Add Core ML and Core AI bundles of v2-20261005

.gitattributes

@@ -35,3 +35,8 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
 onnx/onnx/g2p/model.fp32.onnx.data filter=lfs diff=lfs merge=lfs -text
 onnx-int8/onnx/g2p/model.fp32.onnx.data filter=lfs diff=lfs merge=lfs -text
+coreai/coreai/duration_predictor.aimodel/main.mlirb filter=lfs diff=lfs merge=lfs -text
+coreai/coreai/g2p.aimodel/main.mlirb filter=lfs diff=lfs merge=lfs -text
+coreai/coreai/vector_context_encoder.aimodel/main.mlirb filter=lfs diff=lfs merge=lfs -text
+coreai/coreai/vector_estimator.aimodel/main.mlirb filter=lfs diff=lfs merge=lfs -text
+coreai/coreai/vocoder.aimodel/main.mlirb filter=lfs diff=lfs merge=lfs -text

README.md

@@ -3,12 +3,12 @@ license: apache-2.0
 language: [en, es, it, fr, de, vi]
 library_name: litert
 pipeline_tag: text-to-speech
-tags: [text-to-speech, tts, litert, onnx, pytorch, multilingual, expressive-tts, duration-flow, whisper]
+tags: [text-to-speech, tts, litert, onnx, coreml, coreai, pytorch, multilingual, expressive-tts, duration-flow, whisper]
 ---
 
 # Scylla's Band v2
 
-Release **`v2-20261005`**. The [Scylla's Band runtime](https://github.com/lowkeytea/scyllasband) supports this release.
+Release **`v2-20261005`**. The [Scylla's Band runtime](https://github.com/lowkeytea/scyllasband) supports this release. Tag `v2-20261005-apple` adds Core ML and Core AI bundles of the same model; the runtime downloads from that tag.
 
 The model exposes **energy, tension, valence and assertiveness** on **0–4 scales with neutral at 2**, plus **binary whisper**. Combining the controls changes the generated delivery; a request is not a guarantee of a particular perceived emotion or exact external scorer value.
 
@@ -23,7 +23,7 @@ python -m scyllasband speak --voice ariadne --language en_us \
     --sampler heun --steps 8 -o hello.wav "It is a pleasure to meet you."
 ```
 
-The default download is the LiteRT bundle, installed under `scyllasband/models/litert`. `--flavor onnx` downloads the ONNX Runtime bundle instead (`pip install onnxruntime`). Downloads pin this release tag. The manifest declares graph contract `scyllasband_measured_delivery_v2`; runtimes written for earlier releases cannot run these graphs.
+The default download is the Core ML bundle on an Apple silicon Mac and the LiteRT bundle elsewhere, installed under `scyllasband/models/<bundle>`. A Mac that can build apps for iOS, iPadOS or visionOS 27 is also offered the Core AI bundle those apps use. `--flavor coreml`, `--flavor coreai`, `--flavor litert` or `--flavor onnx` choose explicitly. Downloads pin a release tag. The manifest declares graph contract `scyllasband_measured_delivery_v2`; runtimes written for earlier releases cannot run these graphs.
 
 ## Controls
 
@@ -32,27 +32,31 @@ The default download is the LiteRT bundle, installed under `scyllasband/models/l
 | Energy, tension, valence, assertiveness | 0–4 or unspecified | 2 each |
 | Whisper | on/off or unspecified | off |
 
-With every control at 2, the default, voices speak in a calm, conversational delivery with a neutral tone. Moving a control away from 2 changes the delivery:
+With every control at 2, the default, voices speak in a calm, conversational delivery with a neutral tone. Raising or lowering a control changes the delivery:
 
-| Control | Toward 0 | Toward 4 |
+| Control | Lower | Higher |
 | --- | --- | --- |
-| Energy | Lower, darker voice with more pauses between phrases | Higher, brighter, more projected voice |
+| Energy | Lower, darker voice with a few more pauses | Higher, brighter, more projected voice |
 | Tension | Slightly looser and steadier | Tighter, with wider pitch swings |
-| Valence | Flatter pitch, a little slower, longer pauses | Livelier pitch, a little quicker, shorter pauses |
-| Assertiveness | Hesitant, with longer pauses between phrases | Faster and clipped with few pauses; the strongest control over pace |
+| Valence | Flatter, more downcast pitch | Livelier pitch, a little quicker, shorter pauses |
+| Assertiveness | More pauses between phrases | Faster and clipped with fewer pauses; the strongest control over pace |
 
-Whisper on gives quieter, breathy speech. Combining the controls gives familiar deliveries:
+Whisper on gives quieter, breathy speech.
+
+The model learned delivery mostly from settings between about 1.7 and 2.8 for energy, 1.9 and 3.0 for valence, 1.8 and 2.7 for assertiveness, and 1.0 and 3.1 for tension. The controls are most reliable inside those ranges; further out, voices can sound strained or a style can become ambiguous.
+
+Combining the controls gives familiar deliveries:
 
 | Delivery | Settings |
 | --- | --- |
-| Calm and soft | `energy=1.2,tension=0.8,valence=2.6` |
-| Assertive and rapid-fire | `energy=2.6,tension=2.2,assertiveness=3.6` |
-| Joyful | `energy=3.2,valence=3.6` |
-| Angry | `energy=3.4,tension=3.4,valence=0.8,assertiveness=3.4` |
-| Sad | `energy=1,tension=1.6,valence=0.8,assertiveness=1.6` |
+| Calm and soft | `energy=1.8,tension=1,valence=2.5,assertiveness=2.1` |
+| Assertive and quick | `energy=2.6,tension=2.4,valence=2.1,assertiveness=2.8` |
+| Joyful | `energy=2.7,tension=2.1,valence=3.1,assertiveness=2.3` |
+| Angry | `energy=2.7,tension=3.1,valence=1.3,assertiveness=2.5` |
+| Sad | `energy=1.7,tension=1.7,valence=1.8,assertiveness=2` |
 | Whispered | `whisper=on` |
 
-The [voice gallery](https://lowkeytea.github.io/scyllasband/#delivery) plays each of these for all ten voices with the exact settings used. How far a control moves the delivery varies with voice, language and text, so start near 2 and listen as you move toward the extremes.
+Angry takes valence a little below the trained range so that it separates clearly from joyful. These are moderate styles; how strongly each comes through varies with voice, language and text. The [voice gallery](https://lowkeytea.github.io/scyllasband/#delivery) plays each of them for all ten voices.
 
 `--delivery auto` omits all coordinates. A per-axis `auto` omits just that coordinate; other unspecified fields in a partial request use neutral defaults. Omission is represented by a separate mask, not by the value 2. It is not an automatic text-to-performance planner.
 
@@ -86,12 +90,14 @@ All ten have Spanish (`es`), Italian (`it`), French (`fr`), German (`de`) and Vi
 
 ## Artifacts
 
-- `litert/`: LiteRT bundle; default download. Transformer and vocoder weights are dynamic-range INT8; each size bucket is a signature of one model file, sharing weights.
+- `litert/` (164 MB): LiteRT bundle; default download. Transformer and vocoder weights are dynamic-range INT8; each size bucket is a signature of one model file, sharing weights.
+- `coreai/` (155 MB): Core AI bundle for apps on iOS, iPadOS and visionOS 27 (`.aimodel` assets); it also runs on macOS 27, where the runtime uses the Core ML bundle by default. INT8 weights per output channel with FP16 compute; size buckets are functions of one asset sharing weights. The flow also runs accurately on the Neural Engine (`controls.coreai.neural_engine_assets`), which is the placement for apps that speak in the background; the GPU is the default.
+- `coreml/` (142 MB): Core ML bundle for Apple silicon with iOS 18 / macOS 15 and later (compiled `.mlmodelc` assets), with INT8 weights and FP16 compute; GPU by default, CPU in background apps.
 - `onnx/`: ONNX Runtime bundle with dynamic INT8 transformer weights; the vocoder stays in full precision.
 - `pytorch/`: duration, vector, adapter/vocoder, autoencoder and G2P checkpoints, with the release configuration assets.
 - `SHA256SUMS.json`: published-file hashes.
 
-Both runtime bundles include the G2P model in a matching precision.
+Every runtime bundle includes the G2P model in a matching precision. The Apple bundles also give the G2P narrower input widths (64–512 tokens), so short phrases are converted faster.
 
 ## Validation and limitations
 
@@ -100,6 +106,7 @@ Release validation separates graph accuracy, frontend/timing behaviour and perce
 - Component comparisons against PyTorch during export, and end-to-end comparison of full passages against the PyTorch reference path.
 - Quantized bundles compared with full precision on real sentences: identical phonemes, phone durations and generated latents within small tolerances, followed by listening across all ten voices.
 - G2P conversions compared with the full-precision G2P on multilingual text.
+- Core ML and Core AI bundles compared with the PyTorch model on every graph and on real passages, on each compute unit they are used with (GPU, CPU, and for Core AI the Neural Engine): identical phonemes apart from two near-tie stress marks also seen in FP16, fewer changed phone durations than the LiteRT bundle, and smaller latent and spectral differences. The Apple exports rewrite three computations into equivalent forms that FP16 and the Neural Engine evaluate exactly (span pooling, the modifier bit decoding, the duration model's combined phone/segment index), and scale the flow's residual stream; in FP32 the rewritten graphs are bit-identical to the trained ones.
 
 Listening is decisive; external emotion-score agreement is approximate. Scorer measurements are imperfect supervision and do not establish emotional ground truth. Training audio is not distributed.
 

SHA256SUMS.json

@@ -4,8 +4,236 @@
   "bytes": 11357
  },
  "README.md": {
-  "sha256": "b7590ee31e834406273b1bf28ea3ee2380aba97f18bf6113d39b60e745add56f",
-  "bytes": 7158
+  "sha256": "b4c317afeaa2544f675f8d48f63c9fa5d809bdc7d7eabeb3b1432ef0fedd4570",
+  "bytes": 9251
+ },
+ "coreai/assets/fixed_references.json": {
+  "sha256": "d3c26e7e90d90fa719254ff1ff05dc4b391cd47231b7556c3c2ed5f63547925b",
+  "bytes": 1115042
+ },
+ "coreai/assets/g2p/export_config.json": {
+  "sha256": "0ec5ea81f346521f3a5c9c4a90b939c6b5143b1a11cc5e74b9902969db56e568",
+  "bytes": 928
+ },
+ "coreai/assets/g2p/language_map.json": {
+  "sha256": "b3db2048b492799f8746b43001e0f7411eb5d359e17cabc6bfce4a7bccbfdd99",
+  "bytes": 844
+ },
+ "coreai/assets/g2p/normalization.json": {
+  "sha256": "f9ca4fe4e4eca5e02b6f2874f5da6ede8b44b25c71c3c6ac00f96d7397024de6",
+  "bytes": 1090
+ },
+ "coreai/assets/g2p/phoneme_dict.json": {
+  "sha256": "9cfb40b2fd31d7a32351c33f90230062554a4dc0a44e89f722de211d42ed1875",
+  "bytes": 93
+ },
+ "coreai/assets/g2p/pronunciation_overrides.json": {
+  "sha256": "a1f311379b76cb44719a2242a93a4046ff5f344cdac613374288234c2817eb59",
+  "bytes": 80
+ },
+ "coreai/assets/g2p/tokenizer.json": {
+  "sha256": "4843d57a422b0a6045bbc6fc9a99afa329126a6742ff95ebcefcda660f3ee474",
+  "bytes": 7769
+ },
+ "coreai/assets/languages.json": {
+  "sha256": "f2af2233d0a1c0dbb6e6236fd57592d12da84262fdd1a336ead1ee83a657e2ae",
+  "bytes": 289
+ },
+ "coreai/assets/phone_vocab.json": {
+  "sha256": "83e6457a69948edbd9a94b404c22fd408c263eb0d0b5acaafa4926e994587721",
+  "bytes": 1793
+ },
+ "coreai/assets/voices.json": {
+  "sha256": "db9aa57e3d5f9f34fed283fb8e79f4c38e218647d33bc4a5c71ea3e8338c1a56",
+  "bytes": 433
+ },
+ "coreai/coreai/duration_predictor.aimodel/main.hash": {
+  "sha256": "ec1450950c5c92a8567b91acb3999a9d53ebcfae92401c4c9893a3ab780020af",
+  "bytes": 32
+ },
+ "coreai/coreai/duration_predictor.aimodel/main.mlirb": {
+  "sha256": "4243d405361c9ebc82a0e233a3cb1bd5d64b5389a96ddd58947f14fb84b3c949",
+  "bytes": 10884924
+ },
+ "coreai/coreai/duration_predictor.aimodel/metadata.json": {
+  "sha256": "3b825889dba08f532b2e922b099aff707fdefc53cf86716392f5867f1b667999",
+  "bytes": 105
+ },
+ "coreai/coreai/g2p.aimodel/main.hash": {
+  "sha256": "300a60ae5cd565c7207ff5877dfa8427af6868a14fb6ccfa195ca17fe0e50a3f",
+  "bytes": 32
+ },
+ "coreai/coreai/g2p.aimodel/main.mlirb": {
+  "sha256": "e7426aa159702a859176420d0c34b493da7d3dd97fa74fdeedb2ee4c4f0dce18",
+  "bytes": 28377271
+ },
+ "coreai/coreai/g2p.aimodel/metadata.json": {
+  "sha256": "07542e963b0d178a540fa3a7927bd54ac3050c4ccc959dc38defe42e33fe7066",
+  "bytes": 105
+ },
+ "coreai/coreai/vector_context_encoder.aimodel/main.hash": {
+  "sha256": "0dfa623d5496abf0b862e73ed8adc2a4e8bf5f4249b1c209543ca44312cdc5f3",
+  "bytes": 32
+ },
+ "coreai/coreai/vector_context_encoder.aimodel/main.mlirb": {
+  "sha256": "9dca40e4aa2ce1cf2099c7c4ee34bbea72f2594d88459210cfa06dfc9592f135",
+  "bytes": 8029277
+ },
+ "coreai/coreai/vector_context_encoder.aimodel/metadata.json": {
+  "sha256": "2adcb7092d18dc35b8686fc8b5a6075c09b852ce0353899bc1459cdbbfb3f6ba",
+  "bytes": 105
+ },
+ "coreai/coreai/vector_estimator.aimodel/main.hash": {
+  "sha256": "eddcebf2eca499422d841f257a6c270866151ee1b5f082ad375dda29264176c3",
+  "bytes": 32
+ },
+ "coreai/coreai/vector_estimator.aimodel/main.mlirb": {
+  "sha256": "4dc8e41c5c9de01f2d3a9fd847a30df8b27ceec9aa0894b235f49181dcd664d2",
+  "bytes": 63149068
+ },
+ "coreai/coreai/vector_estimator.aimodel/metadata.json": {
+  "sha256": "e90fa086d07ef39797bce9b6d6c82429d84b17a5ebb77380d71d94d44eb630f7",
+  "bytes": 105
+ },
+ "coreai/coreai/vocoder.aimodel/main.hash": {
+  "sha256": "62fae3f0d6ca6e5cb648e1c54d47c243d919bdb9e93e89e53f786eb318f90003",
+  "bytes": 32
+ },
+ "coreai/coreai/vocoder.aimodel/main.mlirb": {
+  "sha256": "0e0d07b91327322c55da5081c7e3301565e92c5298dd9a5dfb74b71ee6074669",
+  "bytes": 43811030
+ },
+ "coreai/coreai/vocoder.aimodel/metadata.json": {
+  "sha256": "d1169c91caecc862135efebf9a832fd956459929050ca29415db3e41197ef6b2",
+  "bytes": 105
+ },
+ "coreai/manifest.json": {
+  "sha256": "a2a3b2c822673e98ae44af5280ab3b71039609ea913d1c50392faf37ed97c19d",
+  "bytes": 21685
+ },
+ "coreml/assets/fixed_references.json": {
+  "sha256": "d3c26e7e90d90fa719254ff1ff05dc4b391cd47231b7556c3c2ed5f63547925b",
+  "bytes": 1115042
+ },
+ "coreml/assets/g2p/export_config.json": {
+  "sha256": "0ec5ea81f346521f3a5c9c4a90b939c6b5143b1a11cc5e74b9902969db56e568",
+  "bytes": 928
+ },
+ "coreml/assets/g2p/language_map.json": {
+  "sha256": "b3db2048b492799f8746b43001e0f7411eb5d359e17cabc6bfce4a7bccbfdd99",
+  "bytes": 844
+ },
+ "coreml/assets/g2p/normalization.json": {
+  "sha256": "f9ca4fe4e4eca5e02b6f2874f5da6ede8b44b25c71c3c6ac00f96d7397024de6",
+  "bytes": 1090
+ },
+ "coreml/assets/g2p/phoneme_dict.json": {
+  "sha256": "9cfb40b2fd31d7a32351c33f90230062554a4dc0a44e89f722de211d42ed1875",
+  "bytes": 93
+ },
+ "coreml/assets/g2p/pronunciation_overrides.json": {
+  "sha256": "a1f311379b76cb44719a2242a93a4046ff5f344cdac613374288234c2817eb59",
+  "bytes": 80
+ },
+ "coreml/assets/g2p/tokenizer.json": {
+  "sha256": "4843d57a422b0a6045bbc6fc9a99afa329126a6742ff95ebcefcda660f3ee474",
+  "bytes": 7769
+ },
+ "coreml/assets/languages.json": {
+  "sha256": "f2af2233d0a1c0dbb6e6236fd57592d12da84262fdd1a336ead1ee83a657e2ae",
+  "bytes": 289
+ },
+ "coreml/assets/phone_vocab.json": {
+  "sha256": "83e6457a69948edbd9a94b404c22fd408c263eb0d0b5acaafa4926e994587721",
+  "bytes": 1793
+ },
+ "coreml/assets/voices.json": {
+  "sha256": "db9aa57e3d5f9f34fed283fb8e79f4c38e218647d33bc4a5c71ea3e8338c1a56",
+  "bytes": 433
+ },
+ "coreml/coreml/duration_predictor.mlmodelc/analytics/coremldata.bin": {
+  "sha256": "953e274e93c3abfae6f04977436816dc28223ef61d4e91e4191f67837740a2e8",
+  "bytes": 243
+ },
+ "coreml/coreml/duration_predictor.mlmodelc/coremldata.bin": {
+  "sha256": "c090d28d8df68312410867a49963b5fbda00f5883178969df17b1b0c86e7eff2",
+  "bytes": 554
+ },
+ "coreml/coreml/duration_predictor.mlmodelc/model.mil": {
+  "sha256": "b342ee3bcc2acb6b8bd0de42bebfc6c3d83b40e5512156fb26f909fa9ae3d608",
+  "bytes": 138985
+ },
+ "coreml/coreml/duration_predictor.mlmodelc/weights/weight.bin": {
+  "sha256": "5383e19195d6807cf6632d2c82b51b86aaa947ba51d8042434f73b5039c70ea4",
+  "bytes": 10562880
+ },
+ "coreml/coreml/g2p.mlmodelc/analytics/coremldata.bin": {
+  "sha256": "61936fdfe11637e628b2134025bea29d673c98e93ce6f01c05e7b5da124563b8",
+  "bytes": 243
+ },
+ "coreml/coreml/g2p.mlmodelc/coremldata.bin": {
+  "sha256": "f5fc8bee0186f33fa14d35a59a58ec3393676d07bdf03802a3836d806662cd03",
+  "bytes": 377
+ },
+ "coreml/coreml/g2p.mlmodelc/model.mil": {
+  "sha256": "d464334a231d99b9a3764d50c28657d1720e25eaa45e39fa4daa401bb89452a1",
+  "bytes": 388409
+ },
+ "coreml/coreml/g2p.mlmodelc/weights/weight.bin": {
+  "sha256": "045da2991ed2a42ba027aaa6b83813054f410b6eedfdce82f2ab287ec4ca74e2",
+  "bytes": 27346752
+ },
+ "coreml/coreml/vector_context_encoder.mlmodelc/analytics/coremldata.bin": {
+  "sha256": "f89405fefad88549b521a819805c5c1781e0c930c03b8f670b1e1539aafe88c6",
+  "bytes": 243
+ },
+ "coreml/coreml/vector_context_encoder.mlmodelc/coremldata.bin": {
+  "sha256": "d031d4ef221bc846aec129114aec7366fcb6126dacae44c24c4b344c54dc6c5a",
+  "bytes": 476
+ },
+ "coreml/coreml/vector_context_encoder.mlmodelc/model.mil": {
+  "sha256": "c3c0df163512f0ed8ab4dbb927054f2cb627980c734419bd96e341540660d372",
+  "bytes": 44809
+ },
+ "coreml/coreml/vector_context_encoder.mlmodelc/weights/weight.bin": {
+  "sha256": "7120be9d216af8b1021ad92a4d3d0d136881d55e48debfb348197f12da2a21d6",
+  "bytes": 7572352
+ },
+ "coreml/coreml/vector_estimator.mlmodelc/analytics/coremldata.bin": {
+  "sha256": "4cd08db9de17f4e62caab8acf50c3c28909c28b945acb105f35e45820366621a",
+  "bytes": 243
+ },
+ "coreml/coreml/vector_estimator.mlmodelc/coremldata.bin": {
+  "sha256": "d22410a5a41fc0aae79224588c2da505cd4abae60526d06fd989749001741762",
+  "bytes": 3647
+ },
+ "coreml/coreml/vector_estimator.mlmodelc/model.mil": {
+  "sha256": "48c7230b2cee7088f5431117ea54758e6a37e5f55faa8cdea3aa7fdd34010831",
+  "bytes": 1469007
+ },
+ "coreml/coreml/vector_estimator.mlmodelc/weights/weight.bin": {
+  "sha256": "135922d329c47a98942f1ce104c3e62f9ccc6578b6866d57482839172c9bf1be",
+  "bytes": 60754048
+ },
+ "coreml/coreml/vocoder.mlmodelc/analytics/coremldata.bin": {
+  "sha256": "c8c3075b62c159b4773c0d721e911570449783db5e493eb53e2a0a78049e717e",
+  "bytes": 243
+ },
+ "coreml/coreml/vocoder.mlmodelc/coremldata.bin": {
+  "sha256": "04bf9f9de7edfc6ce58e5928a0a1d6669d09095ebc03cb5ca24dd85a416653df",
+  "bytes": 1190
+ },
+ "coreml/coreml/vocoder.mlmodelc/model.mil": {
+  "sha256": "e7b28b62c2ce75b56bb284333542585cf4abe4ebda50ce72322cbea88dcea53c",
+  "bytes": 969259
+ },
+ "coreml/coreml/vocoder.mlmodelc/weights/weight.bin": {
+  "sha256": "8d1aa1e85be960d5ad32f751b602c38601bc9517a951be5bd42c991449d3bccc",
+  "bytes": 31043584
+ },
+ "coreml/manifest.json": {
+  "sha256": "3519f7fe3c8c98e8672a99451e44ff7ef60ed35317031e82669adbb1b5acf3a0",
+  "bytes": 21577
  },
  "litert/assets/fixed_references.json": {
   "sha256": "d3c26e7e90d90fa719254ff1ff05dc4b391cd47231b7556c3c2ed5f63547925b",

coreai/assets/g2p/export_config.json

@@ -0,0 +1,33 @@
+{
+  "batch_dynamic": true,
+  "format": "scyllasband_g2p",
+  "input_granularity": "phrase",
+  "languages": [
+    "de",
+    "en_gb",
+    "en_us",
+    "es",
+    "fr",
+    "it",
+    "vi"
+  ],
+  "model_type": "transformer",
+  "onnx_export_batch_size": 4,
+  "onnx_export_layout": "batch_first",
+  "onnx_export_max_batch_size": 256,
+  "onnx_exporter": "dynamo",
+  "onnx_fp32_model": "model.onnx",
+  "onnx_model": "model.onnx",
+  "onnx_padded_output": "ctc_blank",
+  "onnx_quantization": "none",
+  "onnx_quantized_parity_tolerance": 3.0,
+  "opset": 18,
+  "phoneme_token_mode": "char",
+  "phrase_rows_per_language": 200000,
+  "preserve_commas": true,
+  "punctuation_silence_target": "explicit_silence",
+  "release_role": "fp32_bundle_frontend",
+  "source_checkpoint": "output/deepphonemizer_multilingual/model_v10_7lang/checkpoints/model_step_250k.pt",
+  "training_mix": "phrase_word_chunked_clean",
+  "word_rows_per_language": 200000
+}

coreai/assets/g2p/language_map.json

@@ -0,0 +1,47 @@
+{
+  "american": "en_us",
+  "american_english": "en_us",
+  "british": "en_gb",
+  "british_english": "en_gb",
+  "de": "de",
+  "de_de": "de",
+  "deu": "de",
+  "en": "en_us",
+  "en-gb": "en_gb",
+  "en-uk": "en_gb",
+  "en-us": "en_us",
+  "en_gb": "en_gb",
+  "en_uk": "en_gb",
+  "en_us": "en_us",
+  "eng": "en_us",
+  "eng_gb": "en_gb",
+  "eng_us": "en_us",
+  "english": "en_us",
+  "english_gb": "en_gb",
+  "english_uk": "en_gb",
+  "english_us": "en_us",
+  "es": "es",
+  "es_es": "es",
+  "es_mx": "es",
+  "fr": "fr",
+  "fr_fr": "fr",
+  "fra": "fr",
+  "fre": "fr",
+  "french": "fr",
+  "gb": "en_gb",
+  "ger": "de",
+  "german": "de",
+  "it": "it",
+  "it_it": "it",
+  "ita": "it",
+  "italian": "it",
+  "spa": "es",
+  "spanish": "es",
+  "uk": "en_gb",
+  "us": "en_us",
+  "vi": "vi",
+  "vi_hn": "vi",
+  "vi_vn": "vi",
+  "vie": "vi",
+  "vietnamese": "vi"
+}

coreai/assets/g2p/normalization.json

@@ -0,0 +1,40 @@
+{
+  "contract": {
+    "colon_semicolon_classes": "preserved",
+    "dot_runs_two_or_more": "ellipsis_three_dots",
+    "homogeneous_terminal_runs": "single_mark",
+    "lexical_hyphen": "space_no_pause",
+    "mixed_question_exclamation": "leading_mark",
+    "non_acoustic_modifiers": [
+      "stress",
+      "length",
+      "unicode_combining"
+    ],
+    "ok_spoken_form": "okay",
+    "punctuation_silence_target": "explicit_silence",
+    "syntactic_dash": "em_dash_pause",
+    "training_phone_source": "espeak",
+    "version": "scyllasband_spoken_text_v2_2026_08_05"
+  },
+  "contract_sha256": "d8fd1b2415d0d100d1c70237a4469117ee6b73d0e4da703e4654c83192779cc4",
+  "contract_version": "scyllasband_spoken_text_v2_2026_08_05",
+  "expand_currency": true,
+  "expand_dates": true,
+  "expand_numbers": true,
+  "expand_ordinals": true,
+  "expand_percentages": true,
+  "expand_times": true,
+  "languages": [
+    "en",
+    "en_us",
+    "en_gb",
+    "es",
+    "it",
+    "fr",
+    "de",
+    "vi"
+  ],
+  "normalize_at_sign": true,
+  "normalize_punctuation": true,
+  "normalizer": "SpokenTextNormalizer"
+}

coreai/assets/g2p/phoneme_dict.json

@@ -0,0 +1,9 @@
+{
+  "de": {},
+  "en_gb": {},
+  "en_us": {},
+  "es": {},
+  "fr": {},
+  "it": {},
+  "vi": {}
+}

coreai/assets/g2p/pronunciation_overrides.json

@@ -0,0 +1,4 @@
+{
+  "format": "scyllasband_g2p_pronunciation_overrides_v1",
+  "languages": {}
+}

coreai/assets/g2p/tokenizer.json

@@ -0,0 +1,517 @@
+{
+  "format": "deepphonemizer_forward_ctc",
+  "languages": [
+    "de",
+    "en_gb",
+    "en_us",
+    "es",
+    "fr",
+    "it",
+    "vi"
+  ],
+  "char_repeats": 4,
+  "lowercase": true,
+  "phoneme_token_mode": "char",
+  "text_symbols": {
+    "_": 0,
+    "<de>": 1,
+    "<en_gb>": 2,
+    "<en_us>": 3,
+    "<es>": 4,
+    "<fr>": 5,
+    "<it>": 6,
+    "<vi>": 7,
+    "<end>": 8,
+    "\b": 9,
+    " ": 10,
+    "!": 11,
+    "\"": 12,
+    "#": 13,
+    "$": 14,
+    "%": 15,
+    "&": 16,
+    "'": 17,
+    "(": 18,
+    ")": 19,
+    "*": 20,
+    "+": 21,
+    ",": 22,
+    "-": 23,
+    ".": 24,
+    "/": 25,
+    "0": 26,
+    "1": 27,
+    "2": 28,
+    "3": 29,
+    "4": 30,
+    "5": 31,
+    "6": 32,
+    "7": 33,
+    "8": 34,
+    "9": 35,
+    ":": 36,
+    ";": 37,
+    "<": 38,
+    "=": 39,
+    ">": 40,
+    "?": 41,
+    "@": 42,
+    "A": 43,
+    "B": 44,
+    "C": 45,
+    "D": 46,
+    "E": 47,
+    "F": 48,
+    "G": 49,
+    "H": 50,
+    "I": 51,
+    "J": 52,
+    "K": 53,
+    "L": 54,
+    "M": 55,
+    "N": 56,
+    "O": 57,
+    "P": 58,
+    "Q": 59,
+    "R": 60,
+    "S": 61,
+    "T": 62,
+    "U": 63,
+    "V": 64,
+    "W": 65,
+    "X": 66,
+    "Y": 67,
+    "Z": 68,
+    "[": 69,
+    "\\": 70,
+    "]": 71,
+    "^": 72,
+    "`": 73,
+    "a": 74,
+    "b": 75,
+    "c": 76,
+    "d": 77,
+    "e": 78,
+    "f": 79,
+    "g": 80,
+    "h": 81,
+    "i": 82,
+    "j": 83,
+    "k": 84,
+    "l": 85,
+    "m": 86,
+    "n": 87,
+    "o": 88,
+    "p": 89,
+    "q": 90,
+    "r": 91,
+    "s": 92,
+    "t": 93,
+    "u": 94,
+    "v": 95,
+    "w": 96,
+    "x": 97,
+    "y": 98,
+    "z": 99,
+    "{": 100,
+    "|": 101,
+    "}": 102,
+    "~": 103,
+    "¡": 104,
+    "¢": 105,
+    "£": 106,
+    "¥": 107,
+    "§": 108,
+    "ª": 109,
+    "«": 110,
+    "­": 111,
+    "®": 112,
+    "°": 113,
+    "±": 114,
+    "²": 115,
+    "³": 116,
+    "´": 117,
+    "¶": 118,
+    "·": 119,
+    "¹": 120,
+    "º": 121,
+    "»": 122,
+    "¼": 123,
+    "½": 124,
+    "¾": 125,
+    "¿": 126,
+    "À": 127,
+    "Á": 128,
+    "Â": 129,
+    "Ä": 130,
+    "Å": 131,
+    "Æ": 132,
+    "Ç": 133,
+    "È": 134,
+    "É": 135,
+    "Ê": 136,
+    "Í": 137,
+    "Ï": 138,
+    "Ñ": 139,
+    "Ó": 140,
+    "Ô": 141,
+    "Ö": 142,
+    "×": 143,
+    "Ù": 144,
+    "Ú": 145,
+    "Ü": 146,
+    "Ý": 147,
+    "Þ": 148,
+    "ß": 149,
+    "à": 150,
+    "á": 151,
+    "â": 152,
+    "ã": 153,
+    "ä": 154,
+    "å": 155,
+    "æ": 156,
+    "ç": 157,
+    "è": 158,
+    "é": 159,
+    "ê": 160,
+    "ë": 161,
+    "ì": 162,
+    "í": 163,
+    "î": 164,
+    "ï": 165,
+    "ð": 166,
+    "ñ": 167,
+    "ò": 168,
+    "ó": 169,
+    "ô": 170,
+    "õ": 171,
+    "ö": 172,
+    "ø": 173,
+    "ù": 174,
+    "ú": 175,
+    "û": 176,
+    "ü": 177,
+    "ý": 178,
+    "þ": 179,
+    "ÿ": 180,
+    "ā": 181,
+    "Ă": 182,
+    "ă": 183,
+    "ą": 184,
+    "ć": 185,
+    "č": 186,
+    "Đ": 187,
+    "đ": 188,
+    "ē": 189,
+    "ė": 190,
+    "ę": 191,
+    "ğ": 192,
+    "ġ": 193,
+    "ĩ": 194,
+    "ī": 195,
+    "ļ": 196,
+    "ł": 197,
+    "ń": 198,
+    "Ō": 199,
+    "ō": 200,
+    "Œ": 201,
+    "œ": 202,
+    "ŕ": 203,
+    "ś": 204,
+    "ş": 205,
+    "š": 206,
+    "ŧ": 207,
+    "ũ": 208,
+    "ū": 209,
+    "ŭ": 210,
+    "ů": 211,
+    "ű": 212,
+    "ż": 213,
+    "Ž": 214,
+    "ž": 215,
+    "ſ": 216,
+    "Ơ": 217,
+    "ơ": 218,
+    "Ư": 219,
+    "ư": 220,
+    "ǁ": 221,
+    "ǂ": 222,
+    "ǐ": 223,
+    "ǘ": 224,
+    "ǧ": 225,
+    "ș": 226,
+    "Ȝ": 227,
+    "ȝ": 228,
+    "ȟ": 229,
+    "ɛ": 230,
+    "ʰ": 231,
+    "ˆ": 232,
+    "ˈ": 233,
+    "ˢ": 234,
+    "̃": 235,
+    "̍": 236,
+    "̣": 237,
+    "̱": 238,
+    "͈": 239,
+    "α": 240,
+    "β": 241,
+    "γ": 242,
+    "δ": 243,
+    "ε": 244,
+    "η": 245,
+    "κ": 246,
+    "μ": 247,
+    "ξ": 248,
+    "π": 249,
+    "ς": 250,
+    "σ": 251,
+    "χ": 252,
+    "ψ": 253,
+    "а": 254,
+    "д": 255,
+    "е": 256,
+    "з": 257,
+    "и": 258,
+    "к": 259,
+    "л": 260,
+    "о": 261,
+    "п": 262,
+    "р": 263,
+    "с": 264,
+    "т": 265,
+    "ч": 266,
+    "ь": 267,
+    "ง": 268,
+    "ย": 269,
+    "า": 270,
+    "่": 271,
+    "ည": 272,
+    "လ": 273,
+    "း": 274,
+    "်": 275,
+    "ᵈ": 276,
+    "ᵗ": 277,
+    "ḅ": 278,
+    "ḍ": 279,
+    "ḳ": 280,
+    "ṅ": 281,
+    "Ṭ": 282,
+    "ẓ": 283,
+    "ạ": 284,
+    "Ả": 285,
+    "ả": 286,
+    "Ấ": 287,
+    "ấ": 288,
+    "ầ": 289,
+    "ẩ": 290,
+    "ẫ": 291,
+    "ậ": 292,
+    "ắ": 293,
+    "ằ": 294,
+    "ẳ": 295,
+    "ẵ": 296,
+    "ặ": 297,
+    "ẹ": 298,
+    "ẻ": 299,
+    "ẽ": 300,
+    "Ế": 301,
+    "ế": 302,
+    "Ề": 303,
+    "ề": 304,
+    "ể": 305,
+    "ễ": 306,
+    "ệ": 307,
+    "ỉ": 308,
+    "ị": 309,
+    "ọ": 310,
+    "ỏ": 311,
+    "Ố": 312,
+    "ố": 313,
+    "Ồ": 314,
+    "ồ": 315,
+    "Ổ": 316,
+    "ổ": 317,
+    "ỗ": 318,
+    "Ộ": 319,
+    "ộ": 320,
+    "ớ": 321,
+    "Ờ": 322,
+    "ờ": 323,
+    "Ở": 324,
+    "ở": 325,
+    "ỡ": 326,
+    "ợ": 327,
+    "ụ": 328,
+    "ủ": 329,
+    "Ứ": 330,
+    "ứ": 331,
+    "Ừ": 332,
+    "ừ": 333,
+    "ử": 334,
+    "ữ": 335,
+    "ự": 336,
+    "ỳ": 337,
+    "ỵ": 338,
+    "Ỷ": 339,
+    "ỷ": 340,
+    "ỹ": 341,
+    "ὖ": 342,
+    "ὼ": 343,
+    "ῆ": 344,
+    "​": 345,
+    "‐": 346,
+    "—": 347,
+    "―": 348,
+    "†": 349,
+    "•": 350,
+    "…": 351,
+    "‰": 352,
+    "′": 353,
+    "″": 354,
+    "⁄": 355,
+    "⁷": 356,
+    "⁻": 357,
+    "₀": 358,
+    "₁": 359,
+    "₂": 360,
+    "₃": 361,
+    "₄": 362,
+    "₅": 363,
+    "₆": 364,
+    "₇": 365,
+    "₈": 366,
+    "€": 367,
+    "₹": 368,
+    "⅓": 369,
+    "⅔": 370,
+    "⅖": 371,
+    "⅙": 372,
+    "⅛": 373,
+    "⅝": 374,
+    "√": 375,
+    "│": 376,
+    "♦": 377,
+    "⚓": 378,
+    "✈": 379,
+    "✠": 380,
+    "✨": 381,
+    "⟙": 382,
+    "⸺": 383,
+    "。": 384,
+    "〈": 385,
+    "な": 386,
+    "三": 387,
+    "次": 388,
+    "肿": 389,
+    "臃": 390,
+    "구": 391,
+    "로": 392,
+    "요": 393,
+    "ﷺ": 394,
+    "️": 395,
+    "𝑥": 396,
+    "𝑦": 397,
+    "🌊": 398,
+    "🌿": 399,
+    "💦": 400,
+    "💨": 401
+  },
+  "phoneme_symbols": {
+    "0": "_",
+    "1": "<de>",
+    "2": "<en_gb>",
+    "3": "<en_us>",
+    "4": "<es>",
+    "5": "<fr>",
+    "6": "<it>",
+    "7": "<vi>",
+    "8": "<end>",
+    "9": " ",
+    "10": "!",
+    "11": ",",
+    "12": "-",
+    "13": ".",
+    "14": "1",
+    "15": "2",
+    "16": "4",
+    "17": "5",
+    "18": "6",
+    "19": "7",
+    "20": ":",
+    "21": ";",
+    "22": "?",
+    "23": "a",
+    "24": "b",
+    "25": "c",
+    "26": "d",
+    "27": "e",
+    "28": "f",
+    "29": "h",
+    "30": "i",
+    "31": "j",
+    "32": "k",
+    "33": "l",
+    "34": "m",
+    "35": "n",
+    "36": "o",
+    "37": "p",
+    "38": "q",
+    "39": "r",
+    "40": "s",
+    "41": "t",
+    "42": "u",
+    "43": "v",
+    "44": "w",
+    "45": "x",
+    "46": "y",
+    "47": "z",
+    "48": "«",
+    "49": "»",
+    "50": "æ",
+    "51": "ç",
+    "52": "ð",
+    "53": "ø",
+    "54": "ŋ",
+    "55": "œ",
+    "56": "ɐ",
+    "57": "ɑ",
+    "58": "ɒ",
+    "59": "ɔ",
+    "60": "ɗ",
+    "61": "ə",
+    "62": "ɚ",
+    "63": "ɛ",
+    "64": "ɜ",
+    "65": "ɟ",
+    "66": "ɡ",
+    "67": "ɣ",
+    "68": "ɨ",
+    "69": "ɪ",
+    "70": "ɬ",
+    "71": "ɯ",
+    "72": "ɲ",
+    "73": "ɹ",
+    "74": "ɾ",
+    "75": "ʁ",
+    "76": "ʃ",
+    "77": "ʊ",
+    "78": "ʌ",
+    "79": "ʎ",
+    "80": "ʐ",
+    "81": "ʒ",
+    "82": "ʔ",
+    "83": "ʝ",
+    "84": "ʲ",
+    "85": "ˈ",
+    "86": "ˌ",
+    "87": "ː",
+    "88": "̃",
+    "89": "̩",
+    "90": "̪",
+    "91": "β",
+    "92": "θ",
+    "93": "ᵻ",
+    "94": "—",
+    "95": "…"
+  }
+}

coreai/assets/languages.json

@@ -0,0 +1,30 @@
+[
+  {
+    "id": "en_us",
+    "index": 0
+  },
+  {
+    "id": "en_gb",
+    "index": 1
+  },
+  {
+    "id": "es",
+    "index": 2
+  },
+  {
+    "id": "it",
+    "index": 3
+  },
+  {
+    "id": "fr",
+    "index": 4
+  },
+  {
+    "id": "de",
+    "index": 5
+  },
+  {
+    "id": "vi",
+    "index": 6
+  }
+]

coreai/assets/phone_vocab.json

@@ -0,0 +1,111 @@
+{
+  "schema_version": 1,
+  "mode": "scyllasband_ipa_phone_symbols",
+  "special_tokens": [
+    "<pad>",
+    "<unk>",
+    "<sil>",
+    "<pause_comma>",
+    "<pause_semicolon>",
+    "<pause_colon>",
+    "<pause_dash>",
+    "<ellipsis>",
+    "<end_stmt>",
+    "<end_question>",
+    "<end_exclaim>",
+    "<ctx_sentence_start>",
+    "<ctx_continuation>",
+    "<ctx_sentence_end>",
+    "<ctx_chunk_continue>"
+  ],
+  "vocab_size": 87,
+  "token_to_id": {
+    "<pad>": 0,
+    "<unk>": 1,
+    "<sil>": 2,
+    "<pause_comma>": 3,
+    "<pause_semicolon>": 4,
+    "<pause_colon>": 5,
+    "<pause_dash>": 6,
+    "<ellipsis>": 7,
+    "<end_stmt>": 8,
+    "<end_question>": 9,
+    "<end_exclaim>": 10,
+    "<ctx_sentence_start>": 11,
+    "<ctx_continuation>": 12,
+    "<ctx_sentence_end>": 13,
+    "<ctx_chunk_continue>": 14,
+    "1": 15,
+    "2": 16,
+    "4": 17,
+    "5": 18,
+    "6": 19,
+    "7": 20,
+    "a": 21,
+    "b": 22,
+    "c": 23,
+    "d": 24,
+    "e": 25,
+    "f": 26,
+    "h": 27,
+    "i": 28,
+    "j": 29,
+    "k": 30,
+    "l": 31,
+    "m": 32,
+    "n": 33,
+    "o": 34,
+    "p": 35,
+    "r": 36,
+    "s": 37,
+    "t": 38,
+    "u": 39,
+    "v": 40,
+    "w": 41,
+    "x": 42,
+    "y": 43,
+    "z": 44,
+    "æ": 45,
+    "ç": 46,
+    "ð": 47,
+    "ø": 48,
+    "ŋ": 49,
+    "œ": 50,
+    "ɐ": 51,
+    "ɑ": 52,
+    "ɒ": 53,
+    "ɔ": 54,
+    "ɗ": 55,
+    "ə": 56,
+    "ɚ": 57,
+    "ɛ": 58,
+    "ɜ": 59,
+    "ɟ": 60,
+    "ɡ": 61,
+    "ɣ": 62,
+    "ɪ": 63,
+    "ɯ": 64,
+    "ɲ": 65,
+    "ɹ": 66,
+    "ɾ": 67,
+    "ʁ": 68,
+    "ʃ": 69,
+    "ʊ": 70,
+    "ʌ": 71,
+    "ʎ": 72,
+    "ʐ": 73,
+    "ʒ": 74,
+    "ʔ": 75,
+    "ʝ": 76,
+    "ʲ": 77,
+    "ˈ": 78,
+    "ˌ": 79,
+    "ː": 80,
+    "̃": 81,
+    "̩": 82,
+    "̪": 83,
+    "β": 84,
+    "θ": 85,
+    "ᵻ": 86
+  }
+}

coreai/assets/voices.json

@@ -0,0 +1,42 @@
+[
+  {
+    "id": "ariadne",
+    "index": 0
+  },
+  {
+    "id": "felix",
+    "index": 1
+  },
+  {
+    "id": "gwen",
+    "index": 2
+  },
+  {
+    "id": "ink",
+    "index": 3
+  },
+  {
+    "id": "max",
+    "index": 4
+  },
+  {
+    "id": "orpheus",
+    "index": 5
+  },
+  {
+    "id": "rex",
+    "index": 6
+  },
+  {
+    "id": "scylla",
+    "index": 7
+  },
+  {
+    "id": "stone",
+    "index": 8
+  },
+  {
+    "id": "tuesday",
+    "index": 9
+  }
+]

coreai/coreai/duration_predictor.aimodel/main.hash

@@ -0,0 +1 @@
+BC�6�����3����KS��m�X�����I

coreai/coreai/duration_predictor.aimodel/main.mlirb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4243d405361c9ebc82a0e233a3cb1bd5d64b5389a96ddd58947f14fb84b3c949
+size 10884924

coreai/coreai/duration_predictor.aimodel/metadata.json

@@ -0,0 +1,5 @@
+{
+  "assetVersion" : "2.0",
+  "producer" : "coreai-core 1.0.0b3",
+  "creationDate" : "20261005T221157Z"
+}

coreai/coreai/g2p.aimodel/main.hash

@@ -0,0 +1 @@
+�Bj�Yp*��vB

coreai/coreai/g2p.aimodel/main.mlirb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e7426aa159702a859176420d0c34b493da7d3dd97fa74fdeedb2ee4c4f0dce18
+size 28377271

coreai/coreai/g2p.aimodel/metadata.json

@@ -0,0 +1,5 @@
+{
+  "producer" : "coreai-core 1.0.0b3",
+  "creationDate" : "20261005T221156Z",
+  "assetVersion" : "2.0"
+}

coreai/coreai/vector_context_encoder.aimodel/main.hash

@@ -0,0 +1 @@
+��@�,�� ����4��r�YM�E�Ϡm����5

coreai/coreai/vector_context_encoder.aimodel/main.mlirb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9dca40e4aa2ce1cf2099c7c4ee34bbea72f2594d88459210cfa06dfc9592f135
+size 8029277

coreai/coreai/vector_context_encoder.aimodel/metadata.json

@@ -0,0 +1,5 @@
+{
+  "creationDate" : "20261005T221159Z",
+  "assetVersion" : "2.0",
+  "producer" : "coreai-core 1.0.0b3"
+}

coreai/coreai/vector_estimator.aimodel/main.hash

@@ -0,0 +1 @@
+M��\��-:��G�

coreai/coreai/vector_estimator.aimodel/main.mlirb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4dc8e41c5c9de01f2d3a9fd847a30df8b27ceec9aa0894b235f49181dcd664d2
+size 63149068

coreai/coreai/vector_estimator.aimodel/metadata.json

@@ -0,0 +1,5 @@
+{
+  "assetVersion" : "2.0",
+  "creationDate" : "20261005T221212Z",
+  "producer" : "coreai-core 1.0.0b3"
+}

coreai/coreai/vocoder.aimodel/main.hash

@@ -0,0 +1 @@
+

coreai/coreai/vocoder.aimodel/main.mlirb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0e0d07b91327322c55da5081c7e3301565e92c5298dd9a5dfb74b71ee6074669
+size 43811030

coreai/coreai/vocoder.aimodel/metadata.json

@@ -0,0 +1,5 @@
+{
+  "creationDate" : "20261005T221220Z",
+  "producer" : "coreai-core 1.0.0b3",
+  "assetVersion" : "2.0"
+}

coreai/manifest.json

@@ -0,0 +1,1003 @@
+{
+  "architecture": "scyllasband-duration-flow",
+  "assets": {
+    "fixed_delivery_references": "assets/fixed_references.json",
+    "g2p_config": "assets/g2p/export_config.json",
+    "g2p_language_map": "assets/g2p/language_map.json",
+    "g2p_normalization": "assets/g2p/normalization.json",
+    "g2p_phoneme_dict": "assets/g2p/phoneme_dict.json",
+    "g2p_pronunciation_overrides": "assets/g2p/pronunciation_overrides.json",
+    "g2p_tokenizer": "assets/g2p/tokenizer.json",
+    "language_index": "assets/languages.json",
+    "phone_vocab": "assets/phone_vocab.json",
+    "voice_index": "assets/voices.json"
+  },
+  "audio": {
+    "hop_length": 256,
+    "latent_dim": 24,
+    "latent_hop_length": 512,
+    "n_mels": 100,
+    "sample_rate": 24000
+  },
+  "components": {
+    "g2p_64": {
+      "artifacts": {
+        "coreai": {
+          "format": "coreai",
+          "path": "coreai/g2p.aimodel",
+          "function": "g2p_64"
+        }
+      },
+      "inputs": [
+        "text"
+      ],
+      "outputs": [
+        "logits"
+      ],
+      "required": false
+    },
+    "g2p_128": {
+      "artifacts": {
+        "coreai": {
+          "format": "coreai",
+          "path": "coreai/g2p.aimodel",
+          "function": "g2p_128"
+        }
+      },
+      "inputs": [
+        "text"
+      ],
+      "outputs": [
+        "logits"
+      ],
+      "required": false
+    },
+    "g2p_256": {
+      "artifacts": {
+        "coreai": {
+          "format": "coreai",
+          "path": "coreai/g2p.aimodel",
+          "function": "g2p_256"
+        }
+      },
+      "inputs": [
+        "text"
+      ],
+      "outputs": [
+        "logits"
+      ],
+      "required": false
+    },
+    "g2p": {
+      "artifacts": {
+        "coreai": {
+          "format": "coreai",
+          "path": "coreai/g2p.aimodel",
+          "function": "g2p"
+        }
+      },
+      "inputs": [
+        "text"
+      ],
+      "outputs": [
+        "logits"
+      ],
+      "required": true
+    },
+    "duration_predictor": {
+      "artifacts": {
+        "coreai": {
+          "format": "coreai",
+          "path": "coreai/duration_predictor.aimodel",
+          "function": "duration_predictor"
+        }
+      },
+      "inputs": [
+        "span_phone_ids",
+        "span_segment_ids",
+        "span_mask",
+        "voice_id",
+        "language_id",
+        "delivery_values",
+        "delivery_present"
+      ],
+      "outputs": [
+        "durations"
+      ],
+      "required": true
+    },
+    "vector_context_encoder": {
+      "artifacts": {
+        "coreai": {
+          "format": "coreai",
+          "path": "coreai/vector_context_encoder.aimodel",
+          "function": "vector_context_encoder"
+        }
+      },
+      "inputs": [
+        "span_context_phone_ids",
+        "span_context_segment_ids",
+        "span_context_mask"
+      ],
+      "outputs": [
+        "span_context_hidden"
+      ],
+      "required": true
+    },
+    "vector_estimator_128": {
+      "artifacts": {
+        "coreai": {
+          "format": "coreai",
+          "path": "coreai/vector_estimator.aimodel",
+          "function": "vector_estimator_128"
+        }
+      },
+      "inputs": [
+        "noise",
+        "time",
+        "expanded_phone_ids",
+        "voice_id",
+        "language_id",
+        "delivery_values",
+        "delivery_present",
+        "latent_mask",
+        "span_context_hidden",
+        "prefix_latents",
+        "prefix_mask",
+        "expanded_boundary_event_ids",
+        "expanded_modifier_event_ids",
+        "expanded_phone_phase",
+        "expanded_phone_log_duration",
+        "expanded_sentence_type_ids"
+      ],
+      "outputs": [
+        "velocity"
+      ],
+      "required": false
+    },
+    "vocoder_128": {
+      "artifacts": {
+        "coreai": {
+          "format": "coreai",
+          "path": "coreai/vocoder.aimodel",
+          "function": "vocoder_128"
+        }
+      },
+      "inputs": [
+        "latents",
+        "latent_mask",
+        "voice_id",
+        "language_id",
+        "emotion_id"
+      ],
+      "outputs": [
+        "audio"
+      ],
+      "required": false
+    },
+    "vector_estimator_256": {
+      "artifacts": {
+        "coreai": {
+          "format": "coreai",
+          "path": "coreai/vector_estimator.aimodel",
+          "function": "vector_estimator_256"
+        }
+      },
+      "inputs": [
+        "noise",
+        "time",
+        "expanded_phone_ids",
+        "voice_id",
+        "language_id",
+        "delivery_values",
+        "delivery_present",
+        "latent_mask",
+        "span_context_hidden",
+        "prefix_latents",
+        "prefix_mask",
+        "expanded_boundary_event_ids",
+        "expanded_modifier_event_ids",
+        "expanded_phone_phase",
+        "expanded_phone_log_duration",
+        "expanded_sentence_type_ids"
+      ],
+      "outputs": [
+        "velocity"
+      ],
+      "required": false
+    },
+    "vocoder_256": {
+      "artifacts": {
+        "coreai": {
+          "format": "coreai",
+          "path": "coreai/vocoder.aimodel",
+          "function": "vocoder_256"
+        }
+      },
+      "inputs": [
+        "latents",
+        "latent_mask",
+        "voice_id",
+        "language_id",
+        "emotion_id"
+      ],
+      "outputs": [
+        "audio"
+      ],
+      "required": false
+    },
+    "vector_estimator_384": {
+      "artifacts": {
+        "coreai": {
+          "format": "coreai",
+          "path": "coreai/vector_estimator.aimodel",
+          "function": "vector_estimator_384"
+        }
+      },
+      "inputs": [
+        "noise",
+        "time",
+        "expanded_phone_ids",
+        "voice_id",
+        "language_id",
+        "delivery_values",
+        "delivery_present",
+        "latent_mask",
+        "span_context_hidden",
+        "prefix_latents",
+        "prefix_mask",
+        "expanded_boundary_event_ids",
+        "expanded_modifier_event_ids",
+        "expanded_phone_phase",
+        "expanded_phone_log_duration",
+        "expanded_sentence_type_ids"
+      ],
+      "outputs": [
+        "velocity"
+      ],
+      "required": false
+    },
+    "vocoder_384": {
+      "artifacts": {
+        "coreai": {
+          "format": "coreai",
+          "path": "coreai/vocoder.aimodel",
+          "function": "vocoder_384"
+        }
+      },
+      "inputs": [
+        "latents",
+        "latent_mask",
+        "voice_id",
+        "language_id",
+        "emotion_id"
+      ],
+      "outputs": [
+        "audio"
+      ],
+      "required": false
+    },
+    "vector_estimator_512": {
+      "artifacts": {
+        "coreai": {
+          "format": "coreai",
+          "path": "coreai/vector_estimator.aimodel",
+          "function": "vector_estimator_512"
+        }
+      },
+      "inputs": [
+        "noise",
+        "time",
+        "expanded_phone_ids",
+        "voice_id",
+        "language_id",
+        "delivery_values",
+        "delivery_present",
+        "latent_mask",
+        "span_context_hidden",
+        "prefix_latents",
+        "prefix_mask",
+        "expanded_boundary_event_ids",
+        "expanded_modifier_event_ids",
+        "expanded_phone_phase",
+        "expanded_phone_log_duration",
+        "expanded_sentence_type_ids"
+      ],
+      "outputs": [
+        "velocity"
+      ],
+      "required": false
+    },
+    "vocoder_512": {
+      "artifacts": {
+        "coreai": {
+          "format": "coreai",
+          "path": "coreai/vocoder.aimodel",
+          "function": "vocoder_512"
+        }
+      },
+      "inputs": [
+        "latents",
+        "latent_mask",
+        "voice_id",
+        "language_id",
+        "emotion_id"
+      ],
+      "outputs": [
+        "audio"
+      ],
+      "required": false
+    },
+    "vector_estimator_768": {
+      "artifacts": {
+        "coreai": {
+          "format": "coreai",
+          "path": "coreai/vector_estimator.aimodel",
+          "function": "vector_estimator_768"
+        }
+      },
+      "inputs": [
+        "noise",
+        "time",
+        "expanded_phone_ids",
+        "voice_id",
+        "language_id",
+        "delivery_values",
+        "delivery_present",
+        "latent_mask",
+        "span_context_hidden",
+        "prefix_latents",
+        "prefix_mask",
+        "expanded_boundary_event_ids",
+        "expanded_modifier_event_ids",
+        "expanded_phone_phase",
+        "expanded_phone_log_duration",
+        "expanded_sentence_type_ids"
+      ],
+      "outputs": [
+        "velocity"
+      ],
+      "required": false
+    },
+    "vocoder_768": {
+      "artifacts": {
+        "coreai": {
+          "format": "coreai",
+          "path": "coreai/vocoder.aimodel",
+          "function": "vocoder_768"
+        }
+      },
+      "inputs": [
+        "latents",
+        "latent_mask",
+        "voice_id",
+        "language_id",
+        "emotion_id"
+      ],
+      "outputs": [
+        "audio"
+      ],
+      "required": false
+    },
+    "vector_estimator": {
+      "artifacts": {
+        "coreai": {
+          "format": "coreai",
+          "path": "coreai/vector_estimator.aimodel",
+          "function": "vector_estimator"
+        }
+      },
+      "inputs": [
+        "noise",
+        "time",
+        "expanded_phone_ids",
+        "voice_id",
+        "language_id",
+        "delivery_values",
+        "delivery_present",
+        "latent_mask",
+        "span_context_hidden",
+        "prefix_latents",
+        "prefix_mask",
+        "expanded_boundary_event_ids",
+        "expanded_modifier_event_ids",
+        "expanded_phone_phase",
+        "expanded_phone_log_duration",
+        "expanded_sentence_type_ids"
+      ],
+      "outputs": [
+        "velocity"
+      ],
+      "required": true
+    },
+    "vocoder": {
+      "artifacts": {
+        "coreai": {
+          "format": "coreai",
+          "path": "coreai/vocoder.aimodel",
+          "function": "vocoder"
+        }
+      },
+      "inputs": [
+        "latents",
+        "latent_mask",
+        "voice_id",
+        "language_id",
+        "emotion_id"
+      ],
+      "outputs": [
+        "audio"
+      ],
+      "required": true
+    }
+  },
+  "contract_version": "1.0.0",
+  "controls": {
+    "chunk_context": {
+      "artificial_continuation_boundaries": [
+        "chunk_continue"
+      ],
+      "artificial_continuation_phone": null,
+      "boundary_after_values": [
+        "sentence_end",
+        "paragraph_end",
+        "clause_continue",
+        "chunk_continue"
+      ],
+      "boundary_before_values": [
+        "sentence_start",
+        "paragraph_start",
+        "clause_continue",
+        "chunk_continue"
+      ],
+      "clause_continuation_fallback_phone": "<pause_comma>",
+      "conditioning": "boundary_embeddings",
+      "context_phone_tokens": {
+        "continuation": "<ctx_continuation>",
+        "sentence_end": "<ctx_sentence_end>",
+        "sentence_start": "<ctx_sentence_start>"
+      },
+      "default_boundary_after": "sentence_end",
+      "default_boundary_before": "sentence_start",
+      "emit_context_phone_tokens": false,
+      "linguistic_continuation_boundaries": [
+        "clause_continue"
+      ],
+      "schema_version": 2
+    },
+    "chunking": {
+      "shared_split_phone": "<sil>",
+      "split_after": [
+        "<end_stmt>",
+        "<end_question>",
+        "<end_exclaim>",
+        "<ellipsis>"
+      ],
+      "unit": "sentence"
+    },
+    "decoding": {
+      "context_frames": 48,
+      "mode": "overlap"
+    },
+    "delivery": {
+      "axes": [
+        "energy",
+        "tension",
+        "valence",
+        "assertiveness"
+      ],
+      "default": "neutral",
+      "enabled": true,
+      "neutral": 2.0,
+      "normalization": "(rating-2)/4",
+      "presence_mask": true,
+      "range": [
+        0.0,
+        4.0
+      ],
+      "release_id": "v2-20261005",
+      "scope": "utterance",
+      "timeline_conditioning": false,
+      "voice_locales": {
+        "ariadne": [
+          "de",
+          "en_us",
+          "es",
+          "fr",
+          "it",
+          "vi"
+        ],
+        "felix": [
+          "de",
+          "en_us",
+          "es",
+          "fr",
+          "it",
+          "vi"
+        ],
+        "gwen": [
+          "de",
+          "en_us",
+          "es",
+          "fr",
+          "it",
+          "vi"
+        ],
+        "ink": [
+          "de",
+          "en_gb",
+          "es",
+          "fr",
+          "it",
+          "vi"
+        ],
+        "max": [
+          "de",
+          "en_us",
+          "es",
+          "fr",
+          "it",
+          "vi"
+        ],
+        "orpheus": [
+          "de",
+          "en_gb",
+          "es",
+          "fr",
+          "it",
+          "vi"
+        ],
+        "rex": [
+          "de",
+          "en_us",
+          "es",
+          "fr",
+          "it",
+          "vi"
+        ],
+        "scylla": [
+          "de",
+          "en_us",
+          "es",
+          "fr",
+          "it",
+          "vi"
+        ],
+        "stone": [
+          "de",
+          "en_us",
+          "es",
+          "fr",
+          "it",
+          "vi"
+        ],
+        "tuesday": [
+          "de",
+          "en_gb",
+          "es",
+          "fr",
+          "it",
+          "vi"
+        ]
+      },
+      "whisper": "binary"
+    },
+    "durations": {
+      "rounding": "half_even",
+      "silence_min_frames": 0,
+      "speech_min_frames": 1,
+      "zero_frame_phones": "punctuation_and_modifiers"
+    },
+    "emotions": {
+      "enabled": false
+    },
+    "energy": {
+      "default": 1.0,
+      "max": 1.5,
+      "min": 0.5
+    },
+    "fixed_shapes": {
+      "g2p_text_tokens": 512,
+      "latent_frames": 1024,
+      "phone_frames": 512
+    },
+    "graph_input_contract": "scyllasband_measured_delivery_v2",
+    "pitch": {
+      "default": 0.0,
+      "max": 4.0,
+      "min": -4.0
+    },
+    "prefix_conditioning": {
+      "alignment": "right",
+      "enabled": true,
+      "max_frames": 96,
+      "source": "previous_latents_tail"
+    },
+    "punctuation_silence": {
+      "calibrated_floors": false,
+      "floor_table_ms": {},
+      "silence_phone": "<sil>",
+      "target": "explicit_silence"
+    },
+    "reference_packs": {
+      "enabled": false,
+      "fixed_in_graph": true
+    },
+    "sampler": {
+      "choices": [
+        "euler",
+        "heun"
+      ],
+      "default": "heun"
+    },
+    "span_conditioning": {
+      "context_max_phones": 512,
+      "context_phones_each_side": 180,
+      "duration_inputs": [
+        "span_phone_ids",
+        "span_segment_ids",
+        "span_mask"
+      ],
+      "enabled": true,
+      "encoder_component": "vector_context_encoder",
+      "encoder_inputs": [
+        "span_context_phone_ids",
+        "span_context_segment_ids",
+        "span_context_mask"
+      ],
+      "exclude_context_phones": [
+        "<sil>"
+      ],
+      "hidden_size": 512,
+      "layout": "balanced_before_target_after",
+      "scope": "passage",
+      "segments": 3,
+      "solver_input": "span_context_hidden"
+    },
+    "speed": {
+      "default": 1.0,
+      "max": 1.25,
+      "min": 0.75
+    },
+    "steps": {
+      "default": 8,
+      "max": 64,
+      "min": 1
+    },
+    "target_buckets": {
+      "bucket_margin_frames": 0,
+      "buckets": [
+        {
+          "default": false,
+          "latent_frames": 128,
+          "vector_estimator": "vector_estimator_128",
+          "vocoder": "vocoder_128"
+        },
+        {
+          "default": false,
+          "latent_frames": 256,
+          "vector_estimator": "vector_estimator_256",
+          "vocoder": "vocoder_256"
+        },
+        {
+          "default": false,
+          "latent_frames": 384,
+          "vector_estimator": "vector_estimator_384",
+          "vocoder": "vocoder_384"
+        },
+        {
+          "default": false,
+          "latent_frames": 512,
+          "vector_estimator": "vector_estimator_512",
+          "vocoder": "vocoder_512"
+        },
+        {
+          "default": false,
+          "latent_frames": 768,
+          "vector_estimator": "vector_estimator_768",
+          "vocoder": "vocoder_768"
+        },
+        {
+          "default": true,
+          "latent_frames": 1024,
+          "vector_estimator": "vector_estimator",
+          "vocoder": "vocoder"
+        }
+      ],
+      "default_latent_frames": 1024,
+      "enabled": true,
+      "policy": "smallest_fit"
+    },
+    "text_normalizer": {
+      "prepared_contract_recorded": true,
+      "sha256": "d8fd1b2415d0d100d1c70237a4469117ee6b73d0e4da703e4654c83192779cc4",
+      "version": "scyllasband_spoken_text_v2_2026_08_05"
+    },
+    "timing_events": {
+      "enabled": true,
+      "inputs": [
+        "expanded_boundary_event_ids",
+        "expanded_modifier_event_ids",
+        "expanded_phone_phase",
+        "expanded_phone_log_duration",
+        "expanded_sentence_type_ids"
+      ],
+      "modifier_bits": {
+        "ˈ": 1,
+        "ˌ": 2,
+        "ː": 4,
+        "̃": 8,
+        "̩": 16,
+        "̪": 32
+      },
+      "punctuation_owner": "following_silence_else_first_frame_of_next_phone",
+      "sentence_type_count": 5,
+      "sentence_type_lookahead": "following_context",
+      "sentence_types": {
+        "<ellipsis>": 4,
+        "<end_exclaim>": 3,
+        "<end_question>": 2,
+        "<end_stmt>": 1
+      },
+      "word_start_event_id": 87,
+      "word_start_owner": "first_phone_of_word_after_chunk_first_word"
+    },
+    "word_boundaries": {
+      "duration_phone": "<sil>",
+      "enabled": true,
+      "g2p_output_symbol": " ",
+      "zero_frames_allowed": true
+    },
+    "g2p_buckets": [
+      {
+        "text_tokens": 64,
+        "component": "g2p_64"
+      },
+      {
+        "text_tokens": 128,
+        "component": "g2p_128"
+      },
+      {
+        "text_tokens": 256,
+        "component": "g2p_256"
+      },
+      {
+        "text_tokens": 512,
+        "component": "g2p"
+      }
+    ],
+    "coreai": {
+      "minimum_os": {
+        "ios": "27.0",
+        "macos": "27.0"
+      },
+      "io": {
+        "ids": "int32",
+        "masks": "int32",
+        "values": "float32"
+      },
+      "assets": {
+        "g2p": {
+          "precision": "fp16",
+          "weights": "w8",
+          "bytes": 28377408
+        },
+        "duration_predictor": {
+          "precision": "fp16",
+          "weights": "w8",
+          "bytes": 10885061
+        },
+        "vector_context_encoder": {
+          "precision": "fp16",
+          "weights": "w8",
+          "bytes": 8029414
+        },
+        "vector_estimator": {
+          "precision": "fp16",
+          "weights": "w8",
+          "bytes": 63149205
+        },
+        "vocoder": {
+          "precision": "fp16",
+          "weights": "w8",
+          "bytes": 43811167
+        }
+      },
+      "compute_units": "gpu",
+      "neural_engine_assets": [
+        "vector_estimator"
+      ]
+    }
+  },
+  "dataset_summary": {
+    "controls": "measured_four_axes_plus_binary_whisper",
+    "voice_locale_pairs": 60
+  },
+  "default_language": "en_us",
+  "languages": [
+    "en_us",
+    "en_gb",
+    "es",
+    "it",
+    "fr",
+    "de",
+    "vi"
+  ],
+  "model_name": "Scylla's Band",
+  "model_version": "2",
+  "preferred_backends": [
+    "coreai"
+  ],
+  "voices": [
+    {
+      "default_language": "en_us",
+      "embedding_path": null,
+      "id": "ariadne",
+      "languages": [
+        "de",
+        "en_us",
+        "es",
+        "fr",
+        "it",
+        "vi"
+      ],
+      "name": "ariadne",
+      "styles": [
+        "default"
+      ]
+    },
+    {
+      "default_language": "en_us",
+      "embedding_path": null,
+      "id": "felix",
+      "languages": [
+        "de",
+        "en_us",
+        "es",
+        "fr",
+        "it",
+        "vi"
+      ],
+      "name": "felix",
+      "styles": [
+        "default"
+      ]
+    },
+    {
+      "default_language": "en_us",
+      "embedding_path": null,
+      "id": "gwen",
+      "languages": [
+        "de",
+        "en_us",
+        "es",
+        "fr",
+        "it",
+        "vi"
+      ],
+      "name": "gwen",
+      "styles": [
+        "default"
+      ]
+    },
+    {
+      "default_language": "en_gb",
+      "embedding_path": null,
+      "id": "ink",
+      "languages": [
+        "de",
+        "en_gb",
+        "es",
+        "fr",
+        "it",
+        "vi"
+      ],
+      "name": "ink",
+      "styles": [
+        "default"
+      ]
+    },
+    {
+      "default_language": "en_us",
+      "embedding_path": null,
+      "id": "max",
+      "languages": [
+        "de",
+        "en_us",
+        "es",
+        "fr",
+        "it",
+        "vi"
+      ],
+      "name": "max",
+      "styles": [
+        "default"
+      ]
+    },
+    {
+      "default_language": "en_gb",
+      "embedding_path": null,
+      "id": "orpheus",
+      "languages": [
+        "de",
+        "en_gb",
+        "es",
+        "fr",
+        "it",
+        "vi"
+      ],
+      "name": "orpheus",
+      "styles": [
+        "default"
+      ]
+    },
+    {
+      "default_language": "en_us",
+      "embedding_path": null,
+      "id": "rex",
+      "languages": [
+        "de",
+        "en_us",
+        "es",
+        "fr",
+        "it",
+        "vi"
+      ],
+      "name": "rex",
+      "styles": [
+        "default"
+      ]
+    },
+    {
+      "default_language": "en_us",
+      "embedding_path": null,
+      "id": "scylla",
+      "languages": [
+        "de",
+        "en_us",
+        "es",
+        "fr",
+        "it",
+        "vi"
+      ],
+      "name": "scylla",
+      "styles": [
+        "default"
+      ]
+    },
+    {
+      "default_language": "en_us",
+      "embedding_path": null,
+      "id": "stone",
+      "languages": [
+        "de",
+        "en_us",
+        "es",
+        "fr",
+        "it",
+        "vi"
+      ],
+      "name": "stone",
+      "styles": [
+        "default"
+      ]
+    },
+    {
+      "default_language": "en_gb",
+      "embedding_path": null,
+      "id": "tuesday",
+      "languages": [
+        "de",
+        "en_gb",
+        "es",
+        "fr",
+        "it",
+        "vi"
+      ],
+      "name": "tuesday",
+      "styles": [
+        "default"
+      ]
+    }
+  ]
+}

coreml/assets/g2p/export_config.json

@@ -0,0 +1,33 @@
+{
+  "batch_dynamic": true,
+  "format": "scyllasband_g2p",
+  "input_granularity": "phrase",
+  "languages": [
+    "de",
+    "en_gb",
+    "en_us",
+    "es",
+    "fr",
+    "it",
+    "vi"
+  ],
+  "model_type": "transformer",
+  "onnx_export_batch_size": 4,
+  "onnx_export_layout": "batch_first",
+  "onnx_export_max_batch_size": 256,
+  "onnx_exporter": "dynamo",
+  "onnx_fp32_model": "model.onnx",
+  "onnx_model": "model.onnx",
+  "onnx_padded_output": "ctc_blank",
+  "onnx_quantization": "none",
+  "onnx_quantized_parity_tolerance": 3.0,
+  "opset": 18,
+  "phoneme_token_mode": "char",
+  "phrase_rows_per_language": 200000,
+  "preserve_commas": true,
+  "punctuation_silence_target": "explicit_silence",
+  "release_role": "fp32_bundle_frontend",
+  "source_checkpoint": "output/deepphonemizer_multilingual/model_v10_7lang/checkpoints/model_step_250k.pt",
+  "training_mix": "phrase_word_chunked_clean",
+  "word_rows_per_language": 200000
+}

coreml/assets/g2p/language_map.json

@@ -0,0 +1,47 @@
+{
+  "american": "en_us",
+  "american_english": "en_us",
+  "british": "en_gb",
+  "british_english": "en_gb",
+  "de": "de",
+  "de_de": "de",
+  "deu": "de",
+  "en": "en_us",
+  "en-gb": "en_gb",
+  "en-uk": "en_gb",
+  "en-us": "en_us",
+  "en_gb": "en_gb",
+  "en_uk": "en_gb",
+  "en_us": "en_us",
+  "eng": "en_us",
+  "eng_gb": "en_gb",
+  "eng_us": "en_us",
+  "english": "en_us",
+  "english_gb": "en_gb",
+  "english_uk": "en_gb",
+  "english_us": "en_us",
+  "es": "es",
+  "es_es": "es",
+  "es_mx": "es",
+  "fr": "fr",
+  "fr_fr": "fr",
+  "fra": "fr",
+  "fre": "fr",
+  "french": "fr",
+  "gb": "en_gb",
+  "ger": "de",
+  "german": "de",
+  "it": "it",
+  "it_it": "it",
+  "ita": "it",
+  "italian": "it",
+  "spa": "es",
+  "spanish": "es",
+  "uk": "en_gb",
+  "us": "en_us",
+  "vi": "vi",
+  "vi_hn": "vi",
+  "vi_vn": "vi",
+  "vie": "vi",
+  "vietnamese": "vi"
+}

coreml/assets/g2p/normalization.json

@@ -0,0 +1,40 @@
+{
+  "contract": {
+    "colon_semicolon_classes": "preserved",
+    "dot_runs_two_or_more": "ellipsis_three_dots",
+    "homogeneous_terminal_runs": "single_mark",
+    "lexical_hyphen": "space_no_pause",
+    "mixed_question_exclamation": "leading_mark",
+    "non_acoustic_modifiers": [
+      "stress",
+      "length",
+      "unicode_combining"
+    ],
+    "ok_spoken_form": "okay",
+    "punctuation_silence_target": "explicit_silence",
+    "syntactic_dash": "em_dash_pause",
+    "training_phone_source": "espeak",
+    "version": "scyllasband_spoken_text_v2_2026_08_05"
+  },
+  "contract_sha256": "d8fd1b2415d0d100d1c70237a4469117ee6b73d0e4da703e4654c83192779cc4",
+  "contract_version": "scyllasband_spoken_text_v2_2026_08_05",
+  "expand_currency": true,
+  "expand_dates": true,
+  "expand_numbers": true,
+  "expand_ordinals": true,
+  "expand_percentages": true,
+  "expand_times": true,
+  "languages": [
+    "en",
+    "en_us",
+    "en_gb",
+    "es",
+    "it",
+    "fr",
+    "de",
+    "vi"
+  ],
+  "normalize_at_sign": true,
+  "normalize_punctuation": true,
+  "normalizer": "SpokenTextNormalizer"
+}

coreml/assets/g2p/phoneme_dict.json

@@ -0,0 +1,9 @@
+{
+  "de": {},
+  "en_gb": {},
+  "en_us": {},
+  "es": {},
+  "fr": {},
+  "it": {},
+  "vi": {}
+}

coreml/assets/g2p/pronunciation_overrides.json

@@ -0,0 +1,4 @@
+{
+  "format": "scyllasband_g2p_pronunciation_overrides_v1",
+  "languages": {}
+}

coreml/assets/g2p/tokenizer.json

@@ -0,0 +1,517 @@
+{
+  "format": "deepphonemizer_forward_ctc",
+  "languages": [
+    "de",
+    "en_gb",
+    "en_us",
+    "es",
+    "fr",
+    "it",
+    "vi"
+  ],
+  "char_repeats": 4,
+  "lowercase": true,
+  "phoneme_token_mode": "char",
+  "text_symbols": {
+    "_": 0,
+    "<de>": 1,
+    "<en_gb>": 2,
+    "<en_us>": 3,
+    "<es>": 4,
+    "<fr>": 5,
+    "<it>": 6,
+    "<vi>": 7,
+    "<end>": 8,
+    "\b": 9,
+    " ": 10,
+    "!": 11,
+    "\"": 12,
+    "#": 13,
+    "$": 14,
+    "%": 15,
+    "&": 16,
+    "'": 17,
+    "(": 18,
+    ")": 19,
+    "*": 20,
+    "+": 21,
+    ",": 22,
+    "-": 23,
+    ".": 24,
+    "/": 25,
+    "0": 26,
+    "1": 27,
+    "2": 28,
+    "3": 29,
+    "4": 30,
+    "5": 31,
+    "6": 32,
+    "7": 33,
+    "8": 34,
+    "9": 35,
+    ":": 36,
+    ";": 37,
+    "<": 38,
+    "=": 39,
+    ">": 40,
+    "?": 41,
+    "@": 42,
+    "A": 43,
+    "B": 44,
+    "C": 45,
+    "D": 46,
+    "E": 47,
+    "F": 48,
+    "G": 49,
+    "H": 50,
+    "I": 51,
+    "J": 52,
+    "K": 53,
+    "L": 54,
+    "M": 55,
+    "N": 56,
+    "O": 57,
+    "P": 58,
+    "Q": 59,
+    "R": 60,
+    "S": 61,
+    "T": 62,
+    "U": 63,
+    "V": 64,
+    "W": 65,
+    "X": 66,
+    "Y": 67,
+    "Z": 68,
+    "[": 69,
+    "\\": 70,
+    "]": 71,
+    "^": 72,
+    "`": 73,
+    "a": 74,
+    "b": 75,
+    "c": 76,
+    "d": 77,
+    "e": 78,
+    "f": 79,
+    "g": 80,
+    "h": 81,
+    "i": 82,
+    "j": 83,
+    "k": 84,
+    "l": 85,
+    "m": 86,
+    "n": 87,
+    "o": 88,
+    "p": 89,
+    "q": 90,
+    "r": 91,
+    "s": 92,
+    "t": 93,
+    "u": 94,
+    "v": 95,
+    "w": 96,
+    "x": 97,
+    "y": 98,
+    "z": 99,
+    "{": 100,
+    "|": 101,
+    "}": 102,
+    "~": 103,
+    "¡": 104,
+    "¢": 105,
+    "£": 106,
+    "¥": 107,
+    "§": 108,
+    "ª": 109,
+    "«": 110,
+    "­": 111,
+    "®": 112,
+    "°": 113,
+    "±": 114,
+    "²": 115,
+    "³": 116,
+    "´": 117,
+    "¶": 118,
+    "·": 119,
+    "¹": 120,
+    "º": 121,
+    "»": 122,
+    "¼": 123,
+    "½": 124,
+    "¾": 125,
+    "¿": 126,
+    "À": 127,
+    "Á": 128,
+    "Â": 129,
+    "Ä": 130,
+    "Å": 131,
+    "Æ": 132,
+    "Ç": 133,
+    "È": 134,
+    "É": 135,
+    "Ê": 136,
+    "Í": 137,
+    "Ï": 138,
+    "Ñ": 139,
+    "Ó": 140,
+    "Ô": 141,
+    "Ö": 142,
+    "×": 143,
+    "Ù": 144,
+    "Ú": 145,
+    "Ü": 146,
+    "Ý": 147,
+    "Þ": 148,
+    "ß": 149,
+    "à": 150,
+    "á": 151,
+    "â": 152,
+    "ã": 153,
+    "ä": 154,
+    "å": 155,
+    "æ": 156,
+    "ç": 157,
+    "è": 158,
+    "é": 159,
+    "ê": 160,
+    "ë": 161,
+    "ì": 162,
+    "í": 163,
+    "î": 164,
+    "ï": 165,
+    "ð": 166,
+    "ñ": 167,
+    "ò": 168,
+    "ó": 169,
+    "ô": 170,
+    "õ": 171,
+    "ö": 172,
+    "ø": 173,
+    "ù": 174,
+    "ú": 175,
+    "û": 176,
+    "ü": 177,
+    "ý": 178,
+    "þ": 179,
+    "ÿ": 180,
+    "ā": 181,
+    "Ă": 182,
+    "ă": 183,
+    "ą": 184,
+    "ć": 185,
+    "č": 186,
+    "Đ": 187,
+    "đ": 188,
+    "ē": 189,
+    "ė": 190,
+    "ę": 191,
+    "ğ": 192,
+    "ġ": 193,
+    "ĩ": 194,
+    "ī": 195,
+    "ļ": 196,
+    "ł": 197,
+    "ń": 198,
+    "Ō": 199,
+    "ō": 200,
+    "Œ": 201,
+    "œ": 202,
+    "ŕ": 203,
+    "ś": 204,
+    "ş": 205,
+    "š": 206,
+    "ŧ": 207,
+    "ũ": 208,
+    "ū": 209,
+    "ŭ": 210,
+    "ů": 211,
+    "ű": 212,
+    "ż": 213,
+    "Ž": 214,
+    "ž": 215,
+    "ſ": 216,
+    "Ơ": 217,
+    "ơ": 218,
+    "Ư": 219,
+    "ư": 220,
+    "ǁ": 221,
+    "ǂ": 222,
+    "ǐ": 223,
+    "ǘ": 224,
+    "ǧ": 225,
+    "ș": 226,
+    "Ȝ": 227,
+    "ȝ": 228,
+    "ȟ": 229,
+    "ɛ": 230,
+    "ʰ": 231,
+    "ˆ": 232,
+    "ˈ": 233,
+    "ˢ": 234,
+    "̃": 235,
+    "̍": 236,
+    "̣": 237,
+    "̱": 238,
+    "͈": 239,
+    "α": 240,
+    "β": 241,
+    "γ": 242,
+    "δ": 243,
+    "ε": 244,
+    "η": 245,
+    "κ": 246,
+    "μ": 247,
+    "ξ": 248,
+    "π": 249,
+    "ς": 250,
+    "σ": 251,
+    "χ": 252,
+    "ψ": 253,
+    "а": 254,
+    "д": 255,
+    "е": 256,
+    "з": 257,
+    "и": 258,
+    "к": 259,
+    "л": 260,
+    "о": 261,
+    "п": 262,
+    "р": 263,
+    "с": 264,
+    "т": 265,
+    "ч": 266,
+    "ь": 267,
+    "ง": 268,
+    "ย": 269,
+    "า": 270,
+    "่": 271,
+    "ည": 272,
+    "လ": 273,
+    "း": 274,
+    "်": 275,
+    "ᵈ": 276,
+    "ᵗ": 277,
+    "ḅ": 278,
+    "ḍ": 279,
+    "ḳ": 280,
+    "ṅ": 281,
+    "Ṭ": 282,
+    "ẓ": 283,
+    "ạ": 284,
+    "Ả": 285,
+    "ả": 286,
+    "Ấ": 287,
+    "ấ": 288,
+    "ầ": 289,
+    "ẩ": 290,
+    "ẫ": 291,
+    "ậ": 292,
+    "ắ": 293,
+    "ằ": 294,
+    "ẳ": 295,
+    "ẵ": 296,
+    "ặ": 297,
+    "ẹ": 298,
+    "ẻ": 299,
+    "ẽ": 300,
+    "Ế": 301,
+    "ế": 302,
+    "Ề": 303,
+    "ề": 304,
+    "ể": 305,
+    "ễ": 306,
+    "ệ": 307,
+    "ỉ": 308,
+    "ị": 309,
+    "ọ": 310,
+    "ỏ": 311,
+    "Ố": 312,
+    "ố": 313,
+    "Ồ": 314,
+    "ồ": 315,
+    "Ổ": 316,
+    "ổ": 317,
+    "ỗ": 318,
+    "Ộ": 319,
+    "ộ": 320,
+    "ớ": 321,
+    "Ờ": 322,
+    "ờ": 323,
+    "Ở": 324,
+    "ở": 325,
+    "ỡ": 326,
+    "ợ": 327,
+    "ụ": 328,
+    "ủ": 329,
+    "Ứ": 330,
+    "ứ": 331,
+    "Ừ": 332,
+    "ừ": 333,
+    "ử": 334,
+    "ữ": 335,
+    "ự": 336,
+    "ỳ": 337,
+    "ỵ": 338,
+    "Ỷ": 339,
+    "ỷ": 340,
+    "ỹ": 341,
+    "ὖ": 342,
+    "ὼ": 343,
+    "ῆ": 344,
+    "​": 345,
+    "‐": 346,
+    "—": 347,
+    "―": 348,
+    "†": 349,
+    "•": 350,
+    "…": 351,
+    "‰": 352,
+    "′": 353,
+    "″": 354,
+    "⁄": 355,
+    "⁷": 356,
+    "⁻": 357,
+    "₀": 358,
+    "₁": 359,
+    "₂": 360,
+    "₃": 361,
+    "₄": 362,
+    "₅": 363,
+    "₆": 364,
+    "₇": 365,
+    "₈": 366,
+    "€": 367,
+    "₹": 368,
+    "⅓": 369,
+    "⅔": 370,
+    "⅖": 371,
+    "⅙": 372,
+    "⅛": 373,
+    "⅝": 374,
+    "√": 375,
+    "│": 376,
+    "♦": 377,
+    "⚓": 378,
+    "✈": 379,
+    "✠": 380,
+    "✨": 381,
+    "⟙": 382,
+    "⸺": 383,
+    "。": 384,
+    "〈": 385,
+    "な": 386,
+    "三": 387,
+    "次": 388,
+    "肿": 389,
+    "臃": 390,
+    "구": 391,
+    "로": 392,
+    "요": 393,
+    "ﷺ": 394,
+    "️": 395,
+    "𝑥": 396,
+    "𝑦": 397,
+    "🌊": 398,
+    "🌿": 399,
+    "💦": 400,
+    "💨": 401
+  },
+  "phoneme_symbols": {
+    "0": "_",
+    "1": "<de>",
+    "2": "<en_gb>",
+    "3": "<en_us>",
+    "4": "<es>",
+    "5": "<fr>",
+    "6": "<it>",
+    "7": "<vi>",
+    "8": "<end>",
+    "9": " ",
+    "10": "!",
+    "11": ",",
+    "12": "-",
+    "13": ".",
+    "14": "1",
+    "15": "2",
+    "16": "4",
+    "17": "5",
+    "18": "6",
+    "19": "7",
+    "20": ":",
+    "21": ";",
+    "22": "?",
+    "23": "a",
+    "24": "b",
+    "25": "c",
+    "26": "d",
+    "27": "e",
+    "28": "f",
+    "29": "h",
+    "30": "i",
+    "31": "j",
+    "32": "k",
+    "33": "l",
+    "34": "m",
+    "35": "n",
+    "36": "o",
+    "37": "p",
+    "38": "q",
+    "39": "r",
+    "40": "s",
+    "41": "t",
+    "42": "u",
+    "43": "v",
+    "44": "w",
+    "45": "x",
+    "46": "y",
+    "47": "z",
+    "48": "«",
+    "49": "»",
+    "50": "æ",
+    "51": "ç",
+    "52": "ð",
+    "53": "ø",
+    "54": "ŋ",
+    "55": "œ",
+    "56": "ɐ",
+    "57": "ɑ",
+    "58": "ɒ",
+    "59": "ɔ",
+    "60": "ɗ",
+    "61": "ə",
+    "62": "ɚ",
+    "63": "ɛ",
+    "64": "ɜ",
+    "65": "ɟ",
+    "66": "ɡ",
+    "67": "ɣ",
+    "68": "ɨ",
+    "69": "ɪ",
+    "70": "ɬ",
+    "71": "ɯ",
+    "72": "ɲ",
+    "73": "ɹ",
+    "74": "ɾ",
+    "75": "ʁ",
+    "76": "ʃ",
+    "77": "ʊ",
+    "78": "ʌ",
+    "79": "ʎ",
+    "80": "ʐ",
+    "81": "ʒ",
+    "82": "ʔ",
+    "83": "ʝ",
+    "84": "ʲ",
+    "85": "ˈ",
+    "86": "ˌ",
+    "87": "ː",
+    "88": "̃",
+    "89": "̩",
+    "90": "̪",
+    "91": "β",
+    "92": "θ",
+    "93": "ᵻ",
+    "94": "—",
+    "95": "…"
+  }
+}

coreml/assets/languages.json

@@ -0,0 +1,30 @@
+[
+  {
+    "id": "en_us",
+    "index": 0
+  },
+  {
+    "id": "en_gb",
+    "index": 1
+  },
+  {
+    "id": "es",
+    "index": 2
+  },
+  {
+    "id": "it",
+    "index": 3
+  },
+  {
+    "id": "fr",
+    "index": 4
+  },
+  {
+    "id": "de",
+    "index": 5
+  },
+  {
+    "id": "vi",
+    "index": 6
+  }
+]

coreml/assets/phone_vocab.json

@@ -0,0 +1,111 @@
+{
+  "schema_version": 1,
+  "mode": "scyllasband_ipa_phone_symbols",
+  "special_tokens": [
+    "<pad>",
+    "<unk>",
+    "<sil>",
+    "<pause_comma>",
+    "<pause_semicolon>",
+    "<pause_colon>",
+    "<pause_dash>",
+    "<ellipsis>",
+    "<end_stmt>",
+    "<end_question>",
+    "<end_exclaim>",
+    "<ctx_sentence_start>",
+    "<ctx_continuation>",
+    "<ctx_sentence_end>",
+    "<ctx_chunk_continue>"
+  ],
+  "vocab_size": 87,
+  "token_to_id": {
+    "<pad>": 0,
+    "<unk>": 1,
+    "<sil>": 2,
+    "<pause_comma>": 3,
+    "<pause_semicolon>": 4,
+    "<pause_colon>": 5,
+    "<pause_dash>": 6,
+    "<ellipsis>": 7,
+    "<end_stmt>": 8,
+    "<end_question>": 9,
+    "<end_exclaim>": 10,
+    "<ctx_sentence_start>": 11,
+    "<ctx_continuation>": 12,
+    "<ctx_sentence_end>": 13,
+    "<ctx_chunk_continue>": 14,
+    "1": 15,
+    "2": 16,
+    "4": 17,
+    "5": 18,
+    "6": 19,
+    "7": 20,
+    "a": 21,
+    "b": 22,
+    "c": 23,
+    "d": 24,
+    "e": 25,
+    "f": 26,
+    "h": 27,
+    "i": 28,
+    "j": 29,
+    "k": 30,
+    "l": 31,
+    "m": 32,
+    "n": 33,
+    "o": 34,
+    "p": 35,
+    "r": 36,
+    "s": 37,
+    "t": 38,
+    "u": 39,
+    "v": 40,
+    "w": 41,
+    "x": 42,
+    "y": 43,
+    "z": 44,
+    "æ": 45,
+    "ç": 46,
+    "ð": 47,
+    "ø": 48,
+    "ŋ": 49,
+    "œ": 50,
+    "ɐ": 51,
+    "ɑ": 52,
+    "ɒ": 53,
+    "ɔ": 54,
+    "ɗ": 55,
+    "ə": 56,
+    "ɚ": 57,
+    "ɛ": 58,
+    "ɜ": 59,
+    "ɟ": 60,
+    "ɡ": 61,
+    "ɣ": 62,
+    "ɪ": 63,
+    "ɯ": 64,
+    "ɲ": 65,
+    "ɹ": 66,
+    "ɾ": 67,
+    "ʁ": 68,
+    "ʃ": 69,
+    "ʊ": 70,
+    "ʌ": 71,
+    "ʎ": 72,
+    "ʐ": 73,
+    "ʒ": 74,
+    "ʔ": 75,
+    "ʝ": 76,
+    "ʲ": 77,
+    "ˈ": 78,
+    "ˌ": 79,
+    "ː": 80,
+    "̃": 81,
+    "̩": 82,
+    "̪": 83,
+    "β": 84,
+    "θ": 85,
+    "ᵻ": 86
+  }
+}

coreml/assets/voices.json

@@ -0,0 +1,42 @@
+[
+  {
+    "id": "ariadne",
+    "index": 0
+  },
+  {
+    "id": "felix",
+    "index": 1
+  },
+  {
+    "id": "gwen",
+    "index": 2
+  },
+  {
+    "id": "ink",
+    "index": 3
+  },
+  {
+    "id": "max",
+    "index": 4
+  },
+  {
+    "id": "orpheus",
+    "index": 5
+  },
+  {
+    "id": "rex",
+    "index": 6
+  },
+  {
+    "id": "scylla",
+    "index": 7
+  },
+  {
+    "id": "stone",
+    "index": 8
+  },
+  {
+    "id": "tuesday",
+    "index": 9
+  }
+]

coreml/coreml/duration_predictor.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:953e274e93c3abfae6f04977436816dc28223ef61d4e91e4191f67837740a2e8
+size 243

coreml/coreml/duration_predictor.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c090d28d8df68312410867a49963b5fbda00f5883178969df17b1b0c86e7eff2
+size 554

coreml/coreml/duration_predictor.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5383e19195d6807cf6632d2c82b51b86aaa947ba51d8042434f73b5039c70ea4
+size 10562880

coreml/coreml/g2p.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:61936fdfe11637e628b2134025bea29d673c98e93ce6f01c05e7b5da124563b8
+size 243

coreml/coreml/g2p.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f5fc8bee0186f33fa14d35a59a58ec3393676d07bdf03802a3836d806662cd03
+size 377

coreml/coreml/g2p.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:045da2991ed2a42ba027aaa6b83813054f410b6eedfdce82f2ab287ec4ca74e2
+size 27346752

coreml/coreml/vector_context_encoder.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f89405fefad88549b521a819805c5c1781e0c930c03b8f670b1e1539aafe88c6
+size 243

coreml/coreml/vector_context_encoder.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d031d4ef221bc846aec129114aec7366fcb6126dacae44c24c4b344c54dc6c5a
+size 476

coreml/coreml/vector_context_encoder.mlmodelc/model.mil

@@ -0,0 +1,267 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3600.16.1"}, {"coremlc-version", "3600.25.2"}})]
+{
+    func main<ios18>(tensor<int32, [1, 512]> span_context_mask, tensor<int32, [1, 512]> span_context_phone_ids, tensor<int32, [1, 512]> span_context_segment_ids) {
+            int32 var_15 = const()[name = string("op_15"), val = int32(0)];
+            tensor<bool, [1, 512]> span_context_mask_1 = not_equal(x = span_context_mask, y = var_15)[name = string("span_context_mask")];
+            int32 var_30 = const()[name = string("op_30"), val = int32(-1)];
+            int32 var_35 = const()[name = string("op_35"), val = int32(2)];
+            int32 var_36 = const()[name = string("op_36"), val = int32(86)];
+            int32 var_38 = const()[name = string("op_38"), val = int32(1)];
+            int32 var_40 = const()[name = string("op_40"), val = int32(0)];
+            tensor<int32, [1, 512]> var_55 = maximum(x = span_context_phone_ids, y = var_40)[name = string("op_55")];
+            tensor<int32, [1, 512]> input_1 = minimum(x = var_55, y = var_36)[name = string("input_1")];
+            tensor<int32, [1, 512]> var_59 = maximum(x = span_context_segment_ids, y = var_40)[name = string("op_59")];
+            tensor<int32, [1, 512]> input_3 = minimum(x = var_59, y = var_35)[name = string("input_3")];
+            string var_64_dtype_0 = const()[name = string("op_64_dtype_0"), val = string("int32")];
+            bool var_65_exclusive_0 = const()[name = string("op_65_exclusive_0"), val = bool(false)];
+            bool var_65_reverse_0 = const()[name = string("op_65_reverse_0"), val = bool(false)];
+            tensor<int32, [1, 512]> var_64 = cast(dtype = var_64_dtype_0, x = span_context_mask_1)[name = string("cast_13")];
+            tensor<int32, [1, 512]> var_65 = cumsum(axis = var_38, exclusive = var_65_exclusive_0, reverse = var_65_reverse_0, x = var_64)[name = string("op_65")];
+            int32 var_66 = const()[name = string("op_66"), val = int32(1)];
+            tensor<int32, [1, 512]> var_67 = sub(x = var_65, y = var_66)[name = string("op_67")];
+            tensor<int32, [1, 512]> input_5 = maximum(x = var_67, y = var_40)[name = string("input_5")];
+            int32 var_70_batch_dims_0 = const()[name = string("op_70_batch_dims_0"), val = int32(0)];
+            bool var_70_validate_indices_0 = const()[name = string("op_70_validate_indices_0"), val = bool(false)];
+            tensor<fp16, [87, 512]> module_flow_phone_embed_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [87, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), scale = tensor<fp16, [87, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44672))))[name = string("module_flow_phone_embed_weight_to_fp16_quantized")];
+            string input_1_to_int16_dtype_0 = const()[name = string("input_1_to_int16_dtype_0"), val = string("int16")];
+            string cast_24_dtype_0 = const()[name = string("cast_24_dtype_0"), val = string("int32")];
+            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
+            tensor<int16, [1, 512]> input_1_to_int16 = cast(dtype = input_1_to_int16_dtype_0, x = input_1)[name = string("cast_12")];
+            tensor<int32, [1, 512]> cast_24 = cast(dtype = cast_24_dtype_0, x = input_1_to_int16)[name = string("cast_11")];
+            tensor<bool, [1, 512]> greater_equal_0 = greater_equal(x = cast_24, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
+            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(87)];
+            tensor<int32, [1, 512]> add_0 = add(x = cast_24, y = slice_by_index_0)[name = string("add_0")];
+            tensor<int32, [1, 512]> select_0 = select(a = cast_24, b = add_0, cond = greater_equal_0)[name = string("select_0")];
+            string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")];
+            string cast_0_dtype_0 = const()[name = string("cast_0_dtype_0"), val = string("int32")];
+            int32 greater_equal_0_y_0_1 = const()[name = string("greater_equal_0_y_0_1"), val = int32(0)];
+            tensor<int16, [1, 512]> select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_10")];
+            tensor<int32, [1, 512]> cast_0 = cast(dtype = cast_0_dtype_0, x = select_0_to_int16)[name = string("cast_9")];
+            tensor<bool, [1, 512]> greater_equal_0_1 = greater_equal(x = cast_0, y = greater_equal_0_y_0_1)[name = string("greater_equal_0_1")];
+            int32 slice_by_index_0_1 = const()[name = string("slice_by_index_0_1"), val = int32(87)];
+            tensor<int32, [1, 512]> add_0_1 = add(x = cast_0, y = slice_by_index_0_1)[name = string("add_0_1")];
+            tensor<int32, [1, 512]> select_0_1 = select(a = cast_0, b = add_0_1, cond = greater_equal_0_1)[name = string("select_0_1")];
+            int32 op_70_cast_fp16_cast_uint16_cast_uint16_axis_0 = const()[name = string("op_70_cast_fp16_cast_uint16_cast_uint16_axis_0"), val = int32(0)];
+            tensor<fp16, [1, 512, 512]> op_70_cast_fp16_cast_uint16_cast_uint16 = gather(axis = op_70_cast_fp16_cast_uint16_cast_uint16_axis_0, batch_dims = var_70_batch_dims_0, indices = select_0_1, validate_indices = var_70_validate_indices_0, x = module_flow_phone_embed_weight_to_fp16_quantized)[name = string("op_70_cast_fp16_cast_uint16_cast_uint16")];
+            int32 var_72_axis_0 = const()[name = string("op_72_axis_0"), val = int32(0)];
+            int32 var_72_batch_dims_0 = const()[name = string("op_72_batch_dims_0"), val = int32(0)];
+            bool var_72_validate_indices_0 = const()[name = string("op_72_validate_indices_0"), val = bool(false)];
+            tensor<fp16, [3, 512]> module_flow_span_segment_embed_weight_to_fp16 = const()[name = string("module_flow_span_segment_embed_weight_to_fp16"), val = tensor<fp16, [3, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44928)))];
+            string input_3_to_uint16_dtype_0 = const()[name = string("input_3_to_uint16_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 512]> input_3_to_uint16 = cast(dtype = input_3_to_uint16_dtype_0, x = input_3)[name = string("cast_8")];
+            tensor<fp16, [1, 512, 512]> var_72_cast_fp16_cast_uint16 = gather(axis = var_72_axis_0, batch_dims = var_72_batch_dims_0, indices = input_3_to_uint16, validate_indices = var_72_validate_indices_0, x = module_flow_span_segment_embed_weight_to_fp16)[name = string("op_72_cast_fp16_cast_uint16")];
+            tensor<fp16, [1, 512, 512]> var_73_cast_fp16 = add(x = op_70_cast_fp16_cast_uint16_cast_uint16, y = var_72_cast_fp16_cast_uint16)[name = string("op_73_cast_fp16")];
+            int32 var_75_axis_0 = const()[name = string("op_75_axis_0"), val = int32(0)];
+            int32 var_75_batch_dims_0 = const()[name = string("op_75_batch_dims_0"), val = int32(0)];
+            bool var_75_validate_indices_0 = const()[name = string("op_75_validate_indices_0"), val = bool(false)];
+            tensor<fp16, [768, 512]> module_flow_span_position_embed_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48064))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441344))))[name = string("module_flow_span_position_embed_weight_to_fp16_quantized")];
+            string input_5_to_uint16_dtype_0 = const()[name = string("input_5_to_uint16_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 512]> input_5_to_uint16 = cast(dtype = input_5_to_uint16_dtype_0, x = input_5)[name = string("cast_7")];
+            tensor<fp16, [1, 512, 512]> var_75_cast_fp16_cast_uint16 = gather(axis = var_75_axis_0, batch_dims = var_75_batch_dims_0, indices = input_5_to_uint16, validate_indices = var_75_validate_indices_0, x = module_flow_span_position_embed_weight_to_fp16_quantized)[name = string("op_75_cast_fp16_cast_uint16")];
+            tensor<fp16, [1, 512, 512]> hidden_1_cast_fp16 = add(x = var_73_cast_fp16, y = var_75_cast_fp16_cast_uint16)[name = string("hidden_1_cast_fp16")];
+            tensor<int32, [1]> var_82_axes_0 = const()[name = string("op_82_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, [1, 512, 1]> var_82 = expand_dims(axes = var_82_axes_0, x = span_context_mask_1)[name = string("op_82")];
+            tensor<bool, [1, 512, 1]> var_83 = logical_not(x = var_82)[name = string("op_83")];
+            fp16 var_18_to_fp16 = const()[name = string("op_18_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 512, 512]> hidden_3_cast_fp16 = select(a = var_18_to_fp16, b = hidden_1_cast_fp16, cond = var_83)[name = string("hidden_3_cast_fp16")];
+            tensor<int32, [1]> input_7_axes_0 = const()[name = string("input_7_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> module_flow_span_ordered_encoder_blocks_0_norm_attention_weight_to_fp16 = const()[name = string("module_flow_span_ordered_encoder_blocks_0_norm_attention_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442944)))];
+            tensor<fp16, [512]> module_flow_span_ordered_encoder_blocks_0_norm_attention_bias_to_fp16 = const()[name = string("module_flow_span_ordered_encoder_blocks_0_norm_attention_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444032)))];
+            fp16 var_28_to_fp16 = const()[name = string("op_28_to_fp16"), val = fp16(0x1.5p-17)];
+            tensor<fp16, [1, 512, 512]> input_7_cast_fp16 = layer_norm(axes = input_7_axes_0, beta = module_flow_span_ordered_encoder_blocks_0_norm_attention_bias_to_fp16, epsilon = var_28_to_fp16, gamma = module_flow_span_ordered_encoder_blocks_0_norm_attention_weight_to_fp16, x = hidden_3_cast_fp16)[name = string("input_7_cast_fp16")];
+            tensor<fp16, [1536, 512]> module_flow_span_ordered_encoder_blocks_0_qkv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445120))), scale = tensor<fp16, [1536, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1231616))))[name = string("module_flow_span_ordered_encoder_blocks_0_qkv_weight_to_fp16_quantized")];
+            tensor<fp16, [1536]> module_flow_span_ordered_encoder_blocks_0_qkv_bias_to_fp16 = const()[name = string("module_flow_span_ordered_encoder_blocks_0_qkv_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1234752)))];
+            tensor<fp16, [1, 512, 1536]> linear_0_cast_fp16 = linear(bias = module_flow_span_ordered_encoder_blocks_0_qkv_bias_to_fp16, weight = module_flow_span_ordered_encoder_blocks_0_qkv_weight_to_fp16_quantized, x = input_7_cast_fp16)[name = string("linear_0_cast_fp16")];
+            tensor<int32, [5]> var_100 = const()[name = string("op_100"), val = tensor<int32, [5]>([1, 512, 3, 8, 64])];
+            tensor<fp16, [1, 512, 3, 8, 64]> qkv_1_cast_fp16 = reshape(shape = var_100, x = linear_0_cast_fp16)[name = string("qkv_1_cast_fp16")];
+            tensor<int32, [3]> var_102_split_sizes_0 = const()[name = string("op_102_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
+            int32 var_102_axis_0 = const()[name = string("op_102_axis_0"), val = int32(2)];
+            tensor<fp16, [1, 512, 1, 8, 64]> var_102_cast_fp16_0, tensor<fp16, [1, 512, 1, 8, 64]> var_102_cast_fp16_1, tensor<fp16, [1, 512, 1, 8, 64]> var_102_cast_fp16_2 = split(axis = var_102_axis_0, split_sizes = var_102_split_sizes_0, x = qkv_1_cast_fp16)[name = string("op_102_cast_fp16")];
+            tensor<int32, [1]> squeeze_0_axes_0 = const()[name = string("squeeze_0_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 512, 8, 64]> squeeze_0_cast_fp16 = squeeze(axes = squeeze_0_axes_0, x = var_102_cast_fp16_0)[name = string("squeeze_0_cast_fp16")];
+            tensor<int32, [1]> squeeze_1_axes_0 = const()[name = string("squeeze_1_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 512, 8, 64]> squeeze_1_cast_fp16 = squeeze(axes = squeeze_1_axes_0, x = var_102_cast_fp16_1)[name = string("squeeze_1_cast_fp16")];
+            tensor<int32, [1]> squeeze_2_axes_0 = const()[name = string("squeeze_2_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 512, 8, 64]> squeeze_2_cast_fp16 = squeeze(axes = squeeze_2_axes_0, x = var_102_cast_fp16_2)[name = string("squeeze_2_cast_fp16")];
+            tensor<int32, [4]> value_3_perm_0 = const()[name = string("value_3_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool var_110_transpose_x_0 = const()[name = string("op_110_transpose_x_0"), val = bool(false)];
+            bool var_110_transpose_y_0 = const()[name = string("op_110_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_8_perm_0 = const()[name = string("transpose_8_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_9_perm_0 = const()[name = string("transpose_9_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 512]> transpose_9 = transpose(perm = transpose_9_perm_0, x = squeeze_1_cast_fp16)[name = string("transpose_18")];
+            tensor<fp16, [1, 8, 512, 64]> transpose_8 = transpose(perm = transpose_8_perm_0, x = squeeze_0_cast_fp16)[name = string("transpose_19")];
+            tensor<fp16, [1, 8, 512, 512]> var_110_cast_fp16 = matmul(transpose_x = var_110_transpose_x_0, transpose_y = var_110_transpose_y_0, x = transpose_8, y = transpose_9)[name = string("op_110_cast_fp16")];
+            fp16 var_111_to_fp16 = const()[name = string("op_111_to_fp16"), val = fp16(0x1p-3)];
+            tensor<fp16, [1, 8, 512, 512]> scores_1_cast_fp16 = mul(x = var_110_cast_fp16, y = var_111_to_fp16)[name = string("scores_1_cast_fp16")];
+            tensor<int32, [1]> var_114_axes_0 = const()[name = string("op_114_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, 512]> var_114 = expand_dims(axes = var_114_axes_0, x = span_context_mask_1)[name = string("op_114")];
+            tensor<int32, [1]> var_115_axes_0 = const()[name = string("op_115_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<bool, [1, 1, 1, 512]> var_115 = expand_dims(axes = var_115_axes_0, x = var_114)[name = string("op_115")];
+            tensor<bool, [1, 1, 1, 512]> var_117 = logical_not(x = var_115)[name = string("op_117")];
+            fp16 var_23_to_fp16 = const()[name = string("op_23_to_fp16"), val = fp16(-inf)];
+            tensor<fp16, [1, 8, 512, 512]> scores_3_cast_fp16 = select(a = var_23_to_fp16, b = scores_1_cast_fp16, cond = var_117)[name = string("scores_3_cast_fp16")];
+            tensor<fp16, [1, 8, 512, 512]> var_119_cast_fp16 = softmax(axis = var_30, x = scores_3_cast_fp16)[name = string("op_119_cast_fp16")];
+            string var_120_to_fp16_dtype_0 = const()[name = string("op_120_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 1, 1, 512]> var_115_to_fp16 = cast(dtype = var_120_to_fp16_dtype_0, x = var_115)[name = string("cast_6")];
+            tensor<fp16, [1, 8, 512, 512]> weights_1_cast_fp16 = mul(x = var_119_cast_fp16, y = var_115_to_fp16)[name = string("weights_1_cast_fp16")];
+            tensor<int32, [1]> var_123_axes_0 = const()[name = string("op_123_axes_0"), val = tensor<int32, [1]>([-1])];
+            bool var_123_keep_dims_0 = const()[name = string("op_123_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 8, 512, 1]> var_123_cast_fp16 = reduce_sum(axes = var_123_axes_0, keep_dims = var_123_keep_dims_0, x = weights_1_cast_fp16)[name = string("op_123_cast_fp16")];
+            fp16 var_21_to_fp16 = const()[name = string("op_21_to_fp16"), val = fp16(0x1.1p-20)];
+            tensor<fp16, [1, 8, 512, 1]> var_124_cast_fp16 = maximum(x = var_123_cast_fp16, y = var_21_to_fp16)[name = string("op_124_cast_fp16")];
+            tensor<fp16, [1, 8, 512, 512]> weights_3_cast_fp16 = real_div(x = weights_1_cast_fp16, y = var_124_cast_fp16)[name = string("weights_3_cast_fp16")];
+            bool var_126_transpose_x_0 = const()[name = string("op_126_transpose_x_0"), val = bool(false)];
+            bool var_126_transpose_y_0 = const()[name = string("op_126_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 512, 64]> value_3_cast_fp16 = transpose(perm = value_3_perm_0, x = squeeze_2_cast_fp16)[name = string("transpose_17")];
+            tensor<fp16, [1, 8, 512, 64]> var_126_cast_fp16 = matmul(transpose_x = var_126_transpose_x_0, transpose_y = var_126_transpose_y_0, x = weights_3_cast_fp16, y = value_3_cast_fp16)[name = string("op_126_cast_fp16")];
+            tensor<int32, [4]> var_127_perm_0 = const()[name = string("op_127_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_128 = const()[name = string("op_128"), val = tensor<int32, [3]>([1, 512, 512])];
+            tensor<fp16, [1, 512, 8, 64]> var_127_cast_fp16 = transpose(perm = var_127_perm_0, x = var_126_cast_fp16)[name = string("transpose_16")];
+            tensor<fp16, [1, 512, 512]> input_9_cast_fp16 = reshape(shape = var_128, x = var_127_cast_fp16)[name = string("input_9_cast_fp16")];
+            tensor<fp16, [512, 512]> module_flow_span_ordered_encoder_blocks_0_attention_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1237888))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1500096))))[name = string("module_flow_span_ordered_encoder_blocks_0_attention_proj_weight_to_fp16_quantized")];
+            tensor<fp16, [512]> module_flow_span_ordered_encoder_blocks_0_attention_proj_bias_to_fp16 = const()[name = string("module_flow_span_ordered_encoder_blocks_0_attention_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1501184)))];
+            tensor<fp16, [1, 512, 512]> linear_1_cast_fp16 = linear(bias = module_flow_span_ordered_encoder_blocks_0_attention_proj_bias_to_fp16, weight = module_flow_span_ordered_encoder_blocks_0_attention_proj_weight_to_fp16_quantized, x = input_9_cast_fp16)[name = string("linear_1_cast_fp16")];
+            tensor<fp16, [1, 512, 512]> input_13_cast_fp16 = add(x = hidden_3_cast_fp16, y = linear_1_cast_fp16)[name = string("input_13_cast_fp16")];
+            tensor<int32, [1]> input_15_axes_0 = const()[name = string("input_15_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> module_flow_span_ordered_encoder_blocks_0_norm_feedforward_weight_to_fp16 = const()[name = string("module_flow_span_ordered_encoder_blocks_0_norm_feedforward_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1502272)))];
+            tensor<fp16, [512]> module_flow_span_ordered_encoder_blocks_0_norm_feedforward_bias_to_fp16 = const()[name = string("module_flow_span_ordered_encoder_blocks_0_norm_feedforward_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1503360)))];
+            tensor<fp16, [1, 512, 512]> input_15_cast_fp16 = layer_norm(axes = input_15_axes_0, beta = module_flow_span_ordered_encoder_blocks_0_norm_feedforward_bias_to_fp16, epsilon = var_28_to_fp16, gamma = module_flow_span_ordered_encoder_blocks_0_norm_feedforward_weight_to_fp16, x = input_13_cast_fp16)[name = string("input_15_cast_fp16")];
+            tensor<fp16, [2048, 512]> module_flow_span_ordered_encoder_blocks_0_feedforward_0_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1504448))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2553088))))[name = string("module_flow_span_ordered_encoder_blocks_0_feedforward_0_weight_to_fp16_quantized")];
+            tensor<fp16, [2048]> module_flow_span_ordered_encoder_blocks_0_feedforward_0_bias_to_fp16 = const()[name = string("module_flow_span_ordered_encoder_blocks_0_feedforward_0_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2557248)))];
+            tensor<fp16, [1, 512, 2048]> linear_2_cast_fp16 = linear(bias = module_flow_span_ordered_encoder_blocks_0_feedforward_0_bias_to_fp16, weight = module_flow_span_ordered_encoder_blocks_0_feedforward_0_weight_to_fp16_quantized, x = input_15_cast_fp16)[name = string("linear_2_cast_fp16")];
+            string input_19_mode_0 = const()[name = string("input_19_mode_0"), val = string("TANH_APPROXIMATION")];
+            tensor<fp16, [1, 512, 2048]> input_19_cast_fp16 = gelu(mode = input_19_mode_0, x = linear_2_cast_fp16)[name = string("input_19_cast_fp16")];
+            tensor<fp16, [512, 2048]> module_flow_span_ordered_encoder_blocks_0_feedforward_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2561408))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3610048))))[name = string("module_flow_span_ordered_encoder_blocks_0_feedforward_2_weight_to_fp16_quantized")];
+            tensor<fp16, [512]> module_flow_span_ordered_encoder_blocks_0_feedforward_2_bias_to_fp16 = const()[name = string("module_flow_span_ordered_encoder_blocks_0_feedforward_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3611136)))];
+            tensor<fp16, [1, 512, 512]> linear_3_cast_fp16 = linear(bias = module_flow_span_ordered_encoder_blocks_0_feedforward_2_bias_to_fp16, weight = module_flow_span_ordered_encoder_blocks_0_feedforward_2_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_3_cast_fp16")];
+            tensor<fp16, [1, 512, 512]> hidden_5_cast_fp16 = add(x = input_13_cast_fp16, y = linear_3_cast_fp16)[name = string("hidden_5_cast_fp16")];
+            tensor<fp16, [1, 512, 512]> hidden_7_cast_fp16 = select(a = var_18_to_fp16, b = hidden_5_cast_fp16, cond = var_83)[name = string("hidden_7_cast_fp16")];
+            tensor<int32, [1]> input_23_axes_0 = const()[name = string("input_23_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> module_flow_span_ordered_encoder_blocks_1_norm_attention_weight_to_fp16 = const()[name = string("module_flow_span_ordered_encoder_blocks_1_norm_attention_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3612224)))];
+            tensor<fp16, [512]> module_flow_span_ordered_encoder_blocks_1_norm_attention_bias_to_fp16 = const()[name = string("module_flow_span_ordered_encoder_blocks_1_norm_attention_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3613312)))];
+            tensor<fp16, [1, 512, 512]> input_23_cast_fp16 = layer_norm(axes = input_23_axes_0, beta = module_flow_span_ordered_encoder_blocks_1_norm_attention_bias_to_fp16, epsilon = var_28_to_fp16, gamma = module_flow_span_ordered_encoder_blocks_1_norm_attention_weight_to_fp16, x = hidden_7_cast_fp16)[name = string("input_23_cast_fp16")];
+            tensor<fp16, [1536, 512]> module_flow_span_ordered_encoder_blocks_1_qkv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1536, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3614400))), scale = tensor<fp16, [1536, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4400896))))[name = string("module_flow_span_ordered_encoder_blocks_1_qkv_weight_to_fp16_quantized")];
+            tensor<fp16, [1536]> module_flow_span_ordered_encoder_blocks_1_qkv_bias_to_fp16 = const()[name = string("module_flow_span_ordered_encoder_blocks_1_qkv_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4404032)))];
+            tensor<fp16, [1, 512, 1536]> linear_4_cast_fp16 = linear(bias = module_flow_span_ordered_encoder_blocks_1_qkv_bias_to_fp16, weight = module_flow_span_ordered_encoder_blocks_1_qkv_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_4_cast_fp16")];
+            tensor<int32, [5]> var_168 = const()[name = string("op_168"), val = tensor<int32, [5]>([1, 512, 3, 8, 64])];
+            tensor<fp16, [1, 512, 3, 8, 64]> qkv_3_cast_fp16 = reshape(shape = var_168, x = linear_4_cast_fp16)[name = string("qkv_3_cast_fp16")];
+            tensor<int32, [3]> var_170_split_sizes_0 = const()[name = string("op_170_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
+            int32 var_170_axis_0 = const()[name = string("op_170_axis_0"), val = int32(2)];
+            tensor<fp16, [1, 512, 1, 8, 64]> var_170_cast_fp16_0, tensor<fp16, [1, 512, 1, 8, 64]> var_170_cast_fp16_1, tensor<fp16, [1, 512, 1, 8, 64]> var_170_cast_fp16_2 = split(axis = var_170_axis_0, split_sizes = var_170_split_sizes_0, x = qkv_3_cast_fp16)[name = string("op_170_cast_fp16")];
+            tensor<int32, [1]> squeeze_3_axes_0 = const()[name = string("squeeze_3_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 512, 8, 64]> squeeze_3_cast_fp16 = squeeze(axes = squeeze_3_axes_0, x = var_170_cast_fp16_0)[name = string("squeeze_3_cast_fp16")];
+            tensor<int32, [1]> squeeze_4_axes_0 = const()[name = string("squeeze_4_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 512, 8, 64]> squeeze_4_cast_fp16 = squeeze(axes = squeeze_4_axes_0, x = var_170_cast_fp16_1)[name = string("squeeze_4_cast_fp16")];
+            tensor<int32, [1]> squeeze_5_axes_0 = const()[name = string("squeeze_5_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 512, 8, 64]> squeeze_5_cast_fp16 = squeeze(axes = squeeze_5_axes_0, x = var_170_cast_fp16_2)[name = string("squeeze_5_cast_fp16")];
+            tensor<int32, [4]> value_perm_0 = const()[name = string("value_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool var_178_transpose_x_0 = const()[name = string("op_178_transpose_x_0"), val = bool(false)];
+            bool var_178_transpose_y_0 = const()[name = string("op_178_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_10_perm_0 = const()[name = string("transpose_10_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_11_perm_0 = const()[name = string("transpose_11_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 512]> transpose_11 = transpose(perm = transpose_11_perm_0, x = squeeze_4_cast_fp16)[name = string("transpose_14")];
+            tensor<fp16, [1, 8, 512, 64]> transpose_10 = transpose(perm = transpose_10_perm_0, x = squeeze_3_cast_fp16)[name = string("transpose_15")];
+            tensor<fp16, [1, 8, 512, 512]> var_178_cast_fp16 = matmul(transpose_x = var_178_transpose_x_0, transpose_y = var_178_transpose_y_0, x = transpose_10, y = transpose_11)[name = string("op_178_cast_fp16")];
+            fp16 var_179_to_fp16 = const()[name = string("op_179_to_fp16"), val = fp16(0x1p-3)];
+            tensor<fp16, [1, 8, 512, 512]> scores_5_cast_fp16 = mul(x = var_178_cast_fp16, y = var_179_to_fp16)[name = string("scores_5_cast_fp16")];
+            tensor<fp16, [1, 8, 512, 512]> scores_cast_fp16 = select(a = var_23_to_fp16, b = scores_5_cast_fp16, cond = var_117)[name = string("scores_cast_fp16")];
+            tensor<fp16, [1, 8, 512, 512]> var_187_cast_fp16 = softmax(axis = var_30, x = scores_cast_fp16)[name = string("op_187_cast_fp16")];
+            tensor<fp16, [1, 8, 512, 512]> weights_5_cast_fp16 = mul(x = var_187_cast_fp16, y = var_115_to_fp16)[name = string("weights_5_cast_fp16")];
+            tensor<int32, [1]> var_191_axes_0 = const()[name = string("op_191_axes_0"), val = tensor<int32, [1]>([-1])];
+            bool var_191_keep_dims_0 = const()[name = string("op_191_keep_dims_0"), val = bool(true)];
+            tensor<fp16, [1, 8, 512, 1]> var_191_cast_fp16 = reduce_sum(axes = var_191_axes_0, keep_dims = var_191_keep_dims_0, x = weights_5_cast_fp16)[name = string("op_191_cast_fp16")];
+            tensor<fp16, [1, 8, 512, 1]> var_192_cast_fp16 = maximum(x = var_191_cast_fp16, y = var_21_to_fp16)[name = string("op_192_cast_fp16")];
+            tensor<fp16, [1, 8, 512, 512]> weights_cast_fp16 = real_div(x = weights_5_cast_fp16, y = var_192_cast_fp16)[name = string("weights_cast_fp16")];
+            bool var_194_transpose_x_0 = const()[name = string("op_194_transpose_x_0"), val = bool(false)];
+            bool var_194_transpose_y_0 = const()[name = string("op_194_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 512, 64]> value_cast_fp16 = transpose(perm = value_perm_0, x = squeeze_5_cast_fp16)[name = string("transpose_13")];
+            tensor<fp16, [1, 8, 512, 64]> var_194_cast_fp16 = matmul(transpose_x = var_194_transpose_x_0, transpose_y = var_194_transpose_y_0, x = weights_cast_fp16, y = value_cast_fp16)[name = string("op_194_cast_fp16")];
+            tensor<int32, [4]> var_195_perm_0 = const()[name = string("op_195_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_196 = const()[name = string("op_196"), val = tensor<int32, [3]>([1, 512, 512])];
+            tensor<fp16, [1, 512, 8, 64]> var_195_cast_fp16 = transpose(perm = var_195_perm_0, x = var_194_cast_fp16)[name = string("transpose_12")];
+            tensor<fp16, [1, 512, 512]> input_25_cast_fp16 = reshape(shape = var_196, x = var_195_cast_fp16)[name = string("input_25_cast_fp16")];
+            tensor<fp16, [512, 512]> module_flow_span_ordered_encoder_blocks_1_attention_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4407168))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4669376))))[name = string("module_flow_span_ordered_encoder_blocks_1_attention_proj_weight_to_fp16_quantized")];
+            tensor<fp16, [512]> module_flow_span_ordered_encoder_blocks_1_attention_proj_bias_to_fp16 = const()[name = string("module_flow_span_ordered_encoder_blocks_1_attention_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4670464)))];
+            tensor<fp16, [1, 512, 512]> linear_5_cast_fp16 = linear(bias = module_flow_span_ordered_encoder_blocks_1_attention_proj_bias_to_fp16, weight = module_flow_span_ordered_encoder_blocks_1_attention_proj_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_5_cast_fp16")];
+            tensor<fp16, [1, 512, 512]> input_29_cast_fp16 = add(x = hidden_7_cast_fp16, y = linear_5_cast_fp16)[name = string("input_29_cast_fp16")];
+            tensor<int32, [1]> input_31_axes_0 = const()[name = string("input_31_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> module_flow_span_ordered_encoder_blocks_1_norm_feedforward_weight_to_fp16 = const()[name = string("module_flow_span_ordered_encoder_blocks_1_norm_feedforward_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4671552)))];
+            tensor<fp16, [512]> module_flow_span_ordered_encoder_blocks_1_norm_feedforward_bias_to_fp16 = const()[name = string("module_flow_span_ordered_encoder_blocks_1_norm_feedforward_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4672640)))];
+            tensor<fp16, [1, 512, 512]> input_31_cast_fp16 = layer_norm(axes = input_31_axes_0, beta = module_flow_span_ordered_encoder_blocks_1_norm_feedforward_bias_to_fp16, epsilon = var_28_to_fp16, gamma = module_flow_span_ordered_encoder_blocks_1_norm_feedforward_weight_to_fp16, x = input_29_cast_fp16)[name = string("input_31_cast_fp16")];
+            tensor<fp16, [2048, 512]> module_flow_span_ordered_encoder_blocks_1_feedforward_0_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4673728))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5722368))))[name = string("module_flow_span_ordered_encoder_blocks_1_feedforward_0_weight_to_fp16_quantized")];
+            tensor<fp16, [2048]> module_flow_span_ordered_encoder_blocks_1_feedforward_0_bias_to_fp16 = const()[name = string("module_flow_span_ordered_encoder_blocks_1_feedforward_0_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5726528)))];
+            tensor<fp16, [1, 512, 2048]> linear_6_cast_fp16 = linear(bias = module_flow_span_ordered_encoder_blocks_1_feedforward_0_bias_to_fp16, weight = module_flow_span_ordered_encoder_blocks_1_feedforward_0_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_6_cast_fp16")];
+            string input_35_mode_0 = const()[name = string("input_35_mode_0"), val = string("TANH_APPROXIMATION")];
+            tensor<fp16, [1, 512, 2048]> input_35_cast_fp16 = gelu(mode = input_35_mode_0, x = linear_6_cast_fp16)[name = string("input_35_cast_fp16")];
+            tensor<fp16, [512, 2048]> module_flow_span_ordered_encoder_blocks_1_feedforward_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5730688))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6779328))))[name = string("module_flow_span_ordered_encoder_blocks_1_feedforward_2_weight_to_fp16_quantized")];
+            tensor<fp16, [512]> module_flow_span_ordered_encoder_blocks_1_feedforward_2_bias_to_fp16 = const()[name = string("module_flow_span_ordered_encoder_blocks_1_feedforward_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6780416)))];
+            tensor<fp16, [1, 512, 512]> linear_7_cast_fp16 = linear(bias = module_flow_span_ordered_encoder_blocks_1_feedforward_2_bias_to_fp16, weight = module_flow_span_ordered_encoder_blocks_1_feedforward_2_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_7_cast_fp16")];
+            tensor<fp16, [1, 512, 512]> hidden_9_cast_fp16 = add(x = input_29_cast_fp16, y = linear_7_cast_fp16)[name = string("hidden_9_cast_fp16")];
+            tensor<fp16, [1, 512, 512]> hidden_cast_fp16 = select(a = var_18_to_fp16, b = hidden_9_cast_fp16, cond = var_83)[name = string("hidden_cast_fp16")];
+            tensor<bool, [1, 512]> var_224 = equal(x = input_3, y = var_40)[name = string("op_224")];
+            tensor<bool, [1, 512]> var_225 = logical_and(x = span_context_mask_1, y = var_224)[name = string("op_225")];
+            tensor<int32, [1]> var_228_axes_0 = const()[name = string("op_228_axes_0"), val = tensor<int32, [1]>([1])];
+            bool var_228_keep_dims_0 = const()[name = string("op_228_keep_dims_0"), val = bool(true)];
+            string weight_15_to_fp16_dtype_0 = const()[name = string("weight_15_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 512]> var_225_to_fp16 = cast(dtype = weight_15_to_fp16_dtype_0, x = var_225)[name = string("cast_5")];
+            tensor<fp16, [1, 1]> var_228_cast_fp16 = reduce_sum(axes = var_228_axes_0, keep_dims = var_228_keep_dims_0, x = var_225_to_fp16)[name = string("op_228_cast_fp16")];
+            fp16 var_17_to_fp16 = const()[name = string("op_17_to_fp16"), val = fp16(0x1p+0)];
+            tensor<fp16, [1, 1]> var_229_cast_fp16 = maximum(x = var_228_cast_fp16, y = var_17_to_fp16)[name = string("op_229_cast_fp16")];
+            tensor<fp16, [1, 512]> weight_17_cast_fp16 = real_div(x = var_225_to_fp16, y = var_229_cast_fp16)[name = string("weight_17_cast_fp16")];
+            tensor<int32, [1]> var_231_axes_0 = const()[name = string("op_231_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1, 512, 1]> var_231_cast_fp16 = expand_dims(axes = var_231_axes_0, x = weight_17_cast_fp16)[name = string("op_231_cast_fp16")];
+            tensor<fp16, [1, 512, 512]> var_232_cast_fp16 = mul(x = hidden_cast_fp16, y = var_231_cast_fp16)[name = string("op_232_cast_fp16")];
+            tensor<int32, [1]> var_234_axes_0 = const()[name = string("op_234_axes_0"), val = tensor<int32, [1]>([1])];
+            bool var_234_keep_dims_0 = const()[name = string("op_234_keep_dims_0"), val = bool(false)];
+            tensor<fp16, [1, 512]> var_234_cast_fp16 = reduce_sum(axes = var_234_axes_0, keep_dims = var_234_keep_dims_0, x = var_232_cast_fp16)[name = string("op_234_cast_fp16")];
+            tensor<bool, [1, 512]> var_235 = equal(x = input_3, y = var_38)[name = string("op_235")];
+            tensor<bool, [1, 512]> var_236 = logical_and(x = span_context_mask_1, y = var_235)[name = string("op_236")];
+            tensor<int32, [1]> var_239_axes_0 = const()[name = string("op_239_axes_0"), val = tensor<int32, [1]>([1])];
+            bool var_239_keep_dims_0 = const()[name = string("op_239_keep_dims_0"), val = bool(true)];
+            string weight_19_to_fp16_dtype_0 = const()[name = string("weight_19_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 512]> var_236_to_fp16 = cast(dtype = weight_19_to_fp16_dtype_0, x = var_236)[name = string("cast_4")];
+            tensor<fp16, [1, 1]> var_239_cast_fp16 = reduce_sum(axes = var_239_axes_0, keep_dims = var_239_keep_dims_0, x = var_236_to_fp16)[name = string("op_239_cast_fp16")];
+            tensor<fp16, [1, 1]> var_240_cast_fp16 = maximum(x = var_239_cast_fp16, y = var_17_to_fp16)[name = string("op_240_cast_fp16")];
+            tensor<fp16, [1, 512]> weight_21_cast_fp16 = real_div(x = var_236_to_fp16, y = var_240_cast_fp16)[name = string("weight_21_cast_fp16")];
+            tensor<int32, [1]> var_242_axes_0 = const()[name = string("op_242_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1, 512, 1]> var_242_cast_fp16 = expand_dims(axes = var_242_axes_0, x = weight_21_cast_fp16)[name = string("op_242_cast_fp16")];
+            tensor<fp16, [1, 512, 512]> var_243_cast_fp16 = mul(x = hidden_cast_fp16, y = var_242_cast_fp16)[name = string("op_243_cast_fp16")];
+            tensor<int32, [1]> var_245_axes_0 = const()[name = string("op_245_axes_0"), val = tensor<int32, [1]>([1])];
+            bool var_245_keep_dims_0 = const()[name = string("op_245_keep_dims_0"), val = bool(false)];
+            tensor<fp16, [1, 512]> var_245_cast_fp16 = reduce_sum(axes = var_245_axes_0, keep_dims = var_245_keep_dims_0, x = var_243_cast_fp16)[name = string("op_245_cast_fp16")];
+            tensor<bool, [1, 512]> var_246 = equal(x = input_3, y = var_35)[name = string("op_246")];
+            tensor<bool, [1, 512]> var_247 = logical_and(x = span_context_mask_1, y = var_246)[name = string("op_247")];
+            tensor<int32, [1]> var_250_axes_0 = const()[name = string("op_250_axes_0"), val = tensor<int32, [1]>([1])];
+            bool var_250_keep_dims_0 = const()[name = string("op_250_keep_dims_0"), val = bool(true)];
+            string weight_23_to_fp16_dtype_0 = const()[name = string("weight_23_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 512]> var_247_to_fp16 = cast(dtype = weight_23_to_fp16_dtype_0, x = var_247)[name = string("cast_3")];
+            tensor<fp16, [1, 1]> var_250_cast_fp16 = reduce_sum(axes = var_250_axes_0, keep_dims = var_250_keep_dims_0, x = var_247_to_fp16)[name = string("op_250_cast_fp16")];
+            tensor<fp16, [1, 1]> var_251_cast_fp16 = maximum(x = var_250_cast_fp16, y = var_17_to_fp16)[name = string("op_251_cast_fp16")];
+            tensor<fp16, [1, 512]> weight_25_cast_fp16 = real_div(x = var_247_to_fp16, y = var_251_cast_fp16)[name = string("weight_25_cast_fp16")];
+            tensor<int32, [1]> var_253_axes_0 = const()[name = string("op_253_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1, 512, 1]> var_253_cast_fp16 = expand_dims(axes = var_253_axes_0, x = weight_25_cast_fp16)[name = string("op_253_cast_fp16")];
+            tensor<fp16, [1, 512, 512]> var_254_cast_fp16 = mul(x = hidden_cast_fp16, y = var_253_cast_fp16)[name = string("op_254_cast_fp16")];
+            tensor<int32, [1]> var_256_axes_0 = const()[name = string("op_256_axes_0"), val = tensor<int32, [1]>([1])];
+            bool var_256_keep_dims_0 = const()[name = string("op_256_keep_dims_0"), val = bool(false)];
+            tensor<fp16, [1, 512]> var_256_cast_fp16 = reduce_sum(axes = var_256_axes_0, keep_dims = var_256_keep_dims_0, x = var_254_cast_fp16)[name = string("op_256_cast_fp16")];
+            bool input_39_interleave_0 = const()[name = string("input_39_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1536]> input_39_cast_fp16 = concat(axis = var_30, interleave = input_39_interleave_0, values = (var_234_cast_fp16, var_245_cast_fp16, var_256_cast_fp16))[name = string("input_39_cast_fp16")];
+            tensor<fp16, [512, 1536]> module_flow_span_segment_pool_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6781504))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7568000))))[name = string("module_flow_span_segment_pool_proj_weight_to_fp16_quantized")];
+            tensor<fp16, [512]> module_flow_span_segment_pool_proj_bias_to_fp16 = const()[name = string("module_flow_span_segment_pool_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7569088)))];
+            tensor<fp16, [1, 512]> linear_8_cast_fp16 = linear(bias = module_flow_span_segment_pool_proj_bias_to_fp16, weight = module_flow_span_segment_pool_proj_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_8_cast_fp16")];
+            tensor<int32, [1]> pooled_axes_0 = const()[name = string("pooled_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> module_flow_span_context_norm_weight_to_fp16 = const()[name = string("module_flow_span_context_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7570176)))];
+            tensor<fp16, [512]> module_flow_span_context_norm_bias_to_fp16 = const()[name = string("module_flow_span_context_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7571264)))];
+            tensor<fp16, [1, 512]> pooled_cast_fp16 = layer_norm(axes = pooled_axes_0, beta = module_flow_span_context_norm_bias_to_fp16, epsilon = var_28_to_fp16, gamma = module_flow_span_context_norm_weight_to_fp16, x = linear_8_cast_fp16)[name = string("pooled_cast_fp16")];
+            tensor<int32, [1]> var_38_list = const()[name = string("op_38_list"), val = tensor<int32, [1]>([1])];
+            bool reduce_max_0_keep_dims_0 = const()[name = string("reduce_max_0_keep_dims_0"), val = bool(true)];
+            tensor<int32, [1, 1]> reduce_max_0 = reduce_max(axes = var_38_list, keep_dims = reduce_max_0_keep_dims_0, x = var_64)[name = string("reduce_max_0")];
+            string var_266_dtype_0 = const()[name = string("op_266_dtype_0"), val = string("bool")];
+            tensor<bool, [1, 1]> var_266 = cast(dtype = var_266_dtype_0, x = reduce_max_0)[name = string("cast_2")];
+            tensor<bool, [1, 1]> var_267 = logical_not(x = var_266)[name = string("op_267")];
+            tensor<fp16, [1, 512]> var_268_cast_fp16 = select(a = var_18_to_fp16, b = pooled_cast_fp16, cond = var_267)[name = string("op_268_cast_fp16")];
+            string var_268_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_268_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<fp32, [1, 512]> span_context_hidden = cast(dtype = var_268_cast_fp16_to_fp32_dtype_0, x = var_268_cast_fp16)[name = string("cast_1")];
+        } -> (span_context_hidden);
+}