Upload Nemotron-3.5 ASR Streaming 0.6B variant

README.md

@@ -0,0 +1,189 @@
+---
+license: other
+license_name: nvidia-software-and-model-evaluation-license
+license_link: https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-software-and-model-evaluation-license/
+language:
+- en
+- es
+- de
+- fr
+- it
+- pt
+- ru
+- hi
+- zh
+- ja
+- ko
+- ar
+- nl
+- pl
+- tr
+- uk
+- ro
+- el
+- cs
+- hu
+- sv
+- da
+- fi
+- 'no'
+- sk
+- hr
+- bg
+- lt
+- et
+- lv
+- sl
+- th
+- vi
+- id
+- ms
+- bn
+- ur
+- fa
+- ta
+- te
+- mr
+- gu
+- kn
+- ml
+- si
+- ne
+- km
+- sw
+- am
+- ha
+- zu
+- yo
+- ig
+- af
+- rw
+- so
+- ny
+- ln
+- 'or'
+- he
+- ku
+- az
+- ka
+- hy
+- uz
+- tg
+- ky
+- qu
+- ay
+- gn
+- mi
+- haw
+- sm
+- to
+- mt
+tags:
+- automatic-speech-recognition
+- nemotron
+- streaming
+- conformer
+- rnn-t
+- multilingual
+- coreml
+base_model: nvidia/nemotron-3.5-asr-streaming-0.6b
+library_name: coreml
+pipeline_tag: automatic-speech-recognition
+---
+
+# Nemotron-3.5 ASR Streaming 0.6B — CoreML INT8
+
+Cache-aware streaming Conformer + RNN-T from NVIDIA, ported to CoreML (`.mlmodelc`) for on-device inference on Apple Silicon. 600 M params, 76 languages, native punctuation and capitalization. INT8 palettized encoder, FP16 decoder + joint. Ships `.mlmodelc` (compiled) — `.mlpackage` is not shipped because on-device `MLModel.compileModel()` produces non-deterministic output across iOS simulator vs device runtimes.
+
+## Model
+
+| | |
+|---|---|
+| Parameters | 600 M |
+| Architecture | FastConformer-CacheAware-RNN-T with language-conditioning prompt kernel |
+| Languages | 76 (see frontmatter) |
+| Sample rate | 16 kHz mono |
+| Streaming chunk | 320 ms (`att_context_size = [56, 3]`) |
+| Encoder quantization | INT8 palettized |
+| Decoder / joint dtype | FP16 |
+| On-disk size | 612 MB |
+
+## Files
+
+| File | Size | Description |
+|------|------|-------------|
+| `encoder.mlmodelc/` | 565 MB | 24-layer cache-aware Conformer encoder + prompt kernel |
+| `decoder.mlmodelc/` | 29 MB | RNN-T prediction net (2-layer LSTM, 640 dim) |
+| `joint.mlmodelc/` | 18 MB | Joint network (vocab=13088) |
+| `vocab.json` | 230 KB | SentencePiece pieces, id → string |
+| `languages.json` | 2 KB | Language tag → prompt slot (e.g. `"en-US": 0`) |
+| `config.json` | <1 KB | Streaming geometry + dims for the loader |
+
+## Performance
+
+M5 Pro (Apple Silicon), 50 samples per language from FLEURS test, streaming 320 ms chunks, compute units `.all`. Scoring uses Whisper `EnglishTextNormalizer` for English and `BasicTextNormalizer(split_letters=True)` for hi/ja (char-level), `BasicTextNormalizer` for de/fr/ar.
+
+### Accuracy
+
+| lang | WER % | CER % | Δ WER vs fp32 source |
+|------|------:|------:|---------------------:|
+| en_us |  9.59 |  4.26 | +0.26 |
+| de_de | 10.41 |  5.37 | +0.19 |
+| fr_fr | 12.18 |  4.84 | +1.05 |
+| ar_eg | 13.37 |  3.80 | +0.10 |
+| hi_in |  4.42 |  3.61 | −0.84 |
+| ja_jp | 17.66 | 12.09 | +0.69 |
+
+Quantization is essentially lossless. For ja/hi the published WER is char-level (matches NVIDIA's CJK methodology); CER is the more interpretable number for those scripts.
+
+### Streaming throughput + memory
+
+M5 Pro, 60 s long-form en_us audio, single thread, `.all` compute units:
+
+| metric | value |
+|--------|------:|
+| RTF (encode + decode) | 0.068 |
+| p50 chunk latency | 18.6 ms |
+| p99 chunk latency | 23.4 ms |
+| RSS post-load | 1046 MB |
+| RSS peak (mid-stream) | 1238 MB |
+
+## Usage
+
+### Python (coremltools)
+
+```python
+import coremltools as ct
+import numpy as np
+import json
+from huggingface_hub import snapshot_download
+
+bundle = snapshot_download("aufklarer/Nemotron-3.5-ASR-Streaming-0.6B-CoreML-INT8")
+encoder = ct.models.CompiledMLModel(f"{bundle}/encoder.mlmodelc")
+decoder = ct.models.CompiledMLModel(f"{bundle}/decoder.mlmodelc")
+joint   = ct.models.CompiledMLModel(f"{bundle}/joint.mlmodelc")
+
+slots = json.load(open(f"{bundle}/languages.json"))["promptDictionary"]
+lang_mask = np.zeros((1, 128), dtype=np.float32)
+lang_mask[0, slots["en-US"]] = 1.0
+# Feed 320 ms chunks; persist caches across calls (see inference reference below).
+```
+
+### Swift
+
+```swift
+import NemotronStreamingASR
+
+let model = try await NemotronStreamingASRModel.fromPretrained()
+let text = try model.transcribeAudio(audio, sampleRate: 16000, language: "en-US")
+```
+
+## Source
+
+Upstream: [nvidia/nemotron-3.5-asr-streaming-0.6b](https://huggingface.co/nvidia/nemotron-3.5-asr-streaming-0.6b).
+
+## Links
+
+- [speech-swift](https://github.com/soniqo/speech-swift) — Apple SDK with `NemotronStreamingASR`
+- [soniqo.audio](https://soniqo.audio)
+- [blog](https://soniqo.audio/blog)

config.json

@@ -0,0 +1,28 @@
+{
+  "modelName": "nemotron-3.5-asr-streaming-0.6b",
+  "numMelBins": 128,
+  "sampleRate": 16000,
+  "nFFT": 512,
+  "hopLength": 160,
+  "winLength": 400,
+  "preEmphasis": 0.97,
+  "encoderHidden": 1024,
+  "encoderLayers": 24,
+  "subsamplingFactor": 8,
+  "attentionLeftContext": 56,
+  "convCacheSize": 8,
+  "decoderHidden": 640,
+  "decoderLayers": 2,
+  "vocabSize": 13087,
+  "blankTokenId": 13087,
+  "numPrompts": 128,
+  "streaming": {
+    "chunkMs": 320,
+    "chunkSize": 4,
+    "rightContext": 3,
+    "lookaheadMs": 240,
+    "melFrames": 32,
+    "preCacheSize": 9,
+    "outputFrames": 4
+  }
+}

decoder.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4d58297228f44ca23752d3850c60a0559af78a6d4455b348cd8dc6364b8ed0d4
+size 243

decoder.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:15a3baf01a47f013b388b76b940ab5ec33ba84a8bc6025e5d5dba1ec9bf28762
+size 404

decoder.mlmodelc/model.mil

@@ -0,0 +1,53 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.11.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios18>(tensor<fp16, [2, 1, 640]> c, tensor<fp16, [2, 1, 640]> h, tensor<int32, [1, 1]> token) {
+            int32 y_1_axis_0 = const()[name = string("y_1_axis_0"), val = int32(0)];
+            int32 y_1_batch_dims_0 = const()[name = string("y_1_batch_dims_0"), val = int32(0)];
+            bool y_1_validate_indices_0 = const()[name = string("y_1_validate_indices_0"), val = bool(false)];
+            tensor<fp16, [13088, 640]> decoder_prediction_embed_weight_to_fp16 = const()[name = string("decoder_prediction_embed_weight_to_fp16"), val = tensor<fp16, [13088, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            string token_to_int16_dtype_0 = const()[name = string("token_to_int16_dtype_0"), val = string("int16")];
+            tensor<int16, [1, 1]> token_to_int16 = cast(dtype = token_to_int16_dtype_0, x = token)[name = string("cast_6")];
+            tensor<fp16, [1, 1, 640]> y_1_cast_fp16_cast_uint16 = gather(axis = y_1_axis_0, batch_dims = y_1_batch_dims_0, indices = token_to_int16, validate_indices = y_1_validate_indices_0, x = decoder_prediction_embed_weight_to_fp16)[name = string("y_1_cast_fp16_cast_uint16")];
+            tensor<int32, [3]> input_1_perm_0 = const()[name = string("input_1_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            int32 split_0_num_splits_0 = const()[name = string("split_0_num_splits_0"), val = int32(2)];
+            int32 split_0_axis_0 = const()[name = string("split_0_axis_0"), val = int32(0)];
+            tensor<fp16, [1, 1, 640]> split_0_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_0_cast_fp16_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h)[name = string("split_0_cast_fp16")];
+            int32 split_1_num_splits_0 = const()[name = string("split_1_num_splits_0"), val = int32(2)];
+            int32 split_1_axis_0 = const()[name = string("split_1_axis_0"), val = int32(0)];
+            tensor<fp16, [1, 1, 640]> split_1_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_1_cast_fp16_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c)[name = string("split_1_cast_fp16")];
+            tensor<int32, [1]> input0_1_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = string("input0_1_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input0_1_lstm_layer_0_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input0_1_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_0)[name = string("input0_1_lstm_layer_0_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input0_1_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = string("input0_1_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input0_1_lstm_layer_0_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input0_1_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_0)[name = string("input0_1_lstm_layer_0_lstm_c0_squeeze_cast_fp16")];
+            string input0_1_lstm_layer_0_direction_0 = const()[name = string("input0_1_lstm_layer_0_direction_0"), val = string("forward")];
+            bool input0_1_lstm_layer_0_output_sequence_0 = const()[name = string("input0_1_lstm_layer_0_output_sequence_0"), val = bool(true)];
+            string input0_1_lstm_layer_0_recurrent_activation_0 = const()[name = string("input0_1_lstm_layer_0_recurrent_activation_0"), val = string("sigmoid")];
+            string input0_1_lstm_layer_0_cell_activation_0 = const()[name = string("input0_1_lstm_layer_0_cell_activation_0"), val = string("tanh")];
+            string input0_1_lstm_layer_0_activation_0 = const()[name = string("input0_1_lstm_layer_0_activation_0"), val = string("tanh")];
+            tensor<fp16, [2560, 640]> concat_1_to_fp16 = const()[name = string("concat_1_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16752768)))];
+            tensor<fp16, [2560, 640]> concat_2_to_fp16 = const()[name = string("concat_2_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20029632)))];
+            tensor<fp16, [2560]> concat_0_to_fp16 = const()[name = string("concat_0_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23306496)))];
+            tensor<fp16, [1, 1, 640]> input_1_cast_fp16 = transpose(perm = input_1_perm_0, x = y_1_cast_fp16_cast_uint16)[name = string("transpose_1")];
+            tensor<fp16, [1, 1, 640]> input0_1_lstm_layer_0_cast_fp16_0, tensor<fp16, [1, 640]> input0_1_lstm_layer_0_cast_fp16_1, tensor<fp16, [1, 640]> input0_1_lstm_layer_0_cast_fp16_2 = lstm(activation = input0_1_lstm_layer_0_activation_0, bias = concat_0_to_fp16, cell_activation = input0_1_lstm_layer_0_cell_activation_0, direction = input0_1_lstm_layer_0_direction_0, initial_c = input0_1_lstm_layer_0_lstm_c0_squeeze_cast_fp16, initial_h = input0_1_lstm_layer_0_lstm_h0_squeeze_cast_fp16, output_sequence = input0_1_lstm_layer_0_output_sequence_0, recurrent_activation = input0_1_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_1_cast_fp16)[name = string("input0_1_lstm_layer_0_cast_fp16")];
+            tensor<int32, [1]> input0_1_lstm_h0_squeeze_axes_0 = const()[name = string("input0_1_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input0_1_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input0_1_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_1)[name = string("input0_1_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input0_1_lstm_c0_squeeze_axes_0 = const()[name = string("input0_1_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input0_1_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input0_1_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_1)[name = string("input0_1_lstm_c0_squeeze_cast_fp16")];
+            string input0_1_direction_0 = const()[name = string("input0_1_direction_0"), val = string("forward")];
+            bool input0_1_output_sequence_0 = const()[name = string("input0_1_output_sequence_0"), val = bool(true)];
+            string input0_1_recurrent_activation_0 = const()[name = string("input0_1_recurrent_activation_0"), val = string("sigmoid")];
+            string input0_1_cell_activation_0 = const()[name = string("input0_1_cell_activation_0"), val = string("tanh")];
+            string input0_1_activation_0 = const()[name = string("input0_1_activation_0"), val = string("tanh")];
+            tensor<fp16, [2560, 640]> concat_4_to_fp16 = const()[name = string("concat_4_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23311680)))];
+            tensor<fp16, [2560, 640]> concat_5_to_fp16 = const()[name = string("concat_5_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26588544)))];
+            tensor<fp16, [2560]> concat_3_to_fp16 = const()[name = string("concat_3_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29865408)))];
+            tensor<fp16, [1, 1, 640]> input0_1_cast_fp16_0, tensor<fp16, [1, 640]> input0_1_cast_fp16_1, tensor<fp16, [1, 640]> input0_1_cast_fp16_2 = lstm(activation = input0_1_activation_0, bias = concat_3_to_fp16, cell_activation = input0_1_cell_activation_0, direction = input0_1_direction_0, initial_c = input0_1_lstm_c0_squeeze_cast_fp16, initial_h = input0_1_lstm_h0_squeeze_cast_fp16, output_sequence = input0_1_output_sequence_0, recurrent_activation = input0_1_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_ih = concat_4_to_fp16, x = input0_1_lstm_layer_0_cast_fp16_0)[name = string("input0_1_cast_fp16")];
+            int32 var_33_axis_0 = const()[name = string("op_33_axis_0"), val = int32(0)];
+            tensor<fp16, [2, 1, 640]> h_out = stack(axis = var_33_axis_0, values = (input0_1_lstm_layer_0_cast_fp16_1, input0_1_cast_fp16_1))[name = string("op_33_cast_fp16")];
+            int32 var_34_axis_0 = const()[name = string("op_34_axis_0"), val = int32(0)];
+            tensor<fp16, [2, 1, 640]> c_out = stack(axis = var_34_axis_0, values = (input0_1_lstm_layer_0_cast_fp16_2, input0_1_cast_fp16_2))[name = string("op_34_cast_fp16")];
+            tensor<int32, [3]> output_1_perm_0 = const()[name = string("output_1_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<fp16, [1, 1, 640]> decoder_output = transpose(perm = output_1_perm_0, x = input0_1_cast_fp16_0)[name = string("transpose_0")];
+        } -> (decoder_output, h_out, c_out);
+}

decoder.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dcdeccd4ccf46e2675224f9f030d46c1a89e2bda4abb316e901e1a21f1597f8f
+size 29870592

encoder.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cdd661c54b5e6f6389b58567c144cba6fad0d9e2c998d8a9048f87e857f9f32b
+size 243

encoder.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c37a6fb1813130e71e0a19e58c235974234f1b9b3984ec1afa9323bf3c192e8d
+size 701

encoder.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5fab9a684e34183d5f19cd1c63c19eefd7b02cc5a68af1367c6489702e746508
+size 591932416

joint.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e76f3618638fd5cf64805d4ba57942cbc7fcb1b38edd13851a151978871e57c4
+size 243

joint.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1b4924670d16462ab2de77fb7e5f7fa721af139939e5ab2ddcd903734894ce9d
+size 355

joint.mlmodelc/model.mil

@@ -0,0 +1,27 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.11.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios18>(tensor<fp16, [1, 1, 640]> decoder_output, tensor<fp16, [1, 1, 1024]> encoder_output) {
+            int32 var_6 = const()[name = string("op_6"), val = int32(-1)];
+            tensor<fp16, [640, 1024]> joint_enc_weight_to_fp16 = const()[name = string("joint_enc_weight_to_fp16"), val = tensor<fp16, [640, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<fp16, [640]> joint_enc_bias_to_fp16 = const()[name = string("joint_enc_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1310848)))];
+            tensor<fp16, [1, 1, 640]> linear_0_cast_fp16 = linear(bias = joint_enc_bias_to_fp16, weight = joint_enc_weight_to_fp16, x = encoder_output)[name = string("linear_0_cast_fp16")];
+            tensor<fp16, [640, 640]> joint_pred_weight_to_fp16 = const()[name = string("joint_pred_weight_to_fp16"), val = tensor<fp16, [640, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1312192)))];
+            tensor<fp16, [640]> joint_pred_bias_to_fp16 = const()[name = string("joint_pred_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2131456)))];
+            tensor<fp16, [1, 1, 640]> linear_1_cast_fp16 = linear(bias = joint_pred_bias_to_fp16, weight = joint_pred_weight_to_fp16, x = decoder_output)[name = string("linear_1_cast_fp16")];
+            tensor<int32, [1]> f_3_axes_0 = const()[name = string("f_3_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1, 1, 640]> f_3_cast_fp16 = expand_dims(axes = f_3_axes_0, x = linear_0_cast_fp16)[name = string("f_3_cast_fp16")];
+            tensor<int32, [1]> g_3_axes_0 = const()[name = string("g_3_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 640]> g_3_cast_fp16 = expand_dims(axes = g_3_axes_0, x = linear_1_cast_fp16)[name = string("g_3_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_3_cast_fp16 = add(x = f_3_cast_fp16, y = g_3_cast_fp16)[name = string("input_3_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> var_25_cast_fp16 = relu(x = input_3_cast_fp16)[name = string("op_25_cast_fp16")];
+            tensor<fp16, [13088, 640]> joint_joint_net_2_weight_to_fp16 = const()[name = string("joint_joint_net_2_weight_to_fp16"), val = tensor<fp16, [13088, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2132800)))];
+            tensor<fp16, [13088]> joint_joint_net_2_bias_to_fp16 = const()[name = string("joint_joint_net_2_bias_to_fp16"), val = tensor<fp16, [13088]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18885504)))];
+            tensor<fp16, [1, 1, 1, 13088]> linear_2_cast_fp16 = linear(bias = joint_joint_net_2_bias_to_fp16, weight = joint_joint_net_2_weight_to_fp16, x = var_25_cast_fp16)[name = string("linear_2_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 13088]> logits_1_softmax_cast_fp16 = softmax(axis = var_6, x = linear_2_cast_fp16)[name = string("logits_1_softmax_cast_fp16")];
+            fp32 logits_1_epsilon_0 = const()[name = string("logits_1_epsilon_0"), val = fp32(0x1p-149)];
+            tensor<fp16, [1, 1, 1, 13088]> logits_1_cast_fp16 = log(epsilon = logits_1_epsilon_0, x = logits_1_softmax_cast_fp16)[name = string("logits_1_cast_fp16")];
+            tensor<int32, [1]> var_31_axes_0 = const()[name = string("op_31_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1, 13088]> logits = squeeze(axes = var_31_axes_0, x = logits_1_cast_fp16)[name = string("op_31_cast_fp16")];
+        } -> (logits);
+}

joint.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c0ef0a3a6598f962d2aad598dc6850e4428874033419817121e11f1fff4a9cfe
+size 18911744

languages.json

@@ -0,0 +1,127 @@
+{
+  "numPrompts": 128,
+  "promptDictionary": {
+    "en-US": 0,
+    "en": 0,
+    "en-GB": 1,
+    "enGB": 1,
+    "es-ES": 2,
+    "esES": 2,
+    "es-US": 3,
+    "es": 3,
+    "zh-CN": 4,
+    "zh-ZH": 4,
+    "zh-TW": 5,
+    "hi-IN": 6,
+    "hi": 6,
+    "hi-HI": 6,
+    "ar-AR": 7,
+    "ar": 7,
+    "fr-FR": 8,
+    "fr": 8,
+    "de-DE": 9,
+    "de": 9,
+    "ja-JP": 10,
+    "ja-JA": 10,
+    "ru-RU": 11,
+    "ru": 11,
+    "pt-BR": 12,
+    "pt-PT": 13,
+    "pt": 13,
+    "ko-KR": 14,
+    "ko": 14,
+    "ko-KO": 14,
+    "it-IT": 15,
+    "it": 15,
+    "nl-NL": 16,
+    "nl": 16,
+    "pl-PL": 17,
+    "pl": 17,
+    "tr-TR": 18,
+    "tr": 18,
+    "uk-UA": 19,
+    "uk": 19,
+    "ro-RO": 20,
+    "ro": 20,
+    "el-GR": 21,
+    "el": 21,
+    "cs-CZ": 22,
+    "cs": 22,
+    "hu-HU": 23,
+    "hu": 23,
+    "sv-SE": 24,
+    "sv": 24,
+    "da-DK": 25,
+    "da": 25,
+    "fi-FI": 26,
+    "fi": 26,
+    "no-NO": 27,
+    "no": 27,
+    "nb-NO": 103,
+    "nb": 103,
+    "nn-NO": 104,
+    "nn": 104,
+    "sk-SK": 28,
+    "sk": 28,
+    "hr-HR": 29,
+    "hr": 29,
+    "bg-BG": 30,
+    "bg": 30,
+    "lt-LT": 31,
+    "lt": 31,
+    "et-EE": 60,
+    "et": 60,
+    "lv-LV": 61,
+    "lv": 61,
+    "sl-SI": 62,
+    "sl": 62,
+    "th-TH": 32,
+    "vi-VN": 33,
+    "id-ID": 34,
+    "ms-MY": 35,
+    "bn-IN": 36,
+    "ur-PK": 37,
+    "fa-IR": 38,
+    "ta-IN": 39,
+    "te-IN": 40,
+    "mr-IN": 41,
+    "gu-IN": 42,
+    "kn-IN": 43,
+    "ml-IN": 44,
+    "si-LK": 45,
+    "ne-NP": 46,
+    "km-KH": 47,
+    "sw-KE": 48,
+    "am-ET": 49,
+    "ha-NG": 50,
+    "zu-ZA": 51,
+    "yo-NG": 52,
+    "ig-NG": 53,
+    "af-ZA": 54,
+    "rw-RW": 55,
+    "so-SO": 56,
+    "ny-MW": 57,
+    "ln-CD": 58,
+    "or-KE": 59,
+    "he-IL": 64,
+    "ku-TR": 65,
+    "az-AZ": 66,
+    "ka-GE": 67,
+    "hy-AM": 68,
+    "uz-UZ": 69,
+    "tg-TJ": 70,
+    "ky-KG": 71,
+    "qu-PE": 80,
+    "ay-BO": 81,
+    "gn-PY": 82,
+    "nah-MX": 83,
+    "mi-NZ": 96,
+    "haw-US": 97,
+    "sm-WS": 98,
+    "to-TO": 99,
+    "fr-CA": 100,
+    "mt-MT": 102,
+    "auto": 101
+  },
+  "autoSlot": 101
+}