128]> var_78_cast_fp16 = mul(x = var_76_cast_fp16, y = var_77_promoted_to_fp16)[name = tensor<string, []>("op_78_cast_fp16")];
+            tensor<fp16, []> var_79_to_fp16 = const()[name = tensor<string, []>("op_79_to_fp16"), val = tensor<fp16, []>(0x1.92p+1)];
+            tensor<fp16, [1, 128]> freqs_cast_fp16 = mul(x = var_78_cast_fp16, y = var_79_to_fp16)[name = tensor<string, []>("freqs_cast_fp16")];
+            tensor<fp16, [1, 128]> var_81_cast_fp16 = sin(x = freqs_cast_fp16)[name = tensor<string, []>("op_81_cast_fp16")];
+            tensor<fp16, [1, 128]> var_82_cast_fp16 = cos(x = freqs_cast_fp16)[name = tensor<string, []>("op_82_cast_fp16")];
+            tensor<bool, []> input_1_interleave_0 = const()[name = tensor<string, []>("input_1_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 257]> input_1_cast_fp16 = concat(axis = var_55, interleave = input_1_interleave_0, values = (x_5_cast_fp16, var_81_cast_fp16, var_82_cast_fp16))[name = tensor<string, []>("input_1_cast_fp16")];
+            tensor<fp16, [1024, 257]> transformer_to_time_0_1_weight_to_fp16 = const()[name = tensor<string, []>("transformer_to_time_0_1_weight_to_fp16"), val = tensor<fp16, [1024, 257]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(384)))];
+            tensor<fp16, [1024]> transformer_to_time_0_1_bias_to_fp16 = const()[name = tensor<string, []>("transformer_to_time_0_1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(526784)))];
+            tensor<fp16, [1, 1024]> linear_0_cast_fp16 = linear(bias = transformer_to_time_0_1_bias_to_fp16, weight = transformer_to_time_0_1_weight_to_fp16, x = input_1_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")];
+            tensor<string, []> var_88_mode_0 = const()[name = tensor<string, []>("op_88_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 1024]> var_88_cast_fp16 = gelu(mode = var_88_mode_0, x = linear_0_cast_fp16)[name = tensor<string, []>("op_88_cast_fp16")];
+            tensor<string, []> features_to_fp16_dtype_0 = const()[name = tensor<string, []>("features_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1024, 256]> transformer_to_features_0_weight_to_fp16 = const()[name = tensor<string, []>("transformer_to_features_0_weight_to_fp16"), val = tensor<fp16, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(528896)))];
+            tensor<fp16, [1024]> transformer_to_features_0_bias_to_fp16 = const()[name = tensor<string, []>("transformer_to_features_0_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1053248)))];
+            tensor<fp16, [1, 256]> features_to_fp16 = cast(dtype = features_to_fp16_dtype_0, x = features)[name = tensor<string, []>("cast_25")];
+            tensor<fp16, [1, 1024]> linear_1_cast_fp16 = linear(bias = transformer_to_features_0_bias_to_fp16, weight = transformer_to_features_0_weight_to_fp16, x = features_to_fp16)[name = tensor<string, []>("linear_1_cast_fp16")];
+            tensor<string, []> var_94_mode_0 = const()[name = tensor<string, []>("op_94_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 1024]> var_94_cast_fp16 = gelu(mode = var_94_mode_0, x = linear_1_cast_fp16)[name = tensor<string, []>("op_94_cast_fp16")];
+            tensor<int32, []> x_7_axis_0 = const()[name = tensor<string, []>("x_7_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp16, [2, 1, 1024]> x_7_cast_fp16 = stack(axis = x_7_axis_0, values = (var_88_cast_fp16, var_94_cast_fp16))[name = tensor<string, []>("x_7_cast_fp16")];
+            tensor<int32, [3]> var_101 = const()[name = tensor<string, []>("op_101"), val = tensor<int32, [3]>([1, 2, 0])];
+            tensor<int32, [1]> input_7_axes_0 = const()[name = tensor<string, []>("input_7_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<bool, []> input_7_keep_dims_0 = const()[name = tensor<string, []>("input_7_keep_dims_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024, 2]> x_9_cast_fp16 = transpose(perm = var_101, x = x_7_cast_fp16)[name = tensor<string, []>("transpose_41")];
+            tensor<fp16, [1, 1024]> input_7_cast_fp16 = reduce_sum(axes = input_7_axes_0, keep_dims = input_7_keep_dims_0, x = x_9_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
+            tensor<fp16, [1024, 1024]> transformer_to_mapping_0_weight_to_fp16 = const()[name = tensor<string, []>("transformer_to_mapping_0_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1055360)))];
+            tensor<fp16, [1024]> transformer_to_mapping_0_bias_to_fp16 = const()[name = tensor<string, []>("transformer_to_mapping_0_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3152576)))];
+            tensor<fp16, [1, 1024]> linear_2_cast_fp16 = linear(bias = transformer_to_mapping_0_bias_to_fp16, weight = transformer_to_mapping_0_weight_to_fp16, x = input_7_cast_fp16)[name = tensor<string, []>("linear_2_cast_fp16")];
+            tensor<string, []> input_11_mode_0 = const()[name = tensor<string, []>("input_11_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 1024]> input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = linear_2_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")];
+            tensor<fp16, [1024, 1024]> transformer_to_mapping_2_weight_to_fp16 = const()[name = tensor<string, []>("transformer_to_mapping_2_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3154688)))];
+            tensor<fp16, [1024]> transformer_to_mapping_2_bias_to_fp16 = const()[name = tensor<string, []>("transformer_to_mapping_2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5251904)))];
+            tensor<fp16, [1, 1024]> linear_3_cast_fp16 = linear(bias = transformer_to_mapping_2_bias_to_fp16, weight = transformer_to_mapping_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor<string, []>("linear_3_cast_fp16")];
+            tensor<string, []> mapping_1_mode_0 = const()[name = tensor<string, []>("mapping_1_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 1024]> mapping_1_cast_fp16 = gelu(mode = mapping_1_mode_0, x = linear_3_cast_fp16)[name = tensor<string, []>("mapping_1_cast_fp16")];
+            tensor<int32, [3]> var_127_reps_0 = const()[name = tensor<string, []>("op_127_reps_0"), val = tensor<int32, [3]>([1, 512, 1])];
+            tensor<fp16, [1, 512, 256]> var_127_cast_fp16 = tile(reps = var_127_reps_0, x = x_11_cast_fp16)[name = tensor<string, []>("op_127_cast_fp16")];
+            tensor<int32, []> var_129 = const()[name = tensor<string, []>("op_129"), val = tensor<int32, []>(-1)];
+            tensor<bool, []> x_13_interleave_0 = const()[name = tensor<string, []>("x_13_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<string, []> embedding_to_fp16_dtype_0 = const()[name = tensor<string, []>("embedding_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1, 512, 768]> embedding_to_fp16 = cast(dtype = embedding_to_fp16_dtype_0, x = embedding)[name = tensor<string, []>("cast_24")];
+            tensor<fp16, [1, 512, 1024]> x_13_cast_fp16 = concat(axis = var_129, interleave = x_13_interleave_0, values = (var_127_cast_fp16, embedding_to_fp16))[name = tensor<string, []>("x_13_cast_fp16")];
+            tensor<int32, [1]> var_132_axes_0 = const()[name = tensor<string, []>("op_132_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1024]> var_132_cast_fp16 = expand_dims(axes = var_132_axes_0, x = mapping_1_cast_fp16)[name = tensor<string, []>("op_132_cast_fp16")];
+            tensor<int32, [3]> mapping_reps_0 = const()[name = tensor<string, []>("mapping_reps_0"), val = tensor<int32, [3]>([1, 512, 1])];
+            tensor<fp16, [1, 512, 1024]> mapping_cast_fp16 = tile(reps = mapping_reps_0, x = var_132_cast_fp16)[name = tensor<string, []>("mapping_cast_fp16")];
+            tensor<fp16, [1, 512, 1024]> x_15_cast_fp16 = add(x = x_13_cast_fp16, y = mapping_cast_fp16)[name = tensor<string, []>("x_15_cast_fp16")];
+            tensor<int32, []> var_153 = const()[name = tensor<string, []>("op_153"), val = tensor<int32, []>(-1)];
+            tensor<fp16, [2048, 256]> transformer_blocks_0_attention_norm_fc_weight_to_fp16 = const()[name = tensor<string, []>("transformer_blocks_0_attention_norm_fc_weight_to_fp16"), val = tensor<fp16, [2048, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5254016)))];
+            tensor<fp16, [2048]> transformer_blocks_0_attention_norm_fc_bias_to_fp16 = const()[name = tensor<string, []>("transformer_blocks_0_attention_norm_fc_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6302656)))];
+            tensor<fp16, [1, 2048]> linear_4_cast_fp16 = linear(bias = transformer_blocks_0_attention_norm_fc_bias_to_fp16, weight = transformer_blocks_0_attention_norm_fc_weight_to_fp16, x = features_to_fp16)[name = tensor<string, []>("linear_4_cast_fp16")];
+            tensor<int32, [3]> var_172 = const()[name = tensor<string, []>("op_172"), val = tensor<int32, [3]>([1, 2048, 1])];
+            tensor<fp16, [1, 2048, 1]> h_3_cast_fp16 = reshape(shape = var_172, x = linear_4_cast_fp16)[name = tensor<string, []>("h_3_cast_fp16")];
+            tensor<int32, [2]> var_174_split_sizes_0 = const()[name = tensor<string, []>("op_174_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])];
+            tensor<int32, []> var_174_axis_0 = const()[name = tensor<string, []>("op_174_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 1]> var_174_cast_fp16_0, tensor<fp16, [1, 1024, 1]> var_174_cast_fp16_1 = split(axis = var_174_axis_0, split_sizes = var_174_split_sizes_0, x = h_3_cast_fp16)[name = tensor<string, []>("op_174_cast_fp16")];
+            tensor<int32, [3]> gamma_3_perm_0 = const()[name = tensor<string, []>("gamma_3_perm_0"), val = tensor<int32, [3]>([0, -1, 1])];
+            tensor<int32, [3]> beta_3_perm_0 = const()[name = tensor<string, []>("beta_3_perm_0"), val = tensor<int32, [3]>([0, -1, 1])];
+            tensor<int32, [1]> x_19_axes_0 = const()[name = tensor<string, []>("x_19_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, []> var_146_to_fp16 = const()[name = tensor<string, []>("op_146_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 512, 1024]> x_19_cast_fp16 = layer_norm(axes = x_19_axes_0, epsilon = var_146_to_fp16, x = x_15_cast_fp16)[name = tensor<string, []>("x_19_cast_fp16")];
+            tensor<fp16, []> var_180_promoted_to_fp16 = const()[name = tensor<string, []>("op_180_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1, 1, 1024]> gamma_3_cast_fp16 = transpose(perm = gamma_3_perm_0, x = var_174_cast_fp16_0)[name = tensor<string, []>("transpose_40")];
+            tensor<fp16, [1, 1, 1024]> var_181_cast_fp16 = add(x = gamma_3_cast_fp16, y = var_180_promoted_to_fp16)[name = tensor<string, []>("op_181_cast_fp16")];
+            tensor<fp16, [1, 512, 1024]> var_182_cast_fp16 = mul(x = var_181_cast_fp16, y = x_19_cast_fp16)[name = tensor<string, []>("op_182_cast_fp16")];
+            tensor<fp16, [1, 1, 1024]> beta_3_cast_fp16 = transpose(perm = beta_3_perm_0, x = var_174_cast_fp16_1)[name = tensor<string, []>("transpose_39")];
+            tensor<fp16, [1, 512, 1024]> x_21_cast_fp16 = add(x = var_182_cast_fp16, y = beta_3_cast_fp16)[name = tensor<string, []>("x_21_cast_fp16")];
+            tensor<fp16, [2048, 256]> transformer_blocks_0_attention_norm_context_fc_weight_to_fp16 = const()[name = tensor<string, []>("transformer_blocks_0_attention_norm_context_fc_weight_to_fp16"), val = tensor<fp16, [2048, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6306816)))];
+            tensor<fp16, [2048]> transformer_blocks_0_attention_norm_context_fc_bias_to_fp16 = const()[name = tensor<string, []>("transformer_blocks_0_attention_norm_context_fc_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7355456)))];
+            tensor<fp16, [1, 2048]> linear_5_cast_fp16 = linear(bias = transformer_blocks_0_attention_norm_context_fc_bias_to_fp16, weight = transformer_blocks_0_attention_norm_context_fc_weight_to_fp16, x = features_to_fp16)[name = tensor<string, []>("linear_5_cast_fp16")];
+            tensor<int32, [3]> var_194 = const()[name = tensor<string, []>("op_194"), val = tensor<int32, [3]>([1, 2048, 1])];
+            tensor<fp16, [1, 2048, 1]> h_7_cast_fp16 = reshape(shape = var_194, x = linear_5_cast_fp16)[name = tensor<string, []>("h_7_cast_fp16")];
+            tensor<int32, [2]> var_196_split_sizes_0 = const()[name = tensor<string, []>("op_196_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])];
+            tensor<int32, []> var_196_axis_0 = const()[name = tensor<string, []>("op_196_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 1]> var_196_cast_fp16_0, tensor<fp16, [1, 1024, 1]> var_196_cast_fp16_1 = split(axis = var_196_axis_0, split_sizes = var_196_split_sizes_0, x = h_7_cast_fp16)[name = tensor<string, []>("op_196_cast_fp16")];
+            tensor<int32, [3]> gamma_7_perm_0 = const()[name = tensor<string, []>("gamma_7_perm_0"), val = tensor<int32, [3]>([0, -1, 1])];
+            tensor<int32, [3]> beta_7_perm_0 = const()[name = tensor<string, []>("beta_7_perm_0"), val = tensor<int32, [3]>([0, -1, 1])];
+            tensor<fp16, []> var_202_promoted_to_fp16 = const()[name = tensor<string, []>("op_202_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1, 1, 1024]> gamma_7_cast_fp16 = transpose(perm = gamma_7_perm_0, x = var_196_cast_fp16_0)[name = tensor<string, []>("transpose_38")];
+            tensor<fp16, [1, 1, 1024]> var_203_cast_fp16 = add(x = gamma_7_cast_fp16, y = var_202_promoted_to_fp16)[name = tensor<string, []>("op_203_cast_fp16")];
+            tensor<fp16, [1, 512, 1024]> var_204_cast_fp16 = mul(x = var_203_cast_fp16, y = x_19_cast_fp16)[name = tensor<string, []>("op_204_cast_fp16")];
+            tensor<fp16, [1, 1, 1024]> beta_7_cast_fp16 = transpose(perm = beta_7_perm_0, x = var_196_cast_fp16_1)[name = tensor<string, []>("transpose_37")];
+            tensor<fp16, [1, 512, 1024]> x_27_cast_fp16 = add(x = var_204_cast_fp16, y = beta_7_cast_fp16)[name = tensor<string, []>("x_27_cast_fp16")];
+            tensor<fp16, [512, 1024]> transformer_blocks_0_attention_to_q_weight_to_fp16 = const()[name = tensor<string, []>("transformer_blocks_0_attention_to_q_weight_to_fp16"), val = tensor<fp16, [512, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7359616)))];
+            tensor<fp16, [512]> linear_6_bias_0_to_fp16 = const()[name = tensor<string, []>("linear_6_bias_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8408256)))];
+            tensor<fp16, [1, 512, 512]> linear_6_cast_fp16 = linear(bias = linear_6_bias_0_to_fp16, weight = transformer_blocks_0_attention_to_q_weight_to_fp16, x = x_21_cast_fp16)[name = tensor<string, []>("linear_6_cast_fp16")];
+            tensor<fp16, [1024, 1024]> transformer_blocks_0_attention_to_kv_weight_to_fp16 = const()[name = tensor<string, []>("transformer_blocks_0_attention_to_kv_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8409344)))];
+            tensor<fp16, [1024]> linear_7_bias_0_to_fp16 = const()[name = tensor<string, []>("linear_7_bias_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10506560)))];
+            tensor<fp16, [1, 512, 1024]> linear_7_cast_fp16 = linear(bias = linear_7_bias_0_to_fp16, weight = transformer_blocks_0_attention_to_kv_weight_to_fp16, x = x_27_cast_fp16)[name = tensor<string, []>("linear_7_cast_fp16")];
+            tensor<int32, [2]> var_212_split_sizes_0 = const()[name = tensor<string, []>("op_212_split_sizes_0"), val = tensor<int32, [2]>([512, 512])];
+            tensor<int32, []> var_212_axis_0 = const()[name = tensor<string, []>("op_212_axis_0"), val = tensor<int32, []>(-1)];
+            tensor<fp16, [1, 512, 512]> var_212_cast_fp16_0, tensor<fp16, [1, 512, 512]> var_212_cast_fp16_1 = split(axis = var_212_axis_0, split_sizes = var_212_split_sizes_0, x = linear_7_cast_fp16)[name = tensor<string, []>("op_212_cast_fp16")];
+            tensor<int32, [4]> var_221 = const()[name = tensor<string, []>("op_221"), val = tensor<int32, [4]>([1, 512, 8, 64])];
+            tensor<fp16, [1, 512, 8, 64]> x_31_cast_fp16 = reshape(shape = var_221, x = linear_6_cast_fp16)[name = tensor<string, []>("x_31_cast_fp16")];
+            tensor<int32, [4]> var_231 = const()[name = tensor<string, []>("op_231"), val = tensor<int32, [4]>([1, 512, 8, 64])];
+            tensor<fp16, [1, 512, 8, 64]> x_35_cast_fp16 = reshape(shape = var_231, x = var_212_cast_fp16_0)[name = tensor<string, []>("x_35_cast_fp16")];
+            tensor<int32, [4]> var_241 = const()[name = tensor<string, []>("op_241"), val = tensor<int32, [4]>([1, 512, 8, 64])];
+            tensor<fp16, [1, 512, 8, 64]> x_39_cast_fp16 = reshape(shape = var_241, x = var_212_cast_fp16_1)[name = tensor<string, []>("x_39_cast_fp16")];
+            tensor<int32, [4]> var_243 = const()[name = tensor<string, []>("op_243"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> sim_1_transpose_x_0 = const()[name = tensor<string, []>("sim_1_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> sim_1_transpose_y_0 = const()[name = tensor<string, []>("sim_1_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_9_perm_0 = const()[name = tensor<string, []>("transpose_9_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_10_perm_0 = const()[name = tensor<string, []>("transpose_10_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 512]> transpose_10 = transpose(perm = transpose_10_perm_0, x = x_35_cast_fp16)[name = tensor<string, []>("transpose_34")];
+            tensor<fp16, [1, 8, 512, 64]> transpose_9 = transpose(perm = transpose_9_perm_0, x = x_31_cast_fp16)[name = tensor<string, []>("transpose_35")];
+            tensor<fp16, [1, 8, 512, 512]> sim_1_cast_fp16 = matmul(transpose_x = sim_1_transpose_x_0, transpose_y = sim_1_transpose_y_0, x = transpose_9, y = transpose_10)[name = tensor<string, []>("sim_1_cast_fp16")];
+            tensor<fp16, []> var_247_to_fp16 = const()[name = tensor<string, []>("op_247_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 512, 512]> sim_3_cast_fp16 = mul(x = sim_1_cast_fp16, y = var_247_to_fp16)[name = tensor<string, []>("sim_3_cast_fp16")];
+            tensor<fp16, [1, 8, 512, 512]> attn_1_cast_fp16 = softmax(axis = var_153, x = sim_3_cast_fp16)[name = tensor<string, []>("attn_1_cast_fp16")];
+            tensor<bool, []> x_41_transpose_x_0 = const()[name = tensor<string, []>("x_41_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_41_transpose_y_0 = const()[name = tensor<string, []>("x_41_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 512, 64]> v_1_cast_fp16 = transpose(perm = var_243, x = x_39_cast_fp16)[name = tensor<string, []>("transpose_36")];
+            tensor<fp16, [1, 8, 512, 64]> x_41_cast_fp16 = matmul(transpose_x = x_41_transpose_x_0, transpose_y = x_41_transpose_y_0, x = attn_1_cast_fp16, y = v_1_cast_fp16)[name = tensor<string, []>("x_41_cast_fp16")];
+            tensor<int32, [4]> var_269 = const()[name = tensor<string, []>("op_269"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_271 = const()[name = tensor<string, []>("op_271"), val = tensor<int32, [3]>([1, 512, 512])];
+            tensor<fp16, [1, 512, 8, 64]> x_43_cast_fp16 = transpose(perm = var_269, x = x_41_cast_fp16)[name = tensor<string, []>("transpose_33")];
+            tensor<fp16, [1, 512, 512]> input_23_cast_fp16 = reshape(shape = var_271, x = x_43_cast_fp16)[name = tensor<string, []>("input_23_cast_fp16")];
+            tensor<fp16, [1024, 512]> transformer_blocks_0_attention_attention_to_out_weight_to_fp16 = const()[name = tensor<string, []>("transformer_blocks_0_attention_attention_to_out_weight_to_fp16"), val = tensor<fp16, [1024, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10508672)))];
+            tensor<fp16, [1024]> transformer_blocks_0_attention_attention_to_out_bias_to_fp16 = const()[name = tensor<string, []>("transformer_blocks_0_attention_attention_to_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11557312)))];
+            tensor<fp16, [1, 512, 1024]> linear_8_cast_fp16 = linear(bias = transformer_blocks_0_attention_attention_to_out_bias_to_fp16, weight = transformer_blocks_0_attention_attention_to_out_weight_to_fp16, x = input_23_cast_fp16)[name = tensor<string, []>("linear_8_cast_fp16")];
+            tensor<fp16, [1, 512, 1024]> input_25_cast_fp16 = add(x = linear_8_cast_fp16, y = x_15_cast_fp16)[name = tensor<string, []>("input_25_cast_fp16")];
+            tensor<fp16, [2048, 1024]> transformer_blocks_0_feed_forward_0_weight_to_fp16 = const()[name = tensor<string, []>("transformer_blocks_0_feed_forward_0_weight_to_fp16"), val = tensor<fp16, [2048, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11559424)))];
+            tensor<fp16, [2048]> transformer_blocks_0_feed_forward_0_bias_to_fp16 = const()[name = tensor<string, []>("transformer_blocks_0_feed_forward_0_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15753792)))];
+            tensor<fp16, [1, 512, 2048]> linear_9_cast_fp16 = linear(bias = transformer_blocks_0_feed_forward_0_bias_to_fp16, weight = transformer_blocks_0_feed_forward_0_weight_to_fp16, x = input_25_cast_fp16)[name = tensor<string, []>("linear_9_cast_fp16")];
+            tensor<string, []> input_29_mode_0 = const()[name = tensor<string, []>("input_29_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 512, 2048]> input_29_cast_fp16 = gelu(mode = input_29_mode_0, x = linear_9_cast_fp16)[name = tensor<string, []>("input_29_cast_fp16")];
+            tensor<fp16, [1024, 2048]> transformer_blocks_0_feed_forward_2_weight_to_fp16 = const()[name = tensor<string, []>("transformer_blocks_0_feed_forward_2_weight_to_fp16"), val = tensor<fp16, [1024, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15757952)))];
+            tensor<fp16, [1024]> transformer_blocks_0_feed_forward_2_bias_to_fp16 = const()[name = tensor<string, []>("transformer_blocks_0_feed_forward_2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19952320)))];
+            tensor<fp16, [1, 512, 1024]> linear_10_cast_fp16 = linear(bias = transformer_blocks_0_feed_forward_2_bias_to_fp16, weight = transformer_blocks_0_feed_forward_2_weight_to_fp16, x = input_29_cast_fp16)[name = tensor<string, []>("linear_10_cast_fp16")];
+            tensor<fp16, [1, 512, 1024]> x_45_cast_fp16 = add(x = linear_10_cast_fp16, y = input_25_cast_fp16)[name = tensor<string, []>("x_45_cast_fp16")];
+            tensor<fp16, [1, 512, 1024]> x_47_cast_fp16 = add(x = x_45_cast_fp16, y = mapping_cast_fp16)[name = tensor<string, []>("x_47_cast_fp16")];
+            tensor<int32, []> var_298 = const()[name = tensor<string, []>("op_298"), val = tensor<int32, []>(-1)];
+            tensor<fp16, [2048, 256]> transformer_blocks_1_attention_norm_fc_weight_to_fp16 = const()[name = tensor<string, []>("transformer_blocks_1_attention_norm_fc_weight_to_fp16"), val = tensor<fp16, [2048, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19954432)))];
+            tensor<fp16, [2048]> transformer_blocks_1_attention_norm_fc_bias_to_fp16 = const()[name = tensor<string, []>("transformer_blocks_1_attention_norm_fc_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21003072)))];
+            tensor<fp16, [1, 2048]> linear_11_cast_fp16 = linear(bias = transformer_blocks_1_attention_norm_fc_bias_to_fp16, weight = transformer_blocks_1_attention_norm_fc_weight_to_fp16, x = features_to_fp16)[name = tensor<string, []>("linear_11_cast_fp16")];
+            tensor<int32, [3]> var_317 = const()[name = tensor<string, []>("op_317"), val = tensor<int32, [3]>([1, 2048, 1])];
+            tensor<fp16, [1, 2048, 1]> h_11_cast_fp16 = reshape(shape = var_317, x = linear_11_cast_fp16)[name = tensor<string, []>("h_11_cast_fp16")];
+            tensor<int32, [2]> var_319_split_sizes_0 = const()[name = tensor<string, []>("op_319_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])];
+            tensor<int32, []> var_319_axis_0 = const()[name = tensor<string, []>("op_319_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 1]> var_319_cast_fp16_0, tensor<fp16, [1, 1024, 1]> var_319_cast_fp16_1 = split(axis = var_319_axis_0, split_sizes = var_319_split_sizes_0, x = h_11_cast_fp16)[name = tensor<string, []>("op_319_cast_fp16")];
+            tensor<int32, [3]> gamma_11_perm_0 = const()[name = tensor<string, []>("gamma_11_perm_0"), val = tensor<int32, [3]>([0, -1, 1])];
+            tensor<int32, [3]> beta_11_perm_0 = const()[name = tensor<string, []>("beta_11_perm_0"), val = tensor<int32, [3]>([0, -1, 1])];
+            tensor<int32, [1]> x_51_axes_0 = const()[name = tensor<string, []>("x_51_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, []> var_291_to_fp16 = const()[name = tensor<string, []>("op_291_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 512, 1024]> x_51_cast_fp16 = layer_norm(axes = x_51_axes_0, epsilon = var_291_to_fp16, x = x_47_cast_fp16)[name = tensor<string, []>("x_51_cast_fp16")];
+            tensor<fp16, []> var_325_promoted_to_fp16 = const()[name = tensor<string, []>("op_325_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1, 1, 1024]> gamma_11_cast_fp16 = transpose(perm = gamma_11_perm_0, x = var_319_cast_fp16_0)[name = tensor<string, []>("transpose_32")];
+            tensor<fp16, [1, 1, 1024]> var_326_cast_fp16 = add(x = gamma_11_cast_fp16, y = var_325_promoted_to_fp16)[name = tensor<string, []>("op_326_cast_fp16")];
+            tensor<fp16, [1, 512, 1024]> var_327_cast_fp16 = mul(x = var_326_cast_fp16, y = x_51_cast_fp16)[name = tensor<string, []>("op_327_cast_fp16")];
+            tensor<fp16, [1, 1, 1024]> beta_11_cast_fp16 = transpose(perm = beta_11_perm_0, x = var_319_cast_fp16_1)[name = tensor<string, []>("transpose_31")];
+            tensor<fp16, [1, 512, 1024]> x_53_cast_fp16 = add(x = var_327_cast_fp16, y = beta_11_cast_fp16)[name = tensor<string, []>("x_53_cast_fp16")];
+            tensor<fp16, [2048, 256]> transformer_blocks_1_attention_norm_context_fc_weight_to_fp16 = const()[name = tensor<string, []>("transformer_blocks_1_attention_norm_context_fc_weight_to_fp16"), val = tensor<fp16, [2048, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21007232)))];
+            tensor<fp16, [2048]> transformer_blocks_1_attention_norm_context_fc_bias_to_fp16 = const()[name = tensor<string, []>("transformer_blocks_1_attention_norm_context_fc_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22055872)))];
+            tensor<fp16, [1, 2048]> linear_12_cast_fp16 = linear(bias = transformer_blocks_1_attention_norm_context_fc_bias_to_fp16, weight = transformer_blocks_1_attention_norm_context_fc_weight_to_fp16, x = features_to_fp16)[name = tensor<string, []>("linear_12_cast_fp16")];
+            tensor<int32, [3]> var_339 = const()[name = tensor<string, []>("op_339"), val = tensor<int32, [3]>([1, 2048, 1])];
+            tensor<fp16, [1, 2048, 1]> h_15_cast_fp16 = reshape(shape = var_339, x = linear_12_cast_fp16)[name = tensor<string, []>("h_15_cast_fp16")];
+            tensor<int32, [2]> var_341_split_sizes_0 = const()[name = tensor<string, []>("op_341_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])];
+            tensor<int32, []> var_341_axis_0 = const()[name = tensor<string, []>("op_341_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 1]> var_341_cast_fp16_0, tensor<fp16, [1, 1024, 1]> var_341_cast_fp16_1 = split(axis = var_341_axis_0, split_sizes = var_341_split_sizes_0, x = h_15_cast_fp16)[name = tensor<string, []>("op_341_cast_fp16")];
+            tensor<int32, [3]> gamma_15_perm_0 = const()[name = tensor<string, []>("gamma_15_perm_0"), val = tensor<int32, [3]>([0, -1, 1])];
+            tensor<int32, [3]> beta_15_perm_0 = const()[name = tensor<string, []>("beta_15_perm_0"), val = tensor<int32, [3]>([0, -1, 1])];
+            tensor<fp16, []> var_347_promoted_to_fp16 = const()[name = tensor<string, []>("op_347_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1, 1, 1024]> gamma_15_cast_fp16 = transpose(perm = gamma_15_perm_0, x = var_341_cast_fp16_0)[name = tensor<string, []>("transpose_30")];
+            tensor<fp16, [1, 1, 1024]> var_348_cast_fp16 = add(x = gamma_15_cast_fp16, y = var_347_promoted_to_fp16)[name = tensor<string, []>("op_348_cast_fp16")];
+            tensor<fp16, [1, 512, 1024]> var_349_cast_fp16 = mul(x = var_348_cast_fp16, y = x_51_cast_fp16)[name = tensor<string, []>("op_349_cast_fp16")];
+            tensor<fp16, [1, 1, 1024]> beta_15_cast_fp16 = transpose(perm = beta_15_perm_0, x = var_341_cast_fp16_1)[name = tensor<string, []>("transpose_29")];
+            tensor<fp16, [1, 512, 1024]> x_59_cast_fp16 = add(x = var_349_cast_fp16, y = beta_15_cast_fp16)[name = tensor<string, []>("x_59_cast_fp16")];
+            tensor<fp16, [512, 1024]> transformer_blocks_1_attention_to_q_weight_to_fp16 = const()[name = tensor<string, []>("transformer_blocks_1_attention_to_q_weight_to_fp16"), val = tensor<fp16, [512, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22060032)))];
+            tensor<fp16, [1, 512, 512]> linear_13_cast_fp16 = linear(bias = linear_6_bias_0_to_fp16, weight = transformer_blocks_1_attention_to_q_weight_to_fp16, x = x_53_cast_fp16)[name = tensor<string, []>("linear_13_cast_fp16")];
+            tensor<fp16, [1024, 1024]> transformer_blocks_1_attention_to_kv_weight_to_fp16 = const()[name = tensor<string, []>("transformer_blocks_1_attention_to_kv_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23108672)))];
+            tensor<fp16, [1, 512, 1024]> linear_14_cast_fp16 = linear(bias = linear_7_bias_0_to_fp16, weight = transformer_blocks_1_attention_to_kv_weight_to_fp16, x = x_59_cast_fp16)[name = tensor<string, []>("linear_14_cast_fp16")];
+            tensor<int32, [2]> var_357_split_sizes_0 = const()[name = tensor<string, []>("op_357_split_sizes_0"), val = tensor<int32, [2]>([512, 512])];
+            tensor<int32, []> var_357_axis_0 = const()[name = tensor<string, []>("op_357_axis_0"), val = tensor<int32, []>(-1)];
+            tensor<fp16, [1, 512, 512]> var_357_cast_fp16_0, tensor<fp16, [1, 512, 512]> var_357_cast_fp16_1 = split(axis = var_357_axis_0, split_sizes = var_357_split_sizes_0, x = linear_14_cast_fp16)[name = tensor<string, []>("op_357_cast_fp16")];
+            tensor<int32, [4]> var_366 = const()[name = tensor<string, []>("op_366"), val = tensor<int32, [4]>([1, 512, 8, 64])];
+            tensor<fp16, [1, 512, 8, 64]> x_63_cast_fp16 = reshape(shape = var_366, x = linear_13_cast_fp16)[name = tensor<string, []>("x_63_cast_fp16")];
+            tensor<int32, [4]> var_376 = const()[name = tensor<string, []>("op_376"), val = tensor<int32, [4]>([1, 512, 8, 64])];
+            tensor<fp16, [1, 512, 8, 64]> x_67_cast_fp16 = reshape(shape = var_376, x = var_357_cast_fp16_0)[name = tensor<string, []>("x_67_cast_fp16")];
+            tensor<int32, [4]> var_386 = const()[name = tensor<string, []>("op_386"), val = tensor<int32, [4]>([1, 512, 8, 64])];
+            tensor<fp16, [1, 512, 8, 64]> x_71_cast_fp16 = reshape(shape = var_386, x = var_357_cast_fp16_1)[name = tensor<string, []>("x_71_cast_fp16")];
+            tensor<int32, [4]> var_388 = const()[name = tensor<string, []>("op_388"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> sim_5_transpose_x_0 = const()[name = tensor<string, []>("sim_5_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> sim_5_transpose_y_0 = const()[name = tensor<string, []>("sim_5_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_11_perm_0 = const()[name = tensor<string, []>("transpose_11_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_12_perm_0 = const()[name = tensor<string, []>("transpose_12_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 512]> transpose_12 = transpose(perm = transpose_12_perm_0, x = x_67_cast_fp16)[name = tensor<string, []>("transpose_26")];
+            tensor<fp16, [1, 8, 512, 64]> transpose_11 = transpose(perm = transpose_11_perm_0, x = x_63_cast_fp16)[name = tensor<string, []>("transpose_27")];
+            tensor<fp16, [1, 8, 512, 512]> sim_5_cast_fp16 = matmul(transpose_x = sim_5_transpose_x_0, transpose_y = sim_5_transpose_y_0, x = transpose_11, y = transpose_12)[name = tensor<string, []>("sim_5_cast_fp16")];
+            tensor<fp16, []> var_392_to_fp16 = const()[name = tensor<string, []>("op_392_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 512, 512]> sim_7_cast_fp16 = mul(x = sim_5_cast_fp16, y = var_392_to_fp16)[name = tensor<string, []>("sim_7_cast_fp16")];
+            tensor<fp16, [1, 8, 512, 512]> attn_3_cast_fp16 = softmax(axis = var_298, x = sim_7_cast_fp16)[name = tensor<string, []>("attn_3_cast_fp16")];
+            tensor<bool, []> x_73_transpose_x_0 = const()[name = tensor<string, []>("x_73_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_73_transpose_y_0 = const()[name = tensor<string, []>("x_73_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 512, 64]> v_3_cast_fp16 = transpose(perm = var_388, x = x_71_cast_fp16)[name = tensor<string, []>("transpose_28")];
+            tensor<fp16, [1, 8, 512, 64]> x_73_cast_fp16 = matmul(transpose_x = x_73_transpose_x_0, transpose_y = x_73_transpose_y_0, x = attn_3_cast_fp16, y = v_3_cast_fp16)[name = tensor<string, []>("x_73_cast_fp16")];
+            tensor<int32, [4]> var_414 = const()[name = tensor<string, []>("op_414"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_416 = const()[name = tensor<string, []>("op_416"), val = tensor<int32, [3]>([1, 512, 512])];
+            tensor<fp16, [1, 512, 8, 64]> x_75_cast_fp16 = transpose(perm = var_414, x = x_73_cast_fp16)[name = tensor<string, []>("transpose_25")];
+            tensor<fp16, [1, 512, 512]> input_39_cast_fp16 = reshape(shape = var_416, x = x_75_cast_fp16)[name = tensor<string, []>("input_39_cast_fp16")];
+            tensor<fp16, [1024, 512]> transformer_blocks_1_attention_attention_to_out_weight_to_fp16 = const()[name = tensor<string, []>("transformer_blocks_1_attention_attention_to_out_weight_to_fp16"), val = tensor<fp16, [1024, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25205888)))];
+            tensor<fp16, [1024]> transformer_blocks_1_attention_attention_to_out_bias_to_fp16 = const()[name = tensor<string, []>("transformer_blocks_1_attention_attention_to_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26254528)))];
+            tensor<fp16, [1, 512, 1024]> linear_15_cast_fp16 = linear(bias = transformer_blocks_1_attention_attention_to_out_bias_to_fp16, weight = transformer_blocks_1_attention_attention_to_out_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("linear_15_cast_fp16")];
+            tensor<fp16, [1, 512, 1024]> input_41_cast_fp16 = add(x = linear_15_cast_fp16, y = x_47_cast_fp16)[name = tensor<string, []>("input_41_cast_fp16")];
+            tensor<fp16, [2048, 1024]> transformer_blocks_1_feed_forward_0_weight_to_fp16 = const()[name = tensor<string, []>("transformer_blocks_1_feed_forward_0_weight_to_fp16"), val = tensor<fp16, [2048, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26256640)))];
+            tensor<fp16, [2048]> transformer_blocks_1_feed_forward_0_bias_to_fp16 = const()[name = tensor<string, []>("transformer_blocks_1_feed_forward_0_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30451008)))];
+            tensor<fp16, [1, 512, 2048]> linear_16_cast_fp16 = linear(bias = transformer_blocks_1_feed_forward_0_bias_to_fp16, weight = transformer_blocks_1_feed_forward_0_weight_to_fp16, x = input_41_cast_fp16)[name = tensor<string, []>("linear_16_cast_fp16")];
+            tensor<string, []> input_45_mode_0 = const()[name = tensor<string, []>("input_45_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 512, 2048]> input_45_cast_fp16 = gelu(mode = input_45_mode_0, x = linear_16_cast_fp16)[name = tensor<string, []>("input_45_cast_fp16")];
+            tensor<fp16, [1024, 2048]> transformer_blocks_1_feed_forward_2_weight_to_fp16 = const()[name = tensor<string, []>("transformer_blocks_1_feed_forward_2_weight_to_fp16"), val = tensor<fp16, [1024, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30455168)))];
+            tensor<fp16, [1024]> transformer_blocks_1_feed_forward_2_bias_to_fp16 = const()[name = tensor<string, []>("transformer_blocks_1_feed_forward_2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34649536)))];
+            tensor<fp16, [1, 512, 1024]> linear_17_cast_fp16 = linear(bias = transformer_blocks_1_feed_forward_2_bias_to_fp16, weight = transformer_blocks_1_feed_forward_2_weight_to_fp16, x = input_45_cast_fp16)[name = tensor<string, []>("linear_17_cast_fp16")];
+            tensor<fp16, [1, 512, 1024]> x_77_cast_fp16 = add(x = linear_17_cast_fp16, y = input_41_cast_fp16)[name = tensor<string, []>("x_77_cast_fp16")];
+            tensor<fp16, [1, 512, 1024]> x_79_cast_fp16 = add(x = x_77_cast_fp16, y = mapping_cast_fp16)[name = tensor<string, []>("x_79_cast_fp16")];
+            tensor<int32, []> var_443 = const()[name = tensor<string, []>("op_443"), val = tensor<int32, []>(-1)];
+            tensor<fp16, [2048, 256]> transformer_blocks_2_attention_norm_fc_weight_to_fp16 = const()[name = tensor<string, []>("transformer_blocks_2_attention_norm_fc_weight_to_fp16"), val = tensor<fp16, [2048, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34651648)))];
+            tensor<fp16, [2048]> transformer_blocks_2_attention_norm_fc_bias_to_fp16 = const()[name = tensor<string, []>("transformer_blocks_2_attention_norm_fc_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35700288)))];
+            tensor<fp16, [1, 2048]> linear_18_cast_fp16 = linear(bias = transformer_blocks_2_attention_norm_fc_bias_to_fp16, weight = transformer_blocks_2_attention_norm_fc_weight_to_fp16, x = features_to_fp16)[name = tensor<string, []>("linear_18_cast_fp16")];
+            tensor<int32, [3]> var_462 = const()[name = tensor<string, []>("op_462"), val = tensor<int32, [3]>([1, 2048, 1])];
+            tensor<fp16, [1, 2048, 1]> h_19_cast_fp16 = reshape(shape = var_462, x = linear_18_cast_fp16)[name = tensor<string, []>("h_19_cast_fp16")];
+            tensor<int32, [2]> var_464_split_sizes_0 = const()[name = tensor<string, []>("op_464_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])];
+            tensor<int32, []> var_464_axis_0 = const()[name = tensor<string, []>("op_464_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 1]> var_464_cast_fp16_0, tensor<fp16, [1, 1024, 1]> var_464_cast_fp16_1 = split(axis = var_464_axis_0, split_sizes = var_464_split_sizes_0, x = h_19_cast_fp16)[name = tensor<string, []>("op_464_cast_fp16")];
+            tensor<int32, [3]> gamma_19_perm_0 = const()[name = tensor<string, []>("gamma_19_perm_0"), val = tensor<int32, [3]>([0, -1, 1])];
+            tensor<int32, [3]> beta_19_perm_0 = const()[name = tensor<string, []>("beta_19_perm_0"), val = tensor<int32, [3]>([0, -1, 1])];
+            tensor<int32, [1]> x_83_axes_0 = const()[name = tensor<string, []>("x_83_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, []> var_436_to_fp16 = const()[name = tensor<string, []>("op_436_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 512, 1024]> x_83_cast_fp16 = layer_norm(axes = x_83_axes_0, epsilon = var_436_to_fp16, x = x_79_cast_fp16)[name = tensor<string, []>("x_83_cast_fp16")];
+            tensor<fp16, []> var_470_promoted_to_fp16 = const()[name = tensor<string, []>("op_470_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1, 1, 1024]> gamma_19_cast_fp16 = transpose(perm = gamma_19_perm_0, x = var_464_cast_fp16_0)[name = tensor<string, []>("transpose_24")];
+            tensor<fp16, [1, 1, 1024]> var_471_cast_fp16 = add(x = gamma_19_cast_fp16, y = var_470_promoted_to_fp16)[name = tensor<string, []>("op_471_cast_fp16")];
+            tensor<fp16, [1, 512, 1024]> var_472_cast_fp16 = mul(x = var_471_cast_fp16, y = x_83_cast_fp16)[name = tensor<string, []>("op_472_cast_fp16")];
+            tensor<fp16, [1, 1, 1024]> beta_19_cast_fp16 = transpose(perm = beta_19_perm_0, x = var_464_cast_fp16_1)[name = tensor<string, []>("transpose_23")];
+            tensor<fp16, [1, 512, 1024]> x_85_cast_fp16 = add(x = var_472_cast_fp16, y = beta_19_cast_fp16)[name = tensor<string, []>("x_85_cast_fp16")];
+            tensor<fp16, [2048, 256]> transformer_blocks_2_attention_norm_context_fc_weight_to_fp16 = const()[name = tensor<string, []>("transformer_blocks_2_attention_norm_context_fc_weight_to_fp16"), val = tensor<fp16, [2048, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35704448)))];
+            tensor<fp16, [2048]> transformer_blocks_2_attention_norm_context_fc_bias_to_fp16 = const()[name = tensor<string, []>("transformer_blocks_2_attention_norm_context_fc_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36753088)))];
+            tensor<fp16, [1, 2048]> linear_19_cast_fp16 = linear(bias = transformer_blocks_2_attention_norm_context_fc_bias_to_fp16, weight = transformer_blocks_2_attention_norm_context_fc_weight_to_fp16, x = features_to_fp16)[name = tensor<string, []>("linear_19_cast_fp16")];
+            tensor<int32, [3]> var_484 = const()[name = tensor<string, []>("op_484"), val = tensor<int32, [3]>([1, 2048, 1])];
+            tensor<fp16, [1, 2048, 1]> h_cast_fp16 = reshape(shape = var_484, x = linear_19_cast_fp16)[name = tensor<string, []>("h_cast_fp16")];
+            tensor<int32, [2]> var_486_split_sizes_0 = const()[name = tensor<string, []>("op_486_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])];
+            tensor<int32, []> var_486_axis_0 = const()[name = tensor<string, []>("op_486_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1024, 1]> var_486_cast_fp16_0, tensor<fp16, [1, 1024, 1]> var_486_cast_fp16_1 = split(axis = var_486_axis_0, split_sizes = var_486_split_sizes_0, x = h_cast_fp16)[name = tensor<string, []>("op_486_cast_fp16")];
+            tensor<int32, [3]> gamma_perm_0 = const()[name = tensor<string, []>("gamma_perm_0"), val = tensor<int32, [3]>([0, -1, 1])];
+            tensor<int32, [3]> beta_perm_0 = const()[name = tensor<string, []>("beta_perm_0"), val = tensor<int32, [3]>([0, -1, 1])];
+            tensor<fp16, []> var_492_promoted_to_fp16 = const()[name = tensor<string, []>("op_492_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1, 1, 1024]> gamma_cast_fp16 = transpose(perm = gamma_perm_0, x = var_486_cast_fp16_0)[name = tensor<string, []>("transpose_22")];
+            tensor<fp16, [1, 1, 1024]> var_493_cast_fp16 = add(x = gamma_cast_fp16, y = var_492_promoted_to_fp16)[name = tensor<string, []>("op_493_cast_fp16")];
+            tensor<fp16, [1, 512, 1024]> var_494_cast_fp16 = mul(x = var_493_cast_fp16, y = x_83_cast_fp16)[name = tensor<string, []>("op_494_cast_fp16")];
+            tensor<fp16, [1, 1, 1024]> beta_cast_fp16 = transpose(perm = beta_perm_0, x = var_486_cast_fp16_1)[name = tensor<string, []>("transpose_21")];
+            tensor<fp16, [1, 512, 1024]> x_91_cast_fp16 = add(x = var_494_cast_fp16, y = beta_cast_fp16)[name = tensor<string, []>("x_91_cast_fp16")];
+            tensor<fp16, [512, 1024]> transformer_blocks_2_attention_to_q_weight_to_fp16 = const()[name = tensor<string, []>("transformer_blocks_2_attention_to_q_weight_to_fp16"), val = tensor<fp16, [512, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36757248)))];
+            tensor<fp16, [1, 512, 512]> linear_20_cast_fp16 = linear(bias = linear_6_bias_0_to_fp16, weight = transformer_blocks_2_attention_to_q_weight_to_fp16, x = x_85_cast_fp16)[name = tensor<string, []>("linear_20_cast_fp16")];
+            tensor<fp16, [1024, 1024]> transformer_blocks_2_attention_to_kv_weight_to_fp16 = const()[name = tensor<string, []>("transformer_blocks_2_attention_to_kv_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37805888)))];
+            tensor<fp16, [1, 512, 1024]> linear_21_cast_fp16 = linear(bias = linear_7_bias_0_to_fp16, weight = transformer_blocks_2_attention_to_kv_weight_to_fp16, x = x_91_cast_fp16)[name = tensor<string, []>("linear_21_cast_fp16")];
+            tensor<int32, [2]> var_502_split_sizes_0 = const()[name = tensor<string, []>("op_502_split_sizes_0"), val = tensor<int32, [2]>([512, 512])];
+            tensor<int32, []> var_502_axis_0 = const()[name = tensor<string, []>("op_502_axis_0"), val = tensor<int32, []>(-1)];
+            tensor<fp16, [1, 512, 512]> var_502_cast_fp16_0, tensor<fp16, [1, 512, 512]> var_502_cast_fp16_1 = split(axis = var_502_axis_0, split_sizes = var_502_split_sizes_0, x = linear_21_cast_fp16)[name = tensor<string, []>("op_502_cast_fp16")];
+            tensor<int32, [4]> var_511 = const()[name = tensor<string, []>("op_511"), val = tensor<int32, [4]>([1, 512, 8, 64])];
+            tensor<fp16, [1, 512, 8, 64]> x_95_cast_fp16 = reshape(shape = var_511, x = linear_20_cast_fp16)[name = tensor<string, []>("x_95_cast_fp16")];
+            tensor<int32, [4]> var_521 = const()[name = tensor<string, []>("op_521"), val = tensor<int32, [4]>([1, 512, 8, 64])];
+            tensor<fp16, [1, 512, 8, 64]> x_99_cast_fp16 = reshape(shape = var_521, x = var_502_cast_fp16_0)[name = tensor<string, []>("x_99_cast_fp16")];
+            tensor<int32, [4]> var_531 = const()[name = tensor<string, []>("op_531"), val = tensor<int32, [4]>([1, 512, 8, 64])];
+            tensor<fp16, [1, 512, 8, 64]> x_103_cast_fp16 = reshape(shape = var_531, x = var_502_cast_fp16_1)[name = tensor<string, []>("x_103_cast_fp16")];
+            tensor<int32, [4]> var_533 = const()[name = tensor<string, []>("op_533"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> sim_9_transpose_x_0 = const()[name = tensor<string, []>("sim_9_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> sim_9_transpose_y_0 = const()[name = tensor<string, []>("sim_9_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_13_perm_0 = const()[name = tensor<string, []>("transpose_13_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_14_perm_0 = const()[name = tensor<string, []>("transpose_14_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 512]> transpose_14 = transpose(perm = transpose_14_perm_0, x = x_99_cast_fp16)[name = tensor<string, []>("transpose_18")];
+            tensor<fp16, [1, 8, 512, 64]> transpose_13 = transpose(perm = transpose_13_perm_0, x = x_95_cast_fp16)[name = tensor<string, []>("transpose_19")];
+            tensor<fp16, [1, 8, 512, 512]> sim_9_cast_fp16 = matmul(transpose_x = sim_9_transpose_x_0, transpose_y = sim_9_transpose_y_0, x = transpose_13, y = transpose_14)[name = tensor<string, []>("sim_9_cast_fp16")];
+            tensor<fp16, []> var_537_to_fp16 = const()[name = tensor<string, []>("op_537_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 512, 512]> sim_cast_fp16 = mul(x = sim_9_cast_fp16, y = var_537_to_fp16)[name = tensor<string, []>("sim_cast_fp16")];
+            tensor<fp16, [1, 8, 512, 512]> attn_cast_fp16 = softmax(axis = var_443, x = sim_cast_fp16)[name = tensor<string, []>("attn_cast_fp16")];
+            tensor<bool, []> x_105_transpose_x_0 = const()[name = tensor<string, []>("x_105_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_105_transpose_y_0 = const()[name = tensor<string, []>("x_105_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 512, 64]> v_cast_fp16 = transpose(perm = var_533, x = x_103_cast_fp16)[name = tensor<string, []>("transpose_20")];
+            tensor<fp16, [1, 8, 512, 64]> x_105_cast_fp16 = matmul(transpose_x = x_105_transpose_x_0, transpose_y = x_105_transpose_y_0, x = attn_cast_fp16, y = v_cast_fp16)[name = tensor<string, []>("x_105_cast_fp16")];
+            tensor<int32, [4]> var_559 = const()[name = tensor<string, []>("op_559"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_561 = const()[name = tensor<string, []>("op_561"), val = tensor<int32, [3]>([1, 512, 512])];
+            tensor<fp16, [1, 512, 8, 64]> x_107_cast_fp16 = transpose(perm = var_559, x = x_105_cast_fp16)[name = tensor<string, []>("transpose_17")];
+            tensor<fp16, [1, 512, 512]> input_55_cast_fp16 = reshape(shape = var_561, x = x_107_cast_fp16)[name = tensor<string, []>("input_55_cast_fp16")];
+            tensor<fp16, [1024, 512]> transformer_blocks_2_attention_attention_to_out_weight_to_fp16 = const()[name = tensor<string, []>("transformer_blocks_2_attention_attention_to_out_weight_to_fp16"), val = tensor<fp16, [1024, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39903104)))];
+            tensor<fp16, [1024]> transformer_blocks_2_attention_attention_to_out_bias_to_fp16 = const()[name = tensor<string, []>("transformer_blocks_2_attention_attention_to_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40951744)))];
+            tensor<fp16, [1, 512, 1024]> linear_22_cast_fp16 = linear(bias = transformer_blocks_2_attention_attention_to_out_bias_to_fp16, weight = transformer_blocks_2_attention_attention_to_out_weight_to_fp16, x = input_55_cast_fp16)[name = tensor<string, []>("linear_22_cast_fp16")];
+            tensor<fp16, [1, 512, 1024]> input_57_cast_fp16 = add(x = linear_22_cast_fp16, y = x_79_cast_fp16)[name = tensor<string, []>("input_57_cast_fp16")];
+            tensor<fp16, [2048, 1024]> transformer_blocks_2_feed_forward_0_weight_to_fp16 = const()[name = tensor<string, []>("transformer_blocks_2_feed_forward_0_weight_to_fp16"), val = tensor<fp16, [2048, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40953856)))];
+            tensor<fp16, [2048]> transformer_blocks_2_feed_forward_0_bias_to_fp16 = const()[name = tensor<string, []>("transformer_blocks_2_feed_forward_0_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45148224)))];
+            tensor<fp16, [1, 512, 2048]> linear_23_cast_fp16 = linear(bias = transformer_blocks_2_feed_forward_0_bias_to_fp16, weight = transformer_blocks_2_feed_forward_0_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("linear_23_cast_fp16")];
+            tensor<string, []> input_61_mode_0 = const()[name = tensor<string, []>("input_61_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 512, 2048]> input_61_cast_fp16 = gelu(mode = input_61_mode_0, x = linear_23_cast_fp16)[name = tensor<string, []>("input_61_cast_fp16")];
+            tensor<fp16, [1024, 2048]> transformer_blocks_2_feed_forward_2_weight_to_fp16 = const()[name = tensor<string, []>("transformer_blocks_2_feed_forward_2_weight_to_fp16"), val = tensor<fp16, [1024, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45152384)))];
+            tensor<fp16, [1024]> transformer_blocks_2_feed_forward_2_bias_to_fp16 = const()[name = tensor<string, []>("transformer_blocks_2_feed_forward_2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49346752)))];
+            tensor<fp16, [1, 512, 1024]> linear_24_cast_fp16 = linear(bias = transformer_blocks_2_feed_forward_2_bias_to_fp16, weight = transformer_blocks_2_feed_forward_2_weight_to_fp16, x = input_61_cast_fp16)[name = tensor<string, []>("linear_24_cast_fp16")];
+            tensor<fp16, [1, 512, 1024]> x_109_cast_fp16 = add(x = linear_24_cast_fp16, y = input_57_cast_fp16)[name = tensor<string, []>("x_109_cast_fp16")];
+            tensor<int32, [1]> var_581_axes_0 = const()[name = tensor<string, []>("op_581_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, []> var_581_keep_dims_0 = const()[name = tensor<string, []>("op_581_keep_dims_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1024]> var_581_cast_fp16 = reduce_mean(axes = var_581_axes_0, keep_dims = var_581_keep_dims_0, x = x_109_cast_fp16)[name = tensor<string, []>("op_581_cast_fp16")];
+            tensor<int32, [1]> x_111_axes_0 = const()[name = tensor<string, []>("x_111_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1024]> x_111_cast_fp16 = expand_dims(axes = x_111_axes_0, x = var_581_cast_fp16)[name = tensor<string, []>("x_111_cast_fp16")];
+            tensor<int32, [3]> var_590 = const()[name = tensor<string, []>("op_590"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> x_pad_type_0 = const()[name = tensor<string, []>("x_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_strides_0 = const()[name = tensor<string, []>("x_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_pad_0 = const()[name = tensor<string, []>("x_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_dilations_0 = const()[name = tensor<string, []>("x_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_groups_0 = const()[name = tensor<string, []>("x_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [256, 1024, 1]> transformer_to_out_1_weight_to_fp16 = const()[name = tensor<string, []>("transformer_to_out_1_weight_to_fp16"), val = tensor<fp16, [256, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49348864)))];
+            tensor<fp16, [256]> transformer_to_out_1_bias_to_fp16 = const()[name = tensor<string, []>("transformer_to_out_1_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49873216)))];
+            tensor<fp16, [1, 1024, 1]> input_cast_fp16 = transpose(perm = var_590, x = x_111_cast_fp16)[name = tensor<string, []>("transpose_16")];
+            tensor<fp16, [1, 256, 1]> x_cast_fp16 = conv(bias = transformer_to_out_1_bias_to_fp16, dilations = x_dilations_0, groups = x_groups_0, pad = x_pad_0, pad_type = x_pad_type_0, strides = x_strides_0, weight = transformer_to_out_1_weight_to_fp16, x = input_cast_fp16)[name = tensor<string, []>("x_cast_fp16")];
+            tensor<int32, [3]> x_pred_perm_0 = const()[name = tensor<string, []>("x_pred_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
+            tensor<fp16, [1, 1, 256]> var_602_cast_fp16 = mul(x = c_skip_cast_fp16, y = x_noisy_to_fp16)[name = tensor<string, []>("op_602_cast_fp16")];
+            tensor<fp16, [1, 1, 256]> x_pred_cast_fp16 = transpose(perm = x_pred_perm_0, x = x_cast_fp16)[name = tensor<string, []>("transpose_15")];
+            tensor<fp16, [1, 1, 256]> var_603_cast_fp16 = mul(x = c_out_cast_fp16, y = x_pred_cast_fp16)[name = tensor<string, []>("op_603_cast_fp16")];
+            tensor<fp16, [1, 1, 256]> denoised = add(x = var_602_cast_fp16, y = var_603_cast_fp16)[name = tensor<string, []>("op_605_cast_fp16")];
+        } -> (denoised);
+}

compiled/styletts2_diffusion_step_512.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:17ba48a8bdc68851289a23593b223573aaddd1b445e8c77765f5350feed8a251
+size 49873792

compiled/styletts2_f0n_energy.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5467b4fad22d57b3e2e115d3b52104d6d7a0288ae95fe2f7d0789c79cda9bb9d
+size 243

compiled/styletts2_f0n_energy.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:121cb6ca7a0aedd3d608362b9dc12ba3b15a09b81c827b5578b20cf4364ee5d3
+size 427

compiled/styletts2_f0n_energy.mlmodelc/metadata.json

@@ -0,0 +1,98 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "StyleTTS2 F0Ntrain (LibriTTS). ANE-eligible.",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16)",
+        "shortDescription" : "",
+        "shape" : "[]",
+        "name" : "F0",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16)",
+        "shortDescription" : "",
+        "shape" : "[]",
+        "name" : "N",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Float16",
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios17.squeeze" : 4,
+      "Ios17.mul" : 18,
+      "Ios17.linear" : 12,
+      "Ios17.transpose" : 3,
+      "Ios17.conv" : 16,
+      "Ios17.leakyRelu" : 12,
+      "Ios17.add" : 30,
+      "Ios17.convTranspose" : 2,
+      "Ios17.lstm" : 1,
+      "Ios17.sliceByIndex" : 2,
+      "UpsampleNearestNeighbor" : 2,
+      "Ios17.expandDims" : 2,
+      "Ios17.instanceNorm" : 11,
+      "Ios17.reshape" : 12,
+      "Ios17.cast" : 2,
+      "Split" : 12
+    },
+    "computePrecision" : "Mixed (Float16, Float32, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "shortDescription" : "",
+        "dataType" : "Float32",
+        "hasShapeFlexibility" : "1",
+        "isOptional" : "0",
+        "shapeFlexibility" : "1 × 640 × 4096 | 1 × 640 × 256 | 1 × 640 × 512 | 1 × 640 × 1024 | 1 × 640 × 2048",
+        "formattedType" : "MultiArray (Float32 1 × 640 × 4096)",
+        "type" : "MultiArray",
+        "shape" : "[1, 640, 4096]",
+        "name" : "en",
+        "enumeratedShapes" : "[[1, 640, 4096], [1, 640, 256], [1, 640, 512], [1, 640, 1024], [1, 640, 2048]]"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 128]",
+        "name" : "s",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.version" : "8.3.0",
+      "com.github.apple.coremltools.source" : "torch==2.11.0",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "generatedClassName" : "styletts2_f0n_energy",
+    "method" : "predict"
+  }
+]

compiled/styletts2_f0n_energy.mlmodelc/model.mil

@@ -0,0 +1,396 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.11.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios17>(tensor<fp32, [1, 640, ?]> en, tensor<fp32, [1, 128]> s) [FlexibleShapeInformation = tuple<tuple<tensor<string, []>, dict<tensor<string, []>, tensor<int32, [?]>>>, tuple<tensor<string, []>, dict<tensor<string, []>, dict<tensor<string, []>, tensor<int32, [?]>>>>>((("DefaultShapes", {{"en", [1, 640, 4096]}}), ("EnumeratedShapes", {{"en_1_1_1_640_1024_s_1_1_1_1_128_", {{"en", [1, 640, 1024]}, {"s", [1, 128]}}}, {"en_1_1_1_640_2048_s_1_1_1_1_128_", {{"en", [1, 640, 2048]}, {"s", [1, 128]}}}, {"en_1_1_1_640_256_s_1_1_1_1_128_", {{"en", [1, 640, 256]}, {"s", [1, 128]}}}, {"en_1_1_1_640_4096_s_1_1_1_1_128_", {{"en", [1, 640, 4096]}, {"s", [1, 128]}}}, {"en_1_1_1_640_512_s_1_1_1_1_128_", {{"en", [1, 640, 512]}, {"s", [1, 128]}}}})))] {
+            tensor<int32, [3]> transpose_0_perm_0 = const()[name = tensor<string, []>("transpose_0_perm_0"), val = tensor<int32, [3]>([-1, 0, -2])];
+            tensor<string, []> en_to_fp16_dtype_0 = const()[name = tensor<string, []>("en_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<string, []> x_batch_first_direction_0 = const()[name = tensor<string, []>("x_batch_first_direction_0"), val = tensor<string, []>("bidirectional")];
+            tensor<bool, []> x_batch_first_output_sequence_0 = const()[name = tensor<string, []>("x_batch_first_output_sequence_0"), val = tensor<bool, []>(true)];
+            tensor<string, []> x_batch_first_recurrent_activation_0 = const()[name = tensor<string, []>("x_batch_first_recurrent_activation_0"), val = tensor<string, []>("sigmoid")];
+            tensor<string, []> x_batch_first_cell_activation_0 = const()[name = tensor<string, []>("x_batch_first_cell_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<string, []> x_batch_first_activation_0 = const()[name = tensor<string, []>("x_batch_first_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<fp16, [1, 512]> x_batch_first_lstm_h0_reshaped_to_fp16 = const()[name = tensor<string, []>("x_batch_first_lstm_h0_reshaped_to_fp16"), val = tensor<fp16, [1, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<fp16, [1024, 640]> concat_4_to_fp16 = const()[name = tensor<string, []>("concat_4_to_fp16"), val = tensor<fp16, [1024, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1152)))];
+            tensor<fp16, [1024, 256]> concat_5_to_fp16 = const()[name = tensor<string, []>("concat_5_to_fp16"), val = tensor<fp16, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1311936)))];
+            tensor<fp16, [1024]> add_1_to_fp16 = const()[name = tensor<string, []>("add_1_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1836288)))];
+            tensor<fp16, [1024, 640]> concat_6_to_fp16 = const()[name = tensor<string, []>("concat_6_to_fp16"), val = tensor<fp16, [1024, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1838400)))];
+            tensor<fp16, [1024, 256]> concat_7_to_fp16 = const()[name = tensor<string, []>("concat_7_to_fp16"), val = tensor<fp16, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3149184)))];
+            tensor<fp16, [1024]> add_2_to_fp16 = const()[name = tensor<string, []>("add_2_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3673536)))];
+            tensor<fp16, [1, 640, ?]> en_to_fp16 = cast(dtype = en_to_fp16_dtype_0, x = en)[name = tensor<string, []>("cast_27")];
+            tensor<fp16, [?, 1, 640]> transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = en_to_fp16)[name = tensor<string, []>("transpose_3")];
+            tensor<fp16, [?, 1, 512]> x_batch_first_cast_fp16_0, tensor<fp16, [1, 512]> x_batch_first_cast_fp16_1, tensor<fp16, [1, 512]> x_batch_first_cast_fp16_2 = lstm(activation = x_batch_first_activation_0, bias = add_1_to_fp16, bias_back = add_2_to_fp16, cell_activation = x_batch_first_cell_activation_0, direction = x_batch_first_direction_0, initial_c = x_batch_first_lstm_h0_reshaped_to_fp16, initial_h = x_batch_first_lstm_h0_reshaped_to_fp16, output_sequence = x_batch_first_output_sequence_0, recurrent_activation = x_batch_first_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_hh_back = concat_7_to_fp16, weight_ih = concat_4_to_fp16, weight_ih_back = concat_6_to_fp16, x = transpose_0_cast_fp16)[name = tensor<string, []>("x_batch_first_cast_fp16")];
+            tensor<int32, [3]> x_perm_0 = const()[name = tensor<string, []>("x_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = tensor<string, []>("input_3_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
+            tensor<fp32, []> var_53 = const()[name = tensor<string, []>("op_53"), val = tensor<fp32, []>(0x1.99999ap-3)];
+            tensor<string, []> s_to_fp16_dtype_0 = const()[name = tensor<string, []>("s_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1024, 128]> F0_0_norm1_fc_weight_to_fp16 = const()[name = tensor<string, []>("F0_0_norm1_fc_weight_to_fp16"), val = tensor<fp16, [1024, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3675648)))];
+            tensor<fp16, [1024]> F0_0_norm1_fc_bias_to_fp16 = const()[name = tensor<string, []>("F0_0_norm1_fc_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3937856)))];
+            tensor<fp16, [1, 128]> s_to_fp16 = cast(dtype = s_to_fp16_dtype_0, x = s)[name = tensor<string, []>("cast_26")];
+            tensor<fp16, [1, 1024]> linear_0_cast_fp16 = linear(bias = F0_0_norm1_fc_bias_to_fp16, weight = F0_0_norm1_fc_weight_to_fp16, x = s_to_fp16)[name = tensor<string, []>("linear_0_cast_fp16")];
+            tensor<int32, [3]> var_78 = const()[name = tensor<string, []>("op_78"), val = tensor<int32, [3]>([1, 1024, 1])];
+            tensor<fp16, [1, 1024, 1]> h_3_cast_fp16 = reshape(shape = var_78, x = linear_0_cast_fp16)[name = tensor<string, []>("h_3_cast_fp16")];
+            tensor<int32, [2]> var_80_split_sizes_0 = const()[name = tensor<string, []>("op_80_split_sizes_0"), val = tensor<int32, [2]>([512, 512])];
+            tensor<int32, []> var_80_axis_0 = const()[name = tensor<string, []>("op_80_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 1]> var_80_cast_fp16_0, tensor<fp16, [1, 512, 1]> var_80_cast_fp16_1 = split(axis = var_80_axis_0, split_sizes = var_80_split_sizes_0, x = h_3_cast_fp16)[name = tensor<string, []>("op_80_cast_fp16")];
+            tensor<fp16, []> var_82_promoted_to_fp16 = const()[name = tensor<string, []>("op_82_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1, 512, 1]> var_83_cast_fp16 = add(x = var_80_cast_fp16_0, y = var_82_promoted_to_fp16)[name = tensor<string, []>("op_83_cast_fp16")];
+            tensor<fp16, []> var_56_to_fp16 = const()[name = tensor<string, []>("op_56_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, ?, 512]> x_cast_fp16 = transpose(perm = x_perm_0, x = x_batch_first_cast_fp16_0)[name = tensor<string, []>("transpose_2")];
+            tensor<fp16, [1, 512, ?]> input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = x_cast_fp16)[name = tensor<string, []>("transpose_1")];
+            tensor<fp16, [1, 512, ?]> var_84_cast_fp16 = instance_norm(epsilon = var_56_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("op_84_cast_fp16")];
+            tensor<fp16, [1, 512, ?]> var_85_cast_fp16 = mul(x = var_83_cast_fp16, y = var_84_cast_fp16)[name = tensor<string, []>("op_85_cast_fp16")];
+            tensor<fp16, [1, 512, ?]> input_5_cast_fp16 = add(x = var_85_cast_fp16, y = var_80_cast_fp16_1)[name = tensor<string, []>("input_5_cast_fp16")];
+            tensor<fp16, [1, 512, ?]> input_7_cast_fp16 = leaky_relu(alpha = var_53, x = input_5_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
+            tensor<string, []> input_11_pad_type_0 = const()[name = tensor<string, []>("input_11_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [2]> input_11_pad_0 = const()[name = tensor<string, []>("input_11_pad_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [1]> input_11_strides_0 = const()[name = tensor<string, []>("input_11_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> input_11_dilations_0 = const()[name = tensor<string, []>("input_11_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_11_groups_0 = const()[name = tensor<string, []>("input_11_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 3]> weight_1_to_fp16 = const()[name = tensor<string, []>("weight_1_to_fp16"), val = tensor<fp16, [512, 512, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3939968)))];
+            tensor<fp16, [512]> F0_0_conv1_bias_to_fp16 = const()[name = tensor<string, []>("F0_0_conv1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5512896)))];
+            tensor<fp16, [1, 512, ?]> input_11_cast_fp16 = conv(bias = F0_0_conv1_bias_to_fp16, dilations = input_11_dilations_0, groups = input_11_groups_0, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = input_11_strides_0, weight = weight_1_to_fp16, x = input_7_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")];
+            tensor<fp16, [1024, 128]> F0_0_norm2_fc_weight_to_fp16 = const()[name = tensor<string, []>("F0_0_norm2_fc_weight_to_fp16"), val = tensor<fp16, [1024, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5513984)))];
+            tensor<fp16, [1024]> F0_0_norm2_fc_bias_to_fp16 = const()[name = tensor<string, []>("F0_0_norm2_fc_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5776192)))];
+            tensor<fp16, [1, 1024]> linear_1_cast_fp16 = linear(bias = F0_0_norm2_fc_bias_to_fp16, weight = F0_0_norm2_fc_weight_to_fp16, x = s_to_fp16)[name = tensor<string, []>("linear_1_cast_fp16")];
+            tensor<int32, [3]> var_102 = const()[name = tensor<string, []>("op_102"), val = tensor<int32, [3]>([1, 1024, 1])];
+            tensor<fp16, [1, 1024, 1]> h_7_cast_fp16 = reshape(shape = var_102, x = linear_1_cast_fp16)[name = tensor<string, []>("h_7_cast_fp16")];
+            tensor<int32, [2]> var_104_split_sizes_0 = const()[name = tensor<string, []>("op_104_split_sizes_0"), val = tensor<int32, [2]>([512, 512])];
+            tensor<int32, []> var_104_axis_0 = const()[name = tensor<string, []>("op_104_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 1]> var_104_cast_fp16_0, tensor<fp16, [1, 512, 1]> var_104_cast_fp16_1 = split(axis = var_104_axis_0, split_sizes = var_104_split_sizes_0, x = h_7_cast_fp16)[name = tensor<string, []>("op_104_cast_fp16")];
+            tensor<fp16, []> var_106_promoted_to_fp16 = const()[name = tensor<string, []>("op_106_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1, 512, 1]> var_107_cast_fp16 = add(x = var_104_cast_fp16_0, y = var_106_promoted_to_fp16)[name = tensor<string, []>("op_107_cast_fp16")];
+            tensor<fp16, [1, 512, ?]> var_108_cast_fp16 = instance_norm(epsilon = var_56_to_fp16, x = input_11_cast_fp16)[name = tensor<string, []>("op_108_cast_fp16")];
+            tensor<fp16, [1, 512, ?]> var_109_cast_fp16 = mul(x = var_107_cast_fp16, y = var_108_cast_fp16)[name = tensor<string, []>("op_109_cast_fp16")];
+            tensor<fp16, [1, 512, ?]> input_13_cast_fp16 = add(x = var_109_cast_fp16, y = var_104_cast_fp16_1)[name = tensor<string, []>("input_13_cast_fp16")];
+            tensor<fp16, [1, 512, ?]> input_15_cast_fp16 = leaky_relu(alpha = var_53, x = input_13_cast_fp16)[name = tensor<string, []>("input_15_cast_fp16")];
+            tensor<string, []> out_1_pad_type_0 = const()[name = tensor<string, []>("out_1_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [2]> out_1_pad_0 = const()[name = tensor<string, []>("out_1_pad_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [1]> out_1_strides_0 = const()[name = tensor<string, []>("out_1_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> out_1_dilations_0 = const()[name = tensor<string, []>("out_1_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> out_1_groups_0 = const()[name = tensor<string, []>("out_1_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 3]> weight_3_to_fp16 = const()[name = tensor<string, []>("weight_3_to_fp16"), val = tensor<fp16, [512, 512, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5778304)))];
+            tensor<fp16, [512]> F0_0_conv2_bias_to_fp16 = const()[name = tensor<string, []>("F0_0_conv2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7351232)))];
+            tensor<fp16, [1, 512, ?]> out_1_cast_fp16 = conv(bias = F0_0_conv2_bias_to_fp16, dilations = out_1_dilations_0, groups = out_1_groups_0, pad = out_1_pad_0, pad_type = out_1_pad_type_0, strides = out_1_strides_0, weight = weight_3_to_fp16, x = input_15_cast_fp16)[name = tensor<string, []>("out_1_cast_fp16")];
+            tensor<fp16, [1, 512, ?]> var_120_cast_fp16 = add(x = out_1_cast_fp16, y = input_3_cast_fp16)[name = tensor<string, []>("op_120_cast_fp16")];
+            tensor<fp16, []> _inversed_input_19_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_input_19_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-1)];
+            tensor<fp16, [1, 512, ?]> _inversed_input_19_cast_fp16 = mul(x = var_120_cast_fp16, y = _inversed_input_19_y_0_to_fp16)[name = tensor<string, []>("_inversed_input_19_cast_fp16")];
+            tensor<fp32, []> var_126 = const()[name = tensor<string, []>("op_126"), val = tensor<fp32, []>(0x1.99999ap-3)];
+            tensor<fp16, [1024, 128]> F0_1_norm1_fc_weight_to_fp16 = const()[name = tensor<string, []>("F0_1_norm1_fc_weight_to_fp16"), val = tensor<fp16, [1024, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7352320)))];
+            tensor<fp16, [1024]> F0_1_norm1_fc_bias_to_fp16 = const()[name = tensor<string, []>("F0_1_norm1_fc_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7614528)))];
+            tensor<fp16, [1, 1024]> linear_2_cast_fp16 = linear(bias = F0_1_norm1_fc_bias_to_fp16, weight = F0_1_norm1_fc_weight_to_fp16, x = s_to_fp16)[name = tensor<string, []>("linear_2_cast_fp16")];
+            tensor<int32, [3]> var_161 = const()[name = tensor<string, []>("op_161"), val = tensor<int32, [3]>([1, 1024, 1])];
+            tensor<fp16, [1, 1024, 1]> h_11_cast_fp16 = reshape(shape = var_161, x = linear_2_cast_fp16)[name = tensor<string, []>("h_11_cast_fp16")];
+            tensor<int32, [2]> var_163_split_sizes_0 = const()[name = tensor<string, []>("op_163_split_sizes_0"), val = tensor<int32, [2]>([512, 512])];
+            tensor<int32, []> var_163_axis_0 = const()[name = tensor<string, []>("op_163_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 1]> var_163_cast_fp16_0, tensor<fp16, [1, 512, 1]> var_163_cast_fp16_1 = split(axis = var_163_axis_0, split_sizes = var_163_split_sizes_0, x = h_11_cast_fp16)[name = tensor<string, []>("op_163_cast_fp16")];
+            tensor<fp16, []> var_165_promoted_to_fp16 = const()[name = tensor<string, []>("op_165_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1, 512, 1]> var_166_cast_fp16 = add(x = var_163_cast_fp16_0, y = var_165_promoted_to_fp16)[name = tensor<string, []>("op_166_cast_fp16")];
+            tensor<fp16, []> var_130_to_fp16 = const()[name = tensor<string, []>("op_130_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 512, ?]> var_167_cast_fp16 = instance_norm(epsilon = var_130_to_fp16, x = _inversed_input_19_cast_fp16)[name = tensor<string, []>("op_167_cast_fp16")];
+            tensor<fp16, [1, 512, ?]> var_168_cast_fp16 = mul(x = var_166_cast_fp16, y = var_167_cast_fp16)[name = tensor<string, []>("op_168_cast_fp16")];
+            tensor<fp16, [1, 512, ?]> input_21_cast_fp16 = add(x = var_168_cast_fp16, y = var_163_cast_fp16_1)[name = tensor<string, []>("input_21_cast_fp16")];
+            tensor<fp16, [1, 512, ?]> input_23_cast_fp16 = leaky_relu(alpha = var_126, x = input_21_cast_fp16)[name = tensor<string, []>("input_23_cast_fp16")];
+            tensor<string, []> conv_transpose_0_pad_type_0 = const()[name = tensor<string, []>("conv_transpose_0_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [2]> conv_transpose_0_pad_0 = const()[name = tensor<string, []>("conv_transpose_0_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> conv_transpose_0_strides_0 = const()[name = tensor<string, []>("conv_transpose_0_strides_0"), val = tensor<int32, [1]>([2])];
+            tensor<int32, []> conv_transpose_0_groups_0 = const()[name = tensor<string, []>("conv_transpose_0_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> conv_transpose_0_dilations_0 = const()[name = tensor<string, []>("conv_transpose_0_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 3]> var_171_to_fp16 = const()[name = tensor<string, []>("op_171_to_fp16"), val = tensor<fp16, [512, 1, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7616640)))];
+            tensor<fp16, [512]> F0_1_pool_bias_to_fp16 = const()[name = tensor<string, []>("F0_1_pool_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7619776)))];
+            tensor<fp16, [1, 512, ?]> conv_transpose_0_cast_fp16 = conv_transpose(bias = F0_1_pool_bias_to_fp16, dilations = conv_transpose_0_dilations_0, groups = conv_transpose_0_groups_0, pad = conv_transpose_0_pad_0, pad_type = conv_transpose_0_pad_type_0, strides = conv_transpose_0_strides_0, weight = var_171_to_fp16, x = input_23_cast_fp16)[name = tensor<string, []>("conv_transpose_0_cast_fp16")];
+            tensor<int32, [3]> input_25_begin_0 = const()[name = tensor<string, []>("input_25_begin_0"), val = tensor<int32, [3]>([0, 0, 1])];
+            tensor<int32, [3]> input_25_end_0 = const()[name = tensor<string, []>("input_25_end_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<bool, [3]> input_25_begin_mask_0 = const()[name = tensor<string, []>("input_25_begin_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<bool, [3]> input_25_end_mask_0 = const()[name = tensor<string, []>("input_25_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, ?]> input_25_cast_fp16 = slice_by_index(begin = input_25_begin_0, begin_mask = input_25_begin_mask_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = conv_transpose_0_cast_fp16)[name = tensor<string, []>("input_25_cast_fp16")];
+            tensor<string, []> input_29_pad_type_0 = const()[name = tensor<string, []>("input_29_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [2]> input_29_pad_0 = const()[name = tensor<string, []>("input_29_pad_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [1]> input_29_strides_0 = const()[name = tensor<string, []>("input_29_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> input_29_dilations_0 = const()[name = tensor<string, []>("input_29_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_29_groups_0 = const()[name = tensor<string, []>("input_29_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [256, 512, 3]> weight_5_to_fp16 = const()[name = tensor<string, []>("weight_5_to_fp16"), val = tensor<fp16, [256, 512, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7620864)))];
+            tensor<fp16, [256]> F0_1_conv1_bias_to_fp16 = const()[name = tensor<string, []>("F0_1_conv1_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8407360)))];
+            tensor<fp16, [1, 256, ?]> input_29_cast_fp16 = conv(bias = F0_1_conv1_bias_to_fp16, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = weight_5_to_fp16, x = input_25_cast_fp16)[name = tensor<string, []>("input_29_cast_fp16")];
+            tensor<fp16, [512, 128]> F0_1_norm2_fc_weight_to_fp16 = const()[name = tensor<string, []>("F0_1_norm2_fc_weight_to_fp16"), val = tensor<fp16, [512, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8407936)))];
+            tensor<fp16, [512]> F0_1_norm2_fc_bias_to_fp16 = const()[name = tensor<string, []>("F0_1_norm2_fc_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8539072)))];
+            tensor<fp16, [1, 512]> linear_3_cast_fp16 = linear(bias = F0_1_norm2_fc_bias_to_fp16, weight = F0_1_norm2_fc_weight_to_fp16, x = s_to_fp16)[name = tensor<string, []>("linear_3_cast_fp16")];
+            tensor<int32, [3]> var_192 = const()[name = tensor<string, []>("op_192"), val = tensor<int32, [3]>([1, 512, 1])];
+            tensor<fp16, [1, 512, 1]> h_15_cast_fp16 = reshape(shape = var_192, x = linear_3_cast_fp16)[name = tensor<string, []>("h_15_cast_fp16")];
+            tensor<int32, [2]> var_194_split_sizes_0 = const()[name = tensor<string, []>("op_194_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
+            tensor<int32, []> var_194_axis_0 = const()[name = tensor<string, []>("op_194_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 256, 1]> var_194_cast_fp16_0, tensor<fp16, [1, 256, 1]> var_194_cast_fp16_1 = split(axis = var_194_axis_0, split_sizes = var_194_split_sizes_0, x = h_15_cast_fp16)[name = tensor<string, []>("op_194_cast_fp16")];
+            tensor<fp16, []> var_196_promoted_to_fp16 = const()[name = tensor<string, []>("op_196_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1, 256, 1]> var_197_cast_fp16 = add(x = var_194_cast_fp16_0, y = var_196_promoted_to_fp16)[name = tensor<string, []>("op_197_cast_fp16")];
+            tensor<fp16, [1, 256, ?]> var_198_cast_fp16 = instance_norm(epsilon = var_130_to_fp16, x = input_29_cast_fp16)[name = tensor<string, []>("op_198_cast_fp16")];
+            tensor<fp16, [1, 256, ?]> var_199_cast_fp16 = mul(x = var_197_cast_fp16, y = var_198_cast_fp16)[name = tensor<string, []>("op_199_cast_fp16")];
+            tensor<fp16, [1, 256, ?]> input_31_cast_fp16 = add(x = var_199_cast_fp16, y = var_194_cast_fp16_1)[name = tensor<string, []>("input_31_cast_fp16")];
+            tensor<fp16, [1, 256, ?]> input_33_cast_fp16 = leaky_relu(alpha = var_126, x = input_31_cast_fp16)[name = tensor<string, []>("input_33_cast_fp16")];
+            tensor<string, []> out_3_pad_type_0 = const()[name = tensor<string, []>("out_3_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [2]> out_3_pad_0 = const()[name = tensor<string, []>("out_3_pad_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [1]> out_3_strides_0 = const()[name = tensor<string, []>("out_3_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> out_3_dilations_0 = const()[name = tensor<string, []>("out_3_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> out_3_groups_0 = const()[name = tensor<string, []>("out_3_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [256, 256, 3]> weight_7_to_fp16 = const()[name = tensor<string, []>("weight_7_to_fp16"), val = tensor<fp16, [256, 256, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8540160)))];
+            tensor<fp16, [256]> F0_1_conv2_bias_to_fp16 = const()[name = tensor<string, []>("F0_1_conv2_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8933440)))];
+            tensor<fp16, [1, 256, ?]> out_3_cast_fp16 = conv(bias = F0_1_conv2_bias_to_fp16, dilations = out_3_dilations_0, groups = out_3_groups_0, pad = out_3_pad_0, pad_type = out_3_pad_type_0, strides = out_3_strides_0, weight = weight_7_to_fp16, x = input_33_cast_fp16)[name = tensor<string, []>("out_3_cast_fp16")];
+            tensor<int32, [1]> expand_dims_0_axes_0 = const()[name = tensor<string, []>("expand_dims_0_axes_0"), val = tensor<int32, [1]>([3])];
+            tensor<fp16, [1, 512, ?, 1]> expand_dims_0_cast_fp16 = expand_dims(axes = expand_dims_0_axes_0, x = _inversed_input_19_cast_fp16)[name = tensor<string, []>("expand_dims_0_cast_fp16")];
+            tensor<int32, []> upsample_nearest_neighbor_0_scale_factor_height_0 = const()[name = tensor<string, []>("upsample_nearest_neighbor_0_scale_factor_height_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> upsample_nearest_neighbor_0_scale_factor_width_0 = const()[name = tensor<string, []>("upsample_nearest_neighbor_0_scale_factor_width_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, ?, 1]> upsample_nearest_neighbor_0_cast_fp16 = upsample_nearest_neighbor(scale_factor_height = upsample_nearest_neighbor_0_scale_factor_height_0, scale_factor_width = upsample_nearest_neighbor_0_scale_factor_width_0, x = expand_dims_0_cast_fp16)[name = tensor<string, []>("upsample_nearest_neighbor_0_cast_fp16")];
+            tensor<int32, [1]> input_37_axes_0 = const()[name = tensor<string, []>("input_37_axes_0"), val = tensor<int32, [1]>([3])];
+            tensor<fp16, [1, 512, ?]> input_37_cast_fp16 = squeeze(axes = input_37_axes_0, x = upsample_nearest_neighbor_0_cast_fp16)[name = tensor<string, []>("input_37_cast_fp16")];
+            tensor<string, []> var_217_pad_type_0 = const()[name = tensor<string, []>("op_217_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> var_217_strides_0 = const()[name = tensor<string, []>("op_217_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> var_217_pad_0 = const()[name = tensor<string, []>("op_217_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> var_217_dilations_0 = const()[name = tensor<string, []>("op_217_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> var_217_groups_0 = const()[name = tensor<string, []>("op_217_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [256, 512, 1]> weight_9_to_fp16 = const()[name = tensor<string, []>("weight_9_to_fp16"), val = tensor<fp16, [256, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8934016)))];
+            tensor<fp16, [1, 256, ?]> var_217_cast_fp16 = conv(dilations = var_217_dilations_0, groups = var_217_groups_0, pad = var_217_pad_0, pad_type = var_217_pad_type_0, strides = var_217_strides_0, weight = weight_9_to_fp16, x = input_37_cast_fp16)[name = tensor<string, []>("op_217_cast_fp16")];
+            tensor<fp16, [1, 256, ?]> var_218_cast_fp16 = add(x = out_3_cast_fp16, y = var_217_cast_fp16)[name = tensor<string, []>("op_218_cast_fp16")];
+            tensor<fp16, []> _inversed_input_39_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_input_39_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-1)];
+            tensor<fp16, [1, 256, ?]> _inversed_input_39_cast_fp16 = mul(x = var_218_cast_fp16, y = _inversed_input_39_y_0_to_fp16)[name = tensor<string, []>("_inversed_input_39_cast_fp16")];
+            tensor<fp32, []> var_222 = const()[name = tensor<string, []>("op_222"), val = tensor<fp32, []>(0x1.99999ap-3)];
+            tensor<fp16, [512, 128]> F0_2_norm1_fc_weight_to_fp16 = const()[name = tensor<string, []>("F0_2_norm1_fc_weight_to_fp16"), val = tensor<fp16, [512, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9196224)))];
+            tensor<fp16, [512]> F0_2_norm1_fc_bias_to_fp16 = const()[name = tensor<string, []>("F0_2_norm1_fc_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9327360)))];
+            tensor<fp16, [1, 512]> linear_4_cast_fp16 = linear(bias = F0_2_norm1_fc_bias_to_fp16, weight = F0_2_norm1_fc_weight_to_fp16, x = s_to_fp16)[name = tensor<string, []>("linear_4_cast_fp16")];
+            tensor<int32, [3]> var_247 = const()[name = tensor<string, []>("op_247"), val = tensor<int32, [3]>([1, 512, 1])];
+            tensor<fp16, [1, 512, 1]> h_19_cast_fp16 = reshape(shape = var_247, x = linear_4_cast_fp16)[name = tensor<string, []>("h_19_cast_fp16")];
+            tensor<int32, [2]> var_249_split_sizes_0 = const()[name = tensor<string, []>("op_249_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
+            tensor<int32, []> var_249_axis_0 = const()[name = tensor<string, []>("op_249_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 256, 1]> var_249_cast_fp16_0, tensor<fp16, [1, 256, 1]> var_249_cast_fp16_1 = split(axis = var_249_axis_0, split_sizes = var_249_split_sizes_0, x = h_19_cast_fp16)[name = tensor<string, []>("op_249_cast_fp16")];
+            tensor<fp16, []> var_251_promoted_to_fp16 = const()[name = tensor<string, []>("op_251_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1, 256, 1]> var_252_cast_fp16 = add(x = var_249_cast_fp16_0, y = var_251_promoted_to_fp16)[name = tensor<string, []>("op_252_cast_fp16")];
+            tensor<fp16, []> var_225_to_fp16 = const()[name = tensor<string, []>("op_225_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 256, ?]> var_253_cast_fp16 = instance_norm(epsilon = var_225_to_fp16, x = _inversed_input_39_cast_fp16)[name = tensor<string, []>("op_253_cast_fp16")];
+            tensor<fp16, [1, 256, ?]> var_254_cast_fp16 = mul(x = var_252_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_254_cast_fp16")];
+            tensor<fp16, [1, 256, ?]> input_41_cast_fp16 = add(x = var_254_cast_fp16, y = var_249_cast_fp16_1)[name = tensor<string, []>("input_41_cast_fp16")];
+            tensor<fp16, [1, 256, ?]> input_43_cast_fp16 = leaky_relu(alpha = var_222, x = input_41_cast_fp16)[name = tensor<string, []>("input_43_cast_fp16")];
+            tensor<string, []> input_47_pad_type_0 = const()[name = tensor<string, []>("input_47_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [2]> input_47_pad_0 = const()[name = tensor<string, []>("input_47_pad_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [1]> input_47_strides_0 = const()[name = tensor<string, []>("input_47_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> input_47_dilations_0 = const()[name = tensor<string, []>("input_47_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_47_groups_0 = const()[name = tensor<string, []>("input_47_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [256, 256, 3]> weight_11_to_fp16 = const()[name = tensor<string, []>("weight_11_to_fp16"), val = tensor<fp16, [256, 256, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9328448)))];
+            tensor<fp16, [256]> F0_2_conv1_bias_to_fp16 = const()[name = tensor<string, []>("F0_2_conv1_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9721728)))];
+            tensor<fp16, [1, 256, ?]> input_47_cast_fp16 = conv(bias = F0_2_conv1_bias_to_fp16, dilations = input_47_dilations_0, groups = input_47_groups_0, pad = input_47_pad_0, pad_type = input_47_pad_type_0, strides = input_47_strides_0, weight = weight_11_to_fp16, x = input_43_cast_fp16)[name = tensor<string, []>("input_47_cast_fp16")];
+            tensor<fp16, [512, 128]> F0_2_norm2_fc_weight_to_fp16 = const()[name = tensor<string, []>("F0_2_norm2_fc_weight_to_fp16"), val = tensor<fp16, [512, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9722304)))];
+            tensor<fp16, [512]> F0_2_norm2_fc_bias_to_fp16 = const()[name = tensor<string, []>("F0_2_norm2_fc_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9853440)))];
+            tensor<fp16, [1, 512]> linear_5_cast_fp16 = linear(bias = F0_2_norm2_fc_bias_to_fp16, weight = F0_2_norm2_fc_weight_to_fp16, x = s_to_fp16)[name = tensor<string, []>("linear_5_cast_fp16")];
+            tensor<int32, [3]> var_271 = const()[name = tensor<string, []>("op_271"), val = tensor<int32, [3]>([1, 512, 1])];
+            tensor<fp16, [1, 512, 1]> h_23_cast_fp16 = reshape(shape = var_271, x = linear_5_cast_fp16)[name = tensor<string, []>("h_23_cast_fp16")];
+            tensor<int32, [2]> var_273_split_sizes_0 = const()[name = tensor<string, []>("op_273_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
+            tensor<int32, []> var_273_axis_0 = const()[name = tensor<string, []>("op_273_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 256, 1]> var_273_cast_fp16_0, tensor<fp16, [1, 256, 1]> var_273_cast_fp16_1 = split(axis = var_273_axis_0, split_sizes = var_273_split_sizes_0, x = h_23_cast_fp16)[name = tensor<string, []>("op_273_cast_fp16")];
+            tensor<fp16, []> var_275_promoted_to_fp16 = const()[name = tensor<string, []>("op_275_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1, 256, 1]> var_276_cast_fp16 = add(x = var_273_cast_fp16_0, y = var_275_promoted_to_fp16)[name = tensor<string, []>("op_276_cast_fp16")];
+            tensor<fp16, [1, 256, ?]> var_277_cast_fp16 = instance_norm(epsilon = var_225_to_fp16, x = input_47_cast_fp16)[name = tensor<string, []>("op_277_cast_fp16")];
+            tensor<fp16, [1, 256, ?]> var_278_cast_fp16 = mul(x = var_276_cast_fp16, y = var_277_cast_fp16)[name = tensor<string, []>("op_278_cast_fp16")];
+            tensor<fp16, [1, 256, ?]> input_49_cast_fp16 = add(x = var_278_cast_fp16, y = var_273_cast_fp16_1)[name = tensor<string, []>("input_49_cast_fp16")];
+            tensor<fp16, [1, 256, ?]> input_51_cast_fp16 = leaky_relu(alpha = var_222, x = input_49_cast_fp16)[name = tensor<string, []>("input_51_cast_fp16")];
+            tensor<string, []> out_5_pad_type_0 = const()[name = tensor<string, []>("out_5_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [2]> out_5_pad_0 = const()[name = tensor<string, []>("out_5_pad_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [1]> out_5_strides_0 = const()[name = tensor<string, []>("out_5_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> out_5_dilations_0 = const()[name = tensor<string, []>("out_5_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> out_5_groups_0 = const()[name = tensor<string, []>("out_5_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [256, 256, 3]> weight_13_to_fp16 = const()[name = tensor<string, []>("weight_13_to_fp16"), val = tensor<fp16, [256, 256, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9854528)))];
+            tensor<fp16, [256]> F0_2_conv2_bias_to_fp16 = const()[name = tensor<string, []>("F0_2_conv2_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10247808)))];
+            tensor<fp16, [1, 256, ?]> out_5_cast_fp16 = conv(bias = F0_2_conv2_bias_to_fp16, dilations = out_5_dilations_0, groups = out_5_groups_0, pad = out_5_pad_0, pad_type = out_5_pad_type_0, strides = out_5_strides_0, weight = weight_13_to_fp16, x = input_51_cast_fp16)[name = tensor<string, []>("out_5_cast_fp16")];
+            tensor<fp16, [1, 256, ?]> var_289_cast_fp16 = add(x = out_5_cast_fp16, y = _inversed_input_39_cast_fp16)[name = tensor<string, []>("op_289_cast_fp16")];
+            tensor<fp16, []> _inversed_input_55_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_input_55_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-1)];
+            tensor<fp16, [1, 256, ?]> _inversed_input_55_cast_fp16 = mul(x = var_289_cast_fp16, y = _inversed_input_55_y_0_to_fp16)[name = tensor<string, []>("_inversed_input_55_cast_fp16")];
+            tensor<string, []> var_302_pad_type_0 = const()[name = tensor<string, []>("op_302_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> var_302_strides_0 = const()[name = tensor<string, []>("op_302_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> var_302_pad_0 = const()[name = tensor<string, []>("op_302_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> var_302_dilations_0 = const()[name = tensor<string, []>("op_302_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> var_302_groups_0 = const()[name = tensor<string, []>("op_302_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 256, 1]> F0_proj_weight_to_fp16 = const()[name = tensor<string, []>("F0_proj_weight_to_fp16"), val = tensor<fp16, [1, 256, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10248384)))];
+            tensor<fp16, [1]> F0_proj_bias_to_fp16 = const()[name = tensor<string, []>("F0_proj_bias_to_fp16"), val = tensor<fp16, [1]>([0x1.adp-4])];
+            tensor<fp16, [1, 1, ?]> var_302_cast_fp16 = conv(bias = F0_proj_bias_to_fp16, dilations = var_302_dilations_0, groups = var_302_groups_0, pad = var_302_pad_0, pad_type = var_302_pad_type_0, strides = var_302_strides_0, weight = F0_proj_weight_to_fp16, x = _inversed_input_55_cast_fp16)[name = tensor<string, []>("op_302_cast_fp16")];
+            tensor<int32, [1]> var_304_axes_0 = const()[name = tensor<string, []>("op_304_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, ?]> F0 = squeeze(axes = var_304_axes_0, x = var_302_cast_fp16)[name = tensor<string, []>("op_304_cast_fp16")];
+            tensor<fp32, []> var_309 = const()[name = tensor<string, []>("op_309"), val = tensor<fp32, []>(0x1.99999ap-3)];
+            tensor<fp16, [1024, 128]> N_0_norm1_fc_weight_to_fp16 = const()[name = tensor<string, []>("N_0_norm1_fc_weight_to_fp16"), val = tensor<fp16, [1024, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10248960)))];
+            tensor<fp16, [1024]> N_0_norm1_fc_bias_to_fp16 = const()[name = tensor<string, []>("N_0_norm1_fc_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10511168)))];
+            tensor<fp16, [1, 1024]> linear_6_cast_fp16 = linear(bias = N_0_norm1_fc_bias_to_fp16, weight = N_0_norm1_fc_weight_to_fp16, x = s_to_fp16)[name = tensor<string, []>("linear_6_cast_fp16")];
+            tensor<int32, [3]> var_334 = const()[name = tensor<string, []>("op_334"), val = tensor<int32, [3]>([1, 1024, 1])];
+            tensor<fp16, [1, 1024, 1]> h_27_cast_fp16 = reshape(shape = var_334, x = linear_6_cast_fp16)[name = tensor<string, []>("h_27_cast_fp16")];
+            tensor<int32, [2]> var_336_split_sizes_0 = const()[name = tensor<string, []>("op_336_split_sizes_0"), val = tensor<int32, [2]>([512, 512])];
+            tensor<int32, []> var_336_axis_0 = const()[name = tensor<string, []>("op_336_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 1]> var_336_cast_fp16_0, tensor<fp16, [1, 512, 1]> var_336_cast_fp16_1 = split(axis = var_336_axis_0, split_sizes = var_336_split_sizes_0, x = h_27_cast_fp16)[name = tensor<string, []>("op_336_cast_fp16")];
+            tensor<fp16, []> var_338_promoted_to_fp16 = const()[name = tensor<string, []>("op_338_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1, 512, 1]> var_339_cast_fp16 = add(x = var_336_cast_fp16_0, y = var_338_promoted_to_fp16)[name = tensor<string, []>("op_339_cast_fp16")];
+            tensor<fp16, [1, 512, ?]> var_341_cast_fp16 = mul(x = var_339_cast_fp16, y = var_84_cast_fp16)[name = tensor<string, []>("op_341_cast_fp16")];
+            tensor<fp16, [1, 512, ?]> input_59_cast_fp16 = add(x = var_341_cast_fp16, y = var_336_cast_fp16_1)[name = tensor<string, []>("input_59_cast_fp16")];
+            tensor<fp16, [1, 512, ?]> input_61_cast_fp16 = leaky_relu(alpha = var_309, x = input_59_cast_fp16)[name = tensor<string, []>("input_61_cast_fp16")];
+            tensor<string, []> input_65_pad_type_0 = const()[name = tensor<string, []>("input_65_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [2]> input_65_pad_0 = const()[name = tensor<string, []>("input_65_pad_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [1]> input_65_strides_0 = const()[name = tensor<string, []>("input_65_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> input_65_dilations_0 = const()[name = tensor<string, []>("input_65_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_65_groups_0 = const()[name = tensor<string, []>("input_65_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 3]> weight_17_to_fp16 = const()[name = tensor<string, []>("weight_17_to_fp16"), val = tensor<fp16, [512, 512, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10513280)))];
+            tensor<fp16, [512]> N_0_conv1_bias_to_fp16 = const()[name = tensor<string, []>("N_0_conv1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12086208)))];
+            tensor<fp16, [1, 512, ?]> input_65_cast_fp16 = conv(bias = N_0_conv1_bias_to_fp16, dilations = input_65_dilations_0, groups = input_65_groups_0, pad = input_65_pad_0, pad_type = input_65_pad_type_0, strides = input_65_strides_0, weight = weight_17_to_fp16, x = input_61_cast_fp16)[name = tensor<string, []>("input_65_cast_fp16")];
+            tensor<fp16, [1024, 128]> N_0_norm2_fc_weight_to_fp16 = const()[name = tensor<string, []>("N_0_norm2_fc_weight_to_fp16"), val = tensor<fp16, [1024, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12087296)))];
+            tensor<fp16, [1024]> N_0_norm2_fc_bias_to_fp16 = const()[name = tensor<string, []>("N_0_norm2_fc_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12349504)))];
+            tensor<fp16, [1, 1024]> linear_7_cast_fp16 = linear(bias = N_0_norm2_fc_bias_to_fp16, weight = N_0_norm2_fc_weight_to_fp16, x = s_to_fp16)[name = tensor<string, []>("linear_7_cast_fp16")];
+            tensor<int32, [3]> var_358 = const()[name = tensor<string, []>("op_358"), val = tensor<int32, [3]>([1, 1024, 1])];
+            tensor<fp16, [1, 1024, 1]> h_31_cast_fp16 = reshape(shape = var_358, x = linear_7_cast_fp16)[name = tensor<string, []>("h_31_cast_fp16")];
+            tensor<int32, [2]> var_360_split_sizes_0 = const()[name = tensor<string, []>("op_360_split_sizes_0"), val = tensor<int32, [2]>([512, 512])];
+            tensor<int32, []> var_360_axis_0 = const()[name = tensor<string, []>("op_360_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 1]> var_360_cast_fp16_0, tensor<fp16, [1, 512, 1]> var_360_cast_fp16_1 = split(axis = var_360_axis_0, split_sizes = var_360_split_sizes_0, x = h_31_cast_fp16)[name = tensor<string, []>("op_360_cast_fp16")];
+            tensor<fp16, []> var_362_promoted_to_fp16 = const()[name = tensor<string, []>("op_362_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1, 512, 1]> var_363_cast_fp16 = add(x = var_360_cast_fp16_0, y = var_362_promoted_to_fp16)[name = tensor<string, []>("op_363_cast_fp16")];
+            tensor<fp16, []> var_312_to_fp16 = const()[name = tensor<string, []>("op_312_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 512, ?]> var_364_cast_fp16 = instance_norm(epsilon = var_312_to_fp16, x = input_65_cast_fp16)[name = tensor<string, []>("op_364_cast_fp16")];
+            tensor<fp16, [1, 512, ?]> var_365_cast_fp16 = mul(x = var_363_cast_fp16, y = var_364_cast_fp16)[name = tensor<string, []>("op_365_cast_fp16")];
+            tensor<fp16, [1, 512, ?]> input_67_cast_fp16 = add(x = var_365_cast_fp16, y = var_360_cast_fp16_1)[name = tensor<string, []>("input_67_cast_fp16")];
+            tensor<fp16, [1, 512, ?]> input_69_cast_fp16 = leaky_relu(alpha = var_309, x = input_67_cast_fp16)[name = tensor<string, []>("input_69_cast_fp16")];
+            tensor<string, []> out_7_pad_type_0 = const()[name = tensor<string, []>("out_7_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [2]> out_7_pad_0 = const()[name = tensor<string, []>("out_7_pad_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [1]> out_7_strides_0 = const()[name = tensor<string, []>("out_7_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> out_7_dilations_0 = const()[name = tensor<string, []>("out_7_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> out_7_groups_0 = const()[name = tensor<string, []>("out_7_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [512, 512, 3]> weight_19_to_fp16 = const()[name = tensor<string, []>("weight_19_to_fp16"), val = tensor<fp16, [512, 512, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12351616)))];
+            tensor<fp16, [512]> N_0_conv2_bias_to_fp16 = const()[name = tensor<string, []>("N_0_conv2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13924544)))];
+            tensor<fp16, [1, 512, ?]> out_7_cast_fp16 = conv(bias = N_0_conv2_bias_to_fp16, dilations = out_7_dilations_0, groups = out_7_groups_0, pad = out_7_pad_0, pad_type = out_7_pad_type_0, strides = out_7_strides_0, weight = weight_19_to_fp16, x = input_69_cast_fp16)[name = tensor<string, []>("out_7_cast_fp16")];
+            tensor<fp16, [1, 512, ?]> var_376_cast_fp16 = add(x = out_7_cast_fp16, y = input_3_cast_fp16)[name = tensor<string, []>("op_376_cast_fp16")];
+            tensor<fp16, []> _inversed_input_73_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_input_73_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-1)];
+            tensor<fp16, [1, 512, ?]> _inversed_input_73_cast_fp16 = mul(x = var_376_cast_fp16, y = _inversed_input_73_y_0_to_fp16)[name = tensor<string, []>("_inversed_input_73_cast_fp16")];
+            tensor<fp32, []> var_382 = const()[name = tensor<string, []>("op_382"), val = tensor<fp32, []>(0x1.99999ap-3)];
+            tensor<fp16, [1024, 128]> N_1_norm1_fc_weight_to_fp16 = const()[name = tensor<string, []>("N_1_norm1_fc_weight_to_fp16"), val = tensor<fp16, [1024, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13925632)))];
+            tensor<fp16, [1024]> N_1_norm1_fc_bias_to_fp16 = const()[name = tensor<string, []>("N_1_norm1_fc_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14187840)))];
+            tensor<fp16, [1, 1024]> linear_8_cast_fp16 = linear(bias = N_1_norm1_fc_bias_to_fp16, weight = N_1_norm1_fc_weight_to_fp16, x = s_to_fp16)[name = tensor<string, []>("linear_8_cast_fp16")];
+            tensor<int32, [3]> var_417 = const()[name = tensor<string, []>("op_417"), val = tensor<int32, [3]>([1, 1024, 1])];
+            tensor<fp16, [1, 1024, 1]> h_35_cast_fp16 = reshape(shape = var_417, x = linear_8_cast_fp16)[name = tensor<string, []>("h_35_cast_fp16")];
+            tensor<int32, [2]> var_419_split_sizes_0 = const()[name = tensor<string, []>("op_419_split_sizes_0"), val = tensor<int32, [2]>([512, 512])];
+            tensor<int32, []> var_419_axis_0 = const()[name = tensor<string, []>("op_419_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, 1]> var_419_cast_fp16_0, tensor<fp16, [1, 512, 1]> var_419_cast_fp16_1 = split(axis = var_419_axis_0, split_sizes = var_419_split_sizes_0, x = h_35_cast_fp16)[name = tensor<string, []>("op_419_cast_fp16")];
+            tensor<fp16, []> var_421_promoted_to_fp16 = const()[name = tensor<string, []>("op_421_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1, 512, 1]> var_422_cast_fp16 = add(x = var_419_cast_fp16_0, y = var_421_promoted_to_fp16)[name = tensor<string, []>("op_422_cast_fp16")];
+            tensor<fp16, []> var_386_to_fp16 = const()[name = tensor<string, []>("op_386_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 512, ?]> var_423_cast_fp16 = instance_norm(epsilon = var_386_to_fp16, x = _inversed_input_73_cast_fp16)[name = tensor<string, []>("op_423_cast_fp16")];
+            tensor<fp16, [1, 512, ?]> var_424_cast_fp16 = mul(x = var_422_cast_fp16, y = var_423_cast_fp16)[name = tensor<string, []>("op_424_cast_fp16")];
+            tensor<fp16, [1, 512, ?]> input_75_cast_fp16 = add(x = var_424_cast_fp16, y = var_419_cast_fp16_1)[name = tensor<string, []>("input_75_cast_fp16")];
+            tensor<fp16, [1, 512, ?]> input_77_cast_fp16 = leaky_relu(alpha = var_382, x = input_75_cast_fp16)[name = tensor<string, []>("input_77_cast_fp16")];
+            tensor<string, []> conv_transpose_1_pad_type_0 = const()[name = tensor<string, []>("conv_transpose_1_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [2]> conv_transpose_1_pad_0 = const()[name = tensor<string, []>("conv_transpose_1_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> conv_transpose_1_strides_0 = const()[name = tensor<string, []>("conv_transpose_1_strides_0"), val = tensor<int32, [1]>([2])];
+            tensor<int32, []> conv_transpose_1_groups_0 = const()[name = tensor<string, []>("conv_transpose_1_groups_0"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> conv_transpose_1_dilations_0 = const()[name = tensor<string, []>("conv_transpose_1_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [512, 1, 3]> var_427_to_fp16 = const()[name = tensor<string, []>("op_427_to_fp16"), val = tensor<fp16, [512, 1, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14189952)))];
+            tensor<fp16, [512]> N_1_pool_bias_to_fp16 = const()[name = tensor<string, []>("N_1_pool_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14193088)))];
+            tensor<fp16, [1, 512, ?]> conv_transpose_1_cast_fp16 = conv_transpose(bias = N_1_pool_bias_to_fp16, dilations = conv_transpose_1_dilations_0, groups = conv_transpose_1_groups_0, pad = conv_transpose_1_pad_0, pad_type = conv_transpose_1_pad_type_0, strides = conv_transpose_1_strides_0, weight = var_427_to_fp16, x = input_77_cast_fp16)[name = tensor<string, []>("conv_transpose_1_cast_fp16")];
+            tensor<int32, [3]> input_79_begin_0 = const()[name = tensor<string, []>("input_79_begin_0"), val = tensor<int32, [3]>([0, 0, 1])];
+            tensor<int32, [3]> input_79_end_0 = const()[name = tensor<string, []>("input_79_end_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<bool, [3]> input_79_begin_mask_0 = const()[name = tensor<string, []>("input_79_begin_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<bool, [3]> input_79_end_mask_0 = const()[name = tensor<string, []>("input_79_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, ?]> input_79_cast_fp16 = slice_by_index(begin = input_79_begin_0, begin_mask = input_79_begin_mask_0, end = input_79_end_0, end_mask = input_79_end_mask_0, x = conv_transpose_1_cast_fp16)[name = tensor<string, []>("input_79_cast_fp16")];
+            tensor<string, []> input_83_pad_type_0 = const()[name = tensor<string, []>("input_83_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [2]> input_83_pad_0 = const()[name = tensor<string, []>("input_83_pad_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [1]> input_83_strides_0 = const()[name = tensor<string, []>("input_83_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> input_83_dilations_0 = const()[name = tensor<string, []>("input_83_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_83_groups_0 = const()[name = tensor<string, []>("input_83_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [256, 512, 3]> weight_21_to_fp16 = const()[name = tensor<string, []>("weight_21_to_fp16"), val = tensor<fp16, [256, 512, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14194176)))];
+            tensor<fp16, [256]> N_1_conv1_bias_to_fp16 = const()[name = tensor<string, []>("N_1_conv1_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14980672)))];
+            tensor<fp16, [1, 256, ?]> input_83_cast_fp16 = conv(bias = N_1_conv1_bias_to_fp16, dilations = input_83_dilations_0, groups = input_83_groups_0, pad = input_83_pad_0, pad_type = input_83_pad_type_0, strides = input_83_strides_0, weight = weight_21_to_fp16, x = input_79_cast_fp16)[name = tensor<string, []>("input_83_cast_fp16")];
+            tensor<fp16, [512, 128]> N_1_norm2_fc_weight_to_fp16 = const()[name = tensor<string, []>("N_1_norm2_fc_weight_to_fp16"), val = tensor<fp16, [512, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14981248)))];
+            tensor<fp16, [512]> N_1_norm2_fc_bias_to_fp16 = const()[name = tensor<string, []>("N_1_norm2_fc_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15112384)))];
+            tensor<fp16, [1, 512]> linear_9_cast_fp16 = linear(bias = N_1_norm2_fc_bias_to_fp16, weight = N_1_norm2_fc_weight_to_fp16, x = s_to_fp16)[name = tensor<string, []>("linear_9_cast_fp16")];
+            tensor<int32, [3]> var_448 = const()[name = tensor<string, []>("op_448"), val = tensor<int32, [3]>([1, 512, 1])];
+            tensor<fp16, [1, 512, 1]> h_39_cast_fp16 = reshape(shape = var_448, x = linear_9_cast_fp16)[name = tensor<string, []>("h_39_cast_fp16")];
+            tensor<int32, [2]> var_450_split_sizes_0 = const()[name = tensor<string, []>("op_450_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
+            tensor<int32, []> var_450_axis_0 = const()[name = tensor<string, []>("op_450_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 256, 1]> var_450_cast_fp16_0, tensor<fp16, [1, 256, 1]> var_450_cast_fp16_1 = split(axis = var_450_axis_0, split_sizes = var_450_split_sizes_0, x = h_39_cast_fp16)[name = tensor<string, []>("op_450_cast_fp16")];
+            tensor<fp16, []> var_452_promoted_to_fp16 = const()[name = tensor<string, []>("op_452_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1, 256, 1]> var_453_cast_fp16 = add(x = var_450_cast_fp16_0, y = var_452_promoted_to_fp16)[name = tensor<string, []>("op_453_cast_fp16")];
+            tensor<fp16, [1, 256, ?]> var_454_cast_fp16 = instance_norm(epsilon = var_386_to_fp16, x = input_83_cast_fp16)[name = tensor<string, []>("op_454_cast_fp16")];
+            tensor<fp16, [1, 256, ?]> var_455_cast_fp16 = mul(x = var_453_cast_fp16, y = var_454_cast_fp16)[name = tensor<string, []>("op_455_cast_fp16")];
+            tensor<fp16, [1, 256, ?]> input_85_cast_fp16 = add(x = var_455_cast_fp16, y = var_450_cast_fp16_1)[name = tensor<string, []>("input_85_cast_fp16")];
+            tensor<fp16, [1, 256, ?]> input_87_cast_fp16 = leaky_relu(alpha = var_382, x = input_85_cast_fp16)[name = tensor<string, []>("input_87_cast_fp16")];
+            tensor<string, []> out_9_pad_type_0 = const()[name = tensor<string, []>("out_9_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [2]> out_9_pad_0 = const()[name = tensor<string, []>("out_9_pad_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [1]> out_9_strides_0 = const()[name = tensor<string, []>("out_9_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> out_9_dilations_0 = const()[name = tensor<string, []>("out_9_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> out_9_groups_0 = const()[name = tensor<string, []>("out_9_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [256, 256, 3]> weight_23_to_fp16 = const()[name = tensor<string, []>("weight_23_to_fp16"), val = tensor<fp16, [256, 256, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15113472)))];
+            tensor<fp16, [256]> N_1_conv2_bias_to_fp16 = const()[name = tensor<string, []>("N_1_conv2_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15506752)))];
+            tensor<fp16, [1, 256, ?]> out_9_cast_fp16 = conv(bias = N_1_conv2_bias_to_fp16, dilations = out_9_dilations_0, groups = out_9_groups_0, pad = out_9_pad_0, pad_type = out_9_pad_type_0, strides = out_9_strides_0, weight = weight_23_to_fp16, x = input_87_cast_fp16)[name = tensor<string, []>("out_9_cast_fp16")];
+            tensor<int32, [1]> expand_dims_1_axes_0 = const()[name = tensor<string, []>("expand_dims_1_axes_0"), val = tensor<int32, [1]>([3])];
+            tensor<fp16, [1, 512, ?, 1]> expand_dims_1_cast_fp16 = expand_dims(axes = expand_dims_1_axes_0, x = _inversed_input_73_cast_fp16)[name = tensor<string, []>("expand_dims_1_cast_fp16")];
+            tensor<int32, []> upsample_nearest_neighbor_1_scale_factor_height_0 = const()[name = tensor<string, []>("upsample_nearest_neighbor_1_scale_factor_height_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> upsample_nearest_neighbor_1_scale_factor_width_0 = const()[name = tensor<string, []>("upsample_nearest_neighbor_1_scale_factor_width_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 512, ?, 1]> upsample_nearest_neighbor_1_cast_fp16 = upsample_nearest_neighbor(scale_factor_height = upsample_nearest_neighbor_1_scale_factor_height_0, scale_factor_width = upsample_nearest_neighbor_1_scale_factor_width_0, x = expand_dims_1_cast_fp16)[name = tensor<string, []>("upsample_nearest_neighbor_1_cast_fp16")];
+            tensor<int32, [1]> input_91_axes_0 = const()[name = tensor<string, []>("input_91_axes_0"), val = tensor<int32, [1]>([3])];
+            tensor<fp16, [1, 512, ?]> input_91_cast_fp16 = squeeze(axes = input_91_axes_0, x = upsample_nearest_neighbor_1_cast_fp16)[name = tensor<string, []>("input_91_cast_fp16")];
+            tensor<string, []> var_473_pad_type_0 = const()[name = tensor<string, []>("op_473_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> var_473_strides_0 = const()[name = tensor<string, []>("op_473_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> var_473_pad_0 = const()[name = tensor<string, []>("op_473_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> var_473_dilations_0 = const()[name = tensor<string, []>("op_473_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> var_473_groups_0 = const()[name = tensor<string, []>("op_473_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [256, 512, 1]> weight_25_to_fp16 = const()[name = tensor<string, []>("weight_25_to_fp16"), val = tensor<fp16, [256, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15507328)))];
+            tensor<fp16, [1, 256, ?]> var_473_cast_fp16 = conv(dilations = var_473_dilations_0, groups = var_473_groups_0, pad = var_473_pad_0, pad_type = var_473_pad_type_0, strides = var_473_strides_0, weight = weight_25_to_fp16, x = input_91_cast_fp16)[name = tensor<string, []>("op_473_cast_fp16")];
+            tensor<fp16, [1, 256, ?]> var_474_cast_fp16 = add(x = out_9_cast_fp16, y = var_473_cast_fp16)[name = tensor<string, []>("op_474_cast_fp16")];
+            tensor<fp16, []> _inversed_input_93_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_input_93_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-1)];
+            tensor<fp16, [1, 256, ?]> _inversed_input_93_cast_fp16 = mul(x = var_474_cast_fp16, y = _inversed_input_93_y_0_to_fp16)[name = tensor<string, []>("_inversed_input_93_cast_fp16")];
+            tensor<fp32, []> var_478 = const()[name = tensor<string, []>("op_478"), val = tensor<fp32, []>(0x1.99999ap-3)];
+            tensor<fp16, [512, 128]> N_2_norm1_fc_weight_to_fp16 = const()[name = tensor<string, []>("N_2_norm1_fc_weight_to_fp16"), val = tensor<fp16, [512, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15769536)))];
+            tensor<fp16, [512]> N_2_norm1_fc_bias_to_fp16 = const()[name = tensor<string, []>("N_2_norm1_fc_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15900672)))];
+            tensor<fp16, [1, 512]> linear_10_cast_fp16 = linear(bias = N_2_norm1_fc_bias_to_fp16, weight = N_2_norm1_fc_weight_to_fp16, x = s_to_fp16)[name = tensor<string, []>("linear_10_cast_fp16")];
+            tensor<int32, [3]> var_503 = const()[name = tensor<string, []>("op_503"), val = tensor<int32, [3]>([1, 512, 1])];
+            tensor<fp16, [1, 512, 1]> h_43_cast_fp16 = reshape(shape = var_503, x = linear_10_cast_fp16)[name = tensor<string, []>("h_43_cast_fp16")];
+            tensor<int32, [2]> var_505_split_sizes_0 = const()[name = tensor<string, []>("op_505_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
+            tensor<int32, []> var_505_axis_0 = const()[name = tensor<string, []>("op_505_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 256, 1]> var_505_cast_fp16_0, tensor<fp16, [1, 256, 1]> var_505_cast_fp16_1 = split(axis = var_505_axis_0, split_sizes = var_505_split_sizes_0, x = h_43_cast_fp16)[name = tensor<string, []>("op_505_cast_fp16")];
+            tensor<fp16, []> var_507_promoted_to_fp16 = const()[name = tensor<string, []>("op_507_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1, 256, 1]> var_508_cast_fp16 = add(x = var_505_cast_fp16_0, y = var_507_promoted_to_fp16)[name = tensor<string, []>("op_508_cast_fp16")];
+            tensor<fp16, []> var_481_to_fp16 = const()[name = tensor<string, []>("op_481_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 256, ?]> var_509_cast_fp16 = instance_norm(epsilon = var_481_to_fp16, x = _inversed_input_93_cast_fp16)[name = tensor<string, []>("op_509_cast_fp16")];
+            tensor<fp16, [1, 256, ?]> var_510_cast_fp16 = mul(x = var_508_cast_fp16, y = var_509_cast_fp16)[name = tensor<string, []>("op_510_cast_fp16")];
+            tensor<fp16, [1, 256, ?]> input_95_cast_fp16 = add(x = var_510_cast_fp16, y = var_505_cast_fp16_1)[name = tensor<string, []>("input_95_cast_fp16")];
+            tensor<fp16, [1, 256, ?]> input_97_cast_fp16 = leaky_relu(alpha = var_478, x = input_95_cast_fp16)[name = tensor<string, []>("input_97_cast_fp16")];
+            tensor<string, []> input_101_pad_type_0 = const()[name = tensor<string, []>("input_101_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [2]> input_101_pad_0 = const()[name = tensor<string, []>("input_101_pad_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [1]> input_101_strides_0 = const()[name = tensor<string, []>("input_101_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> input_101_dilations_0 = const()[name = tensor<string, []>("input_101_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_101_groups_0 = const()[name = tensor<string, []>("input_101_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [256, 256, 3]> weight_27_to_fp16 = const()[name = tensor<string, []>("weight_27_to_fp16"), val = tensor<fp16, [256, 256, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15901760)))];
+            tensor<fp16, [256]> N_2_conv1_bias_to_fp16 = const()[name = tensor<string, []>("N_2_conv1_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16295040)))];
+            tensor<fp16, [1, 256, ?]> input_101_cast_fp16 = conv(bias = N_2_conv1_bias_to_fp16, dilations = input_101_dilations_0, groups = input_101_groups_0, pad = input_101_pad_0, pad_type = input_101_pad_type_0, strides = input_101_strides_0, weight = weight_27_to_fp16, x = input_97_cast_fp16)[name = tensor<string, []>("input_101_cast_fp16")];
+            tensor<fp16, [512, 128]> N_2_norm2_fc_weight_to_fp16 = const()[name = tensor<string, []>("N_2_norm2_fc_weight_to_fp16"), val = tensor<fp16, [512, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16295616)))];
+            tensor<fp16, [512]> N_2_norm2_fc_bias_to_fp16 = const()[name = tensor<string, []>("N_2_norm2_fc_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16426752)))];
+            tensor<fp16, [1, 512]> linear_11_cast_fp16 = linear(bias = N_2_norm2_fc_bias_to_fp16, weight = N_2_norm2_fc_weight_to_fp16, x = s_to_fp16)[name = tensor<string, []>("linear_11_cast_fp16")];
+            tensor<int32, [3]> var_527 = const()[name = tensor<string, []>("op_527"), val = tensor<int32, [3]>([1, 512, 1])];
+            tensor<fp16, [1, 512, 1]> h_cast_fp16 = reshape(shape = var_527, x = linear_11_cast_fp16)[name = tensor<string, []>("h_cast_fp16")];
+            tensor<int32, [2]> var_529_split_sizes_0 = const()[name = tensor<string, []>("op_529_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
+            tensor<int32, []> var_529_axis_0 = const()[name = tensor<string, []>("op_529_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 256, 1]> var_529_cast_fp16_0, tensor<fp16, [1, 256, 1]> var_529_cast_fp16_1 = split(axis = var_529_axis_0, split_sizes = var_529_split_sizes_0, x = h_cast_fp16)[name = tensor<string, []>("op_529_cast_fp16")];
+            tensor<fp16, []> var_531_promoted_to_fp16 = const()[name = tensor<string, []>("op_531_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1, 256, 1]> var_532_cast_fp16 = add(x = var_529_cast_fp16_0, y = var_531_promoted_to_fp16)[name = tensor<string, []>("op_532_cast_fp16")];
+            tensor<fp16, [1, 256, ?]> var_533_cast_fp16 = instance_norm(epsilon = var_481_to_fp16, x = input_101_cast_fp16)[name = tensor<string, []>("op_533_cast_fp16")];
+            tensor<fp16, [1, 256, ?]> var_534_cast_fp16 = mul(x = var_532_cast_fp16, y = var_533_cast_fp16)[name = tensor<string, []>("op_534_cast_fp16")];
+            tensor<fp16, [1, 256, ?]> input_103_cast_fp16 = add(x = var_534_cast_fp16, y = var_529_cast_fp16_1)[name = tensor<string, []>("input_103_cast_fp16")];
+            tensor<fp16, [1, 256, ?]> input_105_cast_fp16 = leaky_relu(alpha = var_478, x = input_103_cast_fp16)[name = tensor<string, []>("input_105_cast_fp16")];
+            tensor<string, []> out_pad_type_0 = const()[name = tensor<string, []>("out_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [2]> out_pad_0 = const()[name = tensor<string, []>("out_pad_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [1]> out_strides_0 = const()[name = tensor<string, []>("out_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> out_dilations_0 = const()[name = tensor<string, []>("out_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> out_groups_0 = const()[name = tensor<string, []>("out_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [256, 256, 3]> weight_29_to_fp16 = const()[name = tensor<string, []>("weight_29_to_fp16"), val = tensor<fp16, [256, 256, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16427840)))];
+            tensor<fp16, [256]> N_2_conv2_bias_to_fp16 = const()[name = tensor<string, []>("N_2_conv2_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16821120)))];
+            tensor<fp16, [1, 256, ?]> out_cast_fp16 = conv(bias = N_2_conv2_bias_to_fp16, dilations = out_dilations_0, groups = out_groups_0, pad = out_pad_0, pad_type = out_pad_type_0, strides = out_strides_0, weight = weight_29_to_fp16, x = input_105_cast_fp16)[name = tensor<string, []>("out_cast_fp16")];
+            tensor<fp16, [1, 256, ?]> var_545_cast_fp16 = add(x = out_cast_fp16, y = _inversed_input_93_cast_fp16)[name = tensor<string, []>("op_545_cast_fp16")];
+            tensor<fp16, []> _inversed_input_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_input_y_0_to_fp16"), val = tensor<fp16, []>(0x1.6ap-1)];
+            tensor<fp16, [1, 256, ?]> _inversed_input_cast_fp16 = mul(x = var_545_cast_fp16, y = _inversed_input_y_0_to_fp16)[name = tensor<string, []>("_inversed_input_cast_fp16")];
+            tensor<string, []> var_558_pad_type_0 = const()[name = tensor<string, []>("op_558_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> var_558_strides_0 = const()[name = tensor<string, []>("op_558_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> var_558_pad_0 = const()[name = tensor<string, []>("op_558_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> var_558_dilations_0 = const()[name = tensor<string, []>("op_558_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> var_558_groups_0 = const()[name = tensor<string, []>("op_558_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 256, 1]> N_proj_weight_to_fp16 = const()[name = tensor<string, []>("N_proj_weight_to_fp16"), val = tensor<fp16, [1, 256, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16821696)))];
+            tensor<fp16, [1]> N_proj_bias_to_fp16 = const()[name = tensor<string, []>("N_proj_bias_to_fp16"), val = tensor<fp16, [1]>([0x1.6ep-4])];
+            tensor<fp16, [1, 1, ?]> var_558_cast_fp16 = conv(bias = N_proj_bias_to_fp16, dilations = var_558_dilations_0, groups = var_558_groups_0, pad = var_558_pad_0, pad_type = var_558_pad_type_0, strides = var_558_strides_0, weight = N_proj_weight_to_fp16, x = _inversed_input_cast_fp16)[name = tensor<string, []>("op_558_cast_fp16")];
+            tensor<int32, [1]> var_560_axes_0 = const()[name = tensor<string, []>("op_560_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, ?]> N = squeeze(axes = var_560_axes_0, x = var_558_cast_fp16)[name = tensor<string, []>("op_560_cast_fp16")];
+        } -> (F0, N);
+}

compiled/styletts2_f0n_energy.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b21f183d2ff876842ea2df14cdc033c8935a1805382b70b241c4f5a1bf32b3a8
+size 16822272

compiled/styletts2_text_predictor_128.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e33f8046365fa207d90593ce420673cd6921bb7156c5d3340cd5c6290990e4f9
+size 243

compiled/styletts2_text_predictor_128.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9c5b8da13d13caee8b6e7fa3bf35b2d98fe29a782deded2477aea61d92d8c283
+size 528

compiled/styletts2_text_predictor_128.mlmodelc/metadata.json

@@ -0,0 +1,138 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "StyleTTS2 text+predictor (LibriTTS HiFi-GAN) @ T_tok=128.",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 512 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 512, 128]",
+        "name" : "t_en",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 512 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 512, 128]",
+        "name" : "d_en",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 128 × 640)",
+        "shortDescription" : "",
+        "shape" : "[1, 128, 640]",
+        "name" : "d",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 128 × 50)",
+        "shortDescription" : "",
+        "shape" : "[1, 128, 50]",
+        "name" : "pred_dur_log",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 128 × 768)",
+        "shortDescription" : "",
+        "shape" : "[1, 128, 768]",
+        "name" : "fixed_embedding",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 128 × 768)",
+        "shortDescription" : "",
+        "shape" : "[1, 128, 768]",
+        "name" : "bert_dur",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Float16",
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Split" : 3,
+      "Ios17.mul" : 15,
+      "Ios17.linear" : 78,
+      "Ios17.transpose" : 76,
+      "Ios17.matmul" : 24,
+      "Ios17.conv" : 3,
+      "Ios17.leakyRelu" : 3,
+      "Ios17.gather" : 2,
+      "Ios17.add" : 32,
+      "Ios17.lstm" : 5,
+      "Ios17.concat" : 4,
+      "Ios16.softmax" : 12,
+      "Ios16.gelu" : 12,
+      "Ios17.expandDims" : 1,
+      "Tile" : 1,
+      "Ios17.cast" : 3,
+      "Ios17.layerNorm" : 31,
+      "Ios17.reshape" : 51
+    },
+    "computePrecision" : "Mixed (Float16, Float32, Int16, Int32, UInt16)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 128]",
+        "name" : "tokens",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 128]",
+        "name" : "style",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.source_dialect" : "TorchScript",
+      "com.github.apple.coremltools.source" : "torch==2.11.0",
+      "com.github.apple.coremltools.version" : "8.3.0"
+    },
+    "generatedClassName" : "styletts2_text_predictor_128",
+    "method" : "predict"
+  }
+]

compiled/styletts2_text_predictor_128.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:79ce18107b3f5fe1e518342354f55dd071f42a30e18f4512029fd9b3635105ce
+size 39073856

compiled/styletts2_text_predictor_256.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d2a3e12ca039dc9909255950043c9bd39018dea7a926c0e6f9c596043b7c75d6
+size 243

compiled/styletts2_text_predictor_256.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:990616fa15d673099e240788d898fed9aea4f27dfebe3e34a4ea44a5f61416da
+size 528

compiled/styletts2_text_predictor_256.mlmodelc/metadata.json

@@ -0,0 +1,138 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "StyleTTS2 text+predictor (LibriTTS HiFi-GAN) @ T_tok=256.",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 512 × 256)",
+        "shortDescription" : "",
+        "shape" : "[1, 512, 256]",
+        "name" : "t_en",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 512 × 256)",
+        "shortDescription" : "",
+        "shape" : "[1, 512, 256]",
+        "name" : "d_en",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 256 × 640)",
+        "shortDescription" : "",
+        "shape" : "[1, 256, 640]",
+        "name" : "d",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 256 × 50)",
+        "shortDescription" : "",
+        "shape" : "[1, 256, 50]",
+        "name" : "pred_dur_log",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 256 × 768)",
+        "shortDescription" : "",
+        "shape" : "[1, 256, 768]",
+        "name" : "fixed_embedding",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 256 × 768)",
+        "shortDescription" : "",
+        "shape" : "[1, 256, 768]",
+        "name" : "bert_dur",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Float16",
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Split" : 3,
+      "Ios17.mul" : 15,
+      "Ios17.linear" : 78,
+      "Ios17.transpose" : 76,
+      "Ios17.matmul" : 24,
+      "Ios17.conv" : 3,
+      "Ios17.leakyRelu" : 3,
+      "Ios17.gather" : 2,
+      "Ios17.add" : 32,
+      "Ios17.lstm" : 5,
+      "Ios17.concat" : 4,
+      "Ios16.softmax" : 12,
+      "Ios16.gelu" : 12,
+      "Ios17.expandDims" : 1,
+      "Tile" : 1,
+      "Ios17.cast" : 3,
+      "Ios17.layerNorm" : 31,
+      "Ios17.reshape" : 51
+    },
+    "computePrecision" : "Mixed (Float16, Float32, Int16, Int32, UInt16)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1 × 256)",
+        "shortDescription" : "",
+        "shape" : "[1, 256]",
+        "name" : "tokens",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 128]",
+        "name" : "style",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.version" : "8.3.0",
+      "com.github.apple.coremltools.source" : "torch==2.11.0",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "generatedClassName" : "styletts2_text_predictor_256",
+    "method" : "predict"
+  }
+]

compiled/styletts2_text_predictor_256.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4221120586ac22d722b35894d94b7a370d35721fd5ffc5ca93bca4c0da90a868
+size 39336000

compiled/styletts2_text_predictor_32.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7a8c53623b57a193d6fd43afe30a1126d4fb5d8685d668cb7e5545766ee39f80
+size 243

compiled/styletts2_text_predictor_32.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:452c32fad506218658305781fea9f2b734a00c71dae78f0c4bd5ff46a1a68230
+size 520

compiled/styletts2_text_predictor_32.mlmodelc/metadata.json

@@ -0,0 +1,138 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "StyleTTS2 text+predictor (LibriTTS HiFi-GAN) @ T_tok=32.",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 512 × 32)",
+        "shortDescription" : "",
+        "shape" : "[1, 512, 32]",
+        "name" : "t_en",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 512 × 32)",
+        "shortDescription" : "",
+        "shape" : "[1, 512, 32]",
+        "name" : "d_en",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 32 × 640)",
+        "shortDescription" : "",
+        "shape" : "[1, 32, 640]",
+        "name" : "d",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 32 × 50)",
+        "shortDescription" : "",
+        "shape" : "[1, 32, 50]",
+        "name" : "pred_dur_log",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 32 × 768)",
+        "shortDescription" : "",
+        "shape" : "[1, 32, 768]",
+        "name" : "fixed_embedding",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 32 × 768)",
+        "shortDescription" : "",
+        "shape" : "[1, 32, 768]",
+        "name" : "bert_dur",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Float16",
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Split" : 3,
+      "Ios17.mul" : 15,
+      "Ios17.linear" : 78,
+      "Ios17.transpose" : 76,
+      "Ios17.matmul" : 24,
+      "Ios17.conv" : 3,
+      "Ios17.leakyRelu" : 3,
+      "Ios17.gather" : 2,
+      "Ios17.add" : 32,
+      "Ios17.lstm" : 5,
+      "Ios17.concat" : 4,
+      "Ios16.softmax" : 12,
+      "Ios16.gelu" : 12,
+      "Ios17.expandDims" : 1,
+      "Tile" : 1,
+      "Ios17.cast" : 3,
+      "Ios17.layerNorm" : 31,
+      "Ios17.reshape" : 51
+    },
+    "computePrecision" : "Mixed (Float16, Float32, Int16, Int32, UInt16)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1 × 32)",
+        "shortDescription" : "",
+        "shape" : "[1, 32]",
+        "name" : "tokens",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 128]",
+        "name" : "style",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.version" : "8.3.0",
+      "com.github.apple.coremltools.source" : "torch==2.11.0",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "generatedClassName" : "styletts2_text_predictor_32",
+    "method" : "predict"
+  }
+]