Upload folder using huggingface_hub

canary_cross_kv.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3553add8e4c4f4351f2e127d0a9c4b9f0ee7885503db507603fdfcb35f395250
+size 243

canary_cross_kv.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:21cceed24d63e235b0d7a1bc93fbce5c040e9c6a3e4485bc6525ec874086baa7
+size 470

canary_cross_kv.mlmodelc/metadata.json

@@ -0,0 +1,77 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "Canary-1b-v2 cross-attention K\/V precompute (per-window, 8L)",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 8 × 1 × 8 × 188 × 128)",
+        "shortDescription" : "",
+        "shape" : "[8, 1, 8, 188, 128]",
+        "name" : "enc_k",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 8 × 1 × 8 × 188 × 128)",
+        "shortDescription" : "",
+        "shape" : "[8, 1, 8, 188, 128]",
+        "name" : "enc_v",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Float16",
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 9,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios18.linear" : 16,
+      "Ios18.transpose" : 16,
+      "Stack" : 2,
+      "Ios18.cast" : 3,
+      "Ios18.reshape" : 16,
+      "Ios18.mul" : 8
+    },
+    "computePrecision" : "Mixed (Float16, Float32, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "15.0",
+      "tvOS" : "18.0",
+      "visionOS" : "2.0",
+      "watchOS" : "11.0",
+      "iOS" : "18.0",
+      "macCatalyst" : "18.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 188 × 1024)",
+        "shortDescription" : "",
+        "shape" : "[1, 188, 1024]",
+        "name" : "enc_states",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.conversion_date" : "2026-07-01",
+      "com.github.apple.coremltools.source" : "torch==2.7.0",
+      "com.github.apple.coremltools.version" : "9.0b1",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "generatedClassName" : "canary_cross_kv",
+    "method" : "predict"
+  }
+]

canary_cross_kv.mlmodelc/model.mil

@@ -0,0 +1,144 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3600.16.1"}, {"coremlc-version", "3600.22.1"}, {"coremltools-component-torch", "2.7.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0b1"}})]
+{
+    func main<ios18>(tensor<fp32, [1, 188, 1024]> enc_states) {
+            string enc_states_to_fp16_dtype_0 = const()[name = string("enc_states_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1024, 1024]> layers_0_second_sub_layer_key_net_weight_to_fp16 = const()[name = string("layers_0_second_sub_layer_key_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<fp16, [1024]> layers_0_second_sub_layer_key_net_bias_to_fp16 = const()[name = string("layers_0_second_sub_layer_key_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2097280)))];
+            tensor<fp16, [1, 188, 1024]> enc_states_to_fp16 = cast(dtype = enc_states_to_fp16_dtype_0, x = enc_states)[name = string("cast_34")];
+            tensor<fp16, [1, 188, 1024]> linear_0_cast_fp16 = linear(bias = layers_0_second_sub_layer_key_net_bias_to_fp16, weight = layers_0_second_sub_layer_key_net_weight_to_fp16, x = enc_states_to_fp16)[name = string("linear_0_cast_fp16")];
+            tensor<int32, [4]> var_79 = const()[name = string("op_79"), val = tensor<int32, [4]>([1, 188, 8, 128])];
+            tensor<fp16, [1, 188, 8, 128]> var_80_cast_fp16 = reshape(shape = var_79, x = linear_0_cast_fp16)[name = string("op_80_cast_fp16")];
+            tensor<fp16, [1024, 1024]> layers_0_second_sub_layer_value_net_weight_to_fp16 = const()[name = string("layers_0_second_sub_layer_value_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2099392)))];
+            tensor<fp16, [1024]> layers_0_second_sub_layer_value_net_bias_to_fp16 = const()[name = string("layers_0_second_sub_layer_value_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4196608)))];
+            tensor<fp16, [1, 188, 1024]> linear_1_cast_fp16 = linear(bias = layers_0_second_sub_layer_value_net_bias_to_fp16, weight = layers_0_second_sub_layer_value_net_weight_to_fp16, x = enc_states_to_fp16)[name = string("linear_1_cast_fp16")];
+            tensor<int32, [4]> var_100 = const()[name = string("op_100"), val = tensor<int32, [4]>([1, 188, 8, 128])];
+            tensor<fp16, [1, 188, 8, 128]> var_101_cast_fp16 = reshape(shape = var_100, x = linear_1_cast_fp16)[name = string("op_101_cast_fp16")];
+            fp16 _inversed_109_y_0_to_fp16 = const()[name = string("_inversed_109_y_0_to_fp16"), val = fp16(0x1.308p-2)];
+            tensor<fp16, [1, 188, 8, 128]> _inversed_109_cast_fp16 = mul(x = var_80_cast_fp16, y = _inversed_109_y_0_to_fp16)[name = string("_inversed_109_cast_fp16")];
+            tensor<fp16, [1024, 1024]> layers_1_second_sub_layer_key_net_weight_to_fp16 = const()[name = string("layers_1_second_sub_layer_key_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4198720)))];
+            tensor<fp16, [1024]> layers_1_second_sub_layer_key_net_bias_to_fp16 = const()[name = string("layers_1_second_sub_layer_key_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6295936)))];
+            tensor<fp16, [1, 188, 1024]> linear_2_cast_fp16 = linear(bias = layers_1_second_sub_layer_key_net_bias_to_fp16, weight = layers_1_second_sub_layer_key_net_weight_to_fp16, x = enc_states_to_fp16)[name = string("linear_2_cast_fp16")];
+            tensor<int32, [4]> var_123 = const()[name = string("op_123"), val = tensor<int32, [4]>([1, 188, 8, 128])];
+            tensor<fp16, [1, 188, 8, 128]> var_124_cast_fp16 = reshape(shape = var_123, x = linear_2_cast_fp16)[name = string("op_124_cast_fp16")];
+            tensor<fp16, [1024, 1024]> layers_1_second_sub_layer_value_net_weight_to_fp16 = const()[name = string("layers_1_second_sub_layer_value_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6298048)))];
+            tensor<fp16, [1024]> layers_1_second_sub_layer_value_net_bias_to_fp16 = const()[name = string("layers_1_second_sub_layer_value_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8395264)))];
+            tensor<fp16, [1, 188, 1024]> linear_3_cast_fp16 = linear(bias = layers_1_second_sub_layer_value_net_bias_to_fp16, weight = layers_1_second_sub_layer_value_net_weight_to_fp16, x = enc_states_to_fp16)[name = string("linear_3_cast_fp16")];
+            tensor<int32, [4]> var_144 = const()[name = string("op_144"), val = tensor<int32, [4]>([1, 188, 8, 128])];
+            tensor<fp16, [1, 188, 8, 128]> var_145_cast_fp16 = reshape(shape = var_144, x = linear_3_cast_fp16)[name = string("op_145_cast_fp16")];
+            fp16 _inversed_153_y_0_to_fp16 = const()[name = string("_inversed_153_y_0_to_fp16"), val = fp16(0x1.308p-2)];
+            tensor<fp16, [1, 188, 8, 128]> _inversed_153_cast_fp16 = mul(x = var_124_cast_fp16, y = _inversed_153_y_0_to_fp16)[name = string("_inversed_153_cast_fp16")];
+            tensor<fp16, [1024, 1024]> layers_2_second_sub_layer_key_net_weight_to_fp16 = const()[name = string("layers_2_second_sub_layer_key_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8397376)))];
+            tensor<fp16, [1024]> layers_2_second_sub_layer_key_net_bias_to_fp16 = const()[name = string("layers_2_second_sub_layer_key_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10494592)))];
+            tensor<fp16, [1, 188, 1024]> linear_4_cast_fp16 = linear(bias = layers_2_second_sub_layer_key_net_bias_to_fp16, weight = layers_2_second_sub_layer_key_net_weight_to_fp16, x = enc_states_to_fp16)[name = string("linear_4_cast_fp16")];
+            tensor<int32, [4]> var_167 = const()[name = string("op_167"), val = tensor<int32, [4]>([1, 188, 8, 128])];
+            tensor<fp16, [1, 188, 8, 128]> var_168_cast_fp16 = reshape(shape = var_167, x = linear_4_cast_fp16)[name = string("op_168_cast_fp16")];
+            tensor<fp16, [1024, 1024]> layers_2_second_sub_layer_value_net_weight_to_fp16 = const()[name = string("layers_2_second_sub_layer_value_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10496704)))];
+            tensor<fp16, [1024]> layers_2_second_sub_layer_value_net_bias_to_fp16 = const()[name = string("layers_2_second_sub_layer_value_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12593920)))];
+            tensor<fp16, [1, 188, 1024]> linear_5_cast_fp16 = linear(bias = layers_2_second_sub_layer_value_net_bias_to_fp16, weight = layers_2_second_sub_layer_value_net_weight_to_fp16, x = enc_states_to_fp16)[name = string("linear_5_cast_fp16")];
+            tensor<int32, [4]> var_188 = const()[name = string("op_188"), val = tensor<int32, [4]>([1, 188, 8, 128])];
+            tensor<fp16, [1, 188, 8, 128]> var_189_cast_fp16 = reshape(shape = var_188, x = linear_5_cast_fp16)[name = string("op_189_cast_fp16")];
+            fp16 _inversed_197_y_0_to_fp16 = const()[name = string("_inversed_197_y_0_to_fp16"), val = fp16(0x1.308p-2)];
+            tensor<fp16, [1, 188, 8, 128]> _inversed_197_cast_fp16 = mul(x = var_168_cast_fp16, y = _inversed_197_y_0_to_fp16)[name = string("_inversed_197_cast_fp16")];
+            tensor<fp16, [1024, 1024]> layers_3_second_sub_layer_key_net_weight_to_fp16 = const()[name = string("layers_3_second_sub_layer_key_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12596032)))];
+            tensor<fp16, [1024]> layers_3_second_sub_layer_key_net_bias_to_fp16 = const()[name = string("layers_3_second_sub_layer_key_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14693248)))];
+            tensor<fp16, [1, 188, 1024]> linear_6_cast_fp16 = linear(bias = layers_3_second_sub_layer_key_net_bias_to_fp16, weight = layers_3_second_sub_layer_key_net_weight_to_fp16, x = enc_states_to_fp16)[name = string("linear_6_cast_fp16")];
+            tensor<int32, [4]> var_211 = const()[name = string("op_211"), val = tensor<int32, [4]>([1, 188, 8, 128])];
+            tensor<fp16, [1, 188, 8, 128]> var_212_cast_fp16 = reshape(shape = var_211, x = linear_6_cast_fp16)[name = string("op_212_cast_fp16")];
+            tensor<fp16, [1024, 1024]> layers_3_second_sub_layer_value_net_weight_to_fp16 = const()[name = string("layers_3_second_sub_layer_value_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14695360)))];
+            tensor<fp16, [1024]> layers_3_second_sub_layer_value_net_bias_to_fp16 = const()[name = string("layers_3_second_sub_layer_value_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16792576)))];
+            tensor<fp16, [1, 188, 1024]> linear_7_cast_fp16 = linear(bias = layers_3_second_sub_layer_value_net_bias_to_fp16, weight = layers_3_second_sub_layer_value_net_weight_to_fp16, x = enc_states_to_fp16)[name = string("linear_7_cast_fp16")];
+            tensor<int32, [4]> var_232 = const()[name = string("op_232"), val = tensor<int32, [4]>([1, 188, 8, 128])];
+            tensor<fp16, [1, 188, 8, 128]> var_233_cast_fp16 = reshape(shape = var_232, x = linear_7_cast_fp16)[name = string("op_233_cast_fp16")];
+            fp16 _inversed_241_y_0_to_fp16 = const()[name = string("_inversed_241_y_0_to_fp16"), val = fp16(0x1.308p-2)];
+            tensor<fp16, [1, 188, 8, 128]> _inversed_241_cast_fp16 = mul(x = var_212_cast_fp16, y = _inversed_241_y_0_to_fp16)[name = string("_inversed_241_cast_fp16")];
+            tensor<fp16, [1024, 1024]> layers_4_second_sub_layer_key_net_weight_to_fp16 = const()[name = string("layers_4_second_sub_layer_key_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16794688)))];
+            tensor<fp16, [1024]> layers_4_second_sub_layer_key_net_bias_to_fp16 = const()[name = string("layers_4_second_sub_layer_key_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18891904)))];
+            tensor<fp16, [1, 188, 1024]> linear_8_cast_fp16 = linear(bias = layers_4_second_sub_layer_key_net_bias_to_fp16, weight = layers_4_second_sub_layer_key_net_weight_to_fp16, x = enc_states_to_fp16)[name = string("linear_8_cast_fp16")];
+            tensor<int32, [4]> var_255 = const()[name = string("op_255"), val = tensor<int32, [4]>([1, 188, 8, 128])];
+            tensor<fp16, [1, 188, 8, 128]> var_256_cast_fp16 = reshape(shape = var_255, x = linear_8_cast_fp16)[name = string("op_256_cast_fp16")];
+            tensor<fp16, [1024, 1024]> layers_4_second_sub_layer_value_net_weight_to_fp16 = const()[name = string("layers_4_second_sub_layer_value_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18894016)))];
+            tensor<fp16, [1024]> layers_4_second_sub_layer_value_net_bias_to_fp16 = const()[name = string("layers_4_second_sub_layer_value_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20991232)))];
+            tensor<fp16, [1, 188, 1024]> linear_9_cast_fp16 = linear(bias = layers_4_second_sub_layer_value_net_bias_to_fp16, weight = layers_4_second_sub_layer_value_net_weight_to_fp16, x = enc_states_to_fp16)[name = string("linear_9_cast_fp16")];
+            tensor<int32, [4]> var_276 = const()[name = string("op_276"), val = tensor<int32, [4]>([1, 188, 8, 128])];
+            tensor<fp16, [1, 188, 8, 128]> var_277_cast_fp16 = reshape(shape = var_276, x = linear_9_cast_fp16)[name = string("op_277_cast_fp16")];
+            fp16 _inversed_285_y_0_to_fp16 = const()[name = string("_inversed_285_y_0_to_fp16"), val = fp16(0x1.308p-2)];
+            tensor<fp16, [1, 188, 8, 128]> _inversed_285_cast_fp16 = mul(x = var_256_cast_fp16, y = _inversed_285_y_0_to_fp16)[name = string("_inversed_285_cast_fp16")];
+            tensor<fp16, [1024, 1024]> layers_5_second_sub_layer_key_net_weight_to_fp16 = const()[name = string("layers_5_second_sub_layer_key_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20993344)))];
+            tensor<fp16, [1024]> layers_5_second_sub_layer_key_net_bias_to_fp16 = const()[name = string("layers_5_second_sub_layer_key_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23090560)))];
+            tensor<fp16, [1, 188, 1024]> linear_10_cast_fp16 = linear(bias = layers_5_second_sub_layer_key_net_bias_to_fp16, weight = layers_5_second_sub_layer_key_net_weight_to_fp16, x = enc_states_to_fp16)[name = string("linear_10_cast_fp16")];
+            tensor<int32, [4]> var_299 = const()[name = string("op_299"), val = tensor<int32, [4]>([1, 188, 8, 128])];
+            tensor<fp16, [1, 188, 8, 128]> var_300_cast_fp16 = reshape(shape = var_299, x = linear_10_cast_fp16)[name = string("op_300_cast_fp16")];
+            tensor<fp16, [1024, 1024]> layers_5_second_sub_layer_value_net_weight_to_fp16 = const()[name = string("layers_5_second_sub_layer_value_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23092672)))];
+            tensor<fp16, [1024]> layers_5_second_sub_layer_value_net_bias_to_fp16 = const()[name = string("layers_5_second_sub_layer_value_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25189888)))];
+            tensor<fp16, [1, 188, 1024]> linear_11_cast_fp16 = linear(bias = layers_5_second_sub_layer_value_net_bias_to_fp16, weight = layers_5_second_sub_layer_value_net_weight_to_fp16, x = enc_states_to_fp16)[name = string("linear_11_cast_fp16")];
+            tensor<int32, [4]> var_320 = const()[name = string("op_320"), val = tensor<int32, [4]>([1, 188, 8, 128])];
+            tensor<fp16, [1, 188, 8, 128]> var_321_cast_fp16 = reshape(shape = var_320, x = linear_11_cast_fp16)[name = string("op_321_cast_fp16")];
+            fp16 _inversed_329_y_0_to_fp16 = const()[name = string("_inversed_329_y_0_to_fp16"), val = fp16(0x1.308p-2)];
+            tensor<fp16, [1, 188, 8, 128]> _inversed_329_cast_fp16 = mul(x = var_300_cast_fp16, y = _inversed_329_y_0_to_fp16)[name = string("_inversed_329_cast_fp16")];
+            tensor<fp16, [1024, 1024]> layers_6_second_sub_layer_key_net_weight_to_fp16 = const()[name = string("layers_6_second_sub_layer_key_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25192000)))];
+            tensor<fp16, [1024]> layers_6_second_sub_layer_key_net_bias_to_fp16 = const()[name = string("layers_6_second_sub_layer_key_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27289216)))];
+            tensor<fp16, [1, 188, 1024]> linear_12_cast_fp16 = linear(bias = layers_6_second_sub_layer_key_net_bias_to_fp16, weight = layers_6_second_sub_layer_key_net_weight_to_fp16, x = enc_states_to_fp16)[name = string("linear_12_cast_fp16")];
+            tensor<int32, [4]> var_343 = const()[name = string("op_343"), val = tensor<int32, [4]>([1, 188, 8, 128])];
+            tensor<fp16, [1, 188, 8, 128]> var_344_cast_fp16 = reshape(shape = var_343, x = linear_12_cast_fp16)[name = string("op_344_cast_fp16")];
+            tensor<fp16, [1024, 1024]> layers_6_second_sub_layer_value_net_weight_to_fp16 = const()[name = string("layers_6_second_sub_layer_value_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27291328)))];
+            tensor<fp16, [1024]> layers_6_second_sub_layer_value_net_bias_to_fp16 = const()[name = string("layers_6_second_sub_layer_value_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29388544)))];
+            tensor<fp16, [1, 188, 1024]> linear_13_cast_fp16 = linear(bias = layers_6_second_sub_layer_value_net_bias_to_fp16, weight = layers_6_second_sub_layer_value_net_weight_to_fp16, x = enc_states_to_fp16)[name = string("linear_13_cast_fp16")];
+            tensor<int32, [4]> var_364 = const()[name = string("op_364"), val = tensor<int32, [4]>([1, 188, 8, 128])];
+            tensor<fp16, [1, 188, 8, 128]> var_365_cast_fp16 = reshape(shape = var_364, x = linear_13_cast_fp16)[name = string("op_365_cast_fp16")];
+            fp16 _inversed_373_y_0_to_fp16 = const()[name = string("_inversed_373_y_0_to_fp16"), val = fp16(0x1.308p-2)];
+            tensor<fp16, [1, 188, 8, 128]> _inversed_373_cast_fp16 = mul(x = var_344_cast_fp16, y = _inversed_373_y_0_to_fp16)[name = string("_inversed_373_cast_fp16")];
+            tensor<fp16, [1024, 1024]> layers_7_second_sub_layer_key_net_weight_to_fp16 = const()[name = string("layers_7_second_sub_layer_key_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29390656)))];
+            tensor<fp16, [1024]> layers_7_second_sub_layer_key_net_bias_to_fp16 = const()[name = string("layers_7_second_sub_layer_key_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31487872)))];
+            tensor<fp16, [1, 188, 1024]> linear_14_cast_fp16 = linear(bias = layers_7_second_sub_layer_key_net_bias_to_fp16, weight = layers_7_second_sub_layer_key_net_weight_to_fp16, x = enc_states_to_fp16)[name = string("linear_14_cast_fp16")];
+            tensor<int32, [4]> var_387 = const()[name = string("op_387"), val = tensor<int32, [4]>([1, 188, 8, 128])];
+            tensor<fp16, [1, 188, 8, 128]> var_388_cast_fp16 = reshape(shape = var_387, x = linear_14_cast_fp16)[name = string("op_388_cast_fp16")];
+            tensor<fp16, [1024, 1024]> layers_7_second_sub_layer_value_net_weight_to_fp16 = const()[name = string("layers_7_second_sub_layer_value_net_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31489984)))];
+            tensor<fp16, [1024]> layers_7_second_sub_layer_value_net_bias_to_fp16 = const()[name = string("layers_7_second_sub_layer_value_net_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33587200)))];
+            tensor<fp16, [1, 188, 1024]> linear_15_cast_fp16 = linear(bias = layers_7_second_sub_layer_value_net_bias_to_fp16, weight = layers_7_second_sub_layer_value_net_weight_to_fp16, x = enc_states_to_fp16)[name = string("linear_15_cast_fp16")];
+            tensor<int32, [4]> var_408 = const()[name = string("op_408"), val = tensor<int32, [4]>([1, 188, 8, 128])];
+            tensor<fp16, [1, 188, 8, 128]> var_409_cast_fp16 = reshape(shape = var_408, x = linear_15_cast_fp16)[name = string("op_409_cast_fp16")];
+            fp16 _inversed_417_y_0_to_fp16 = const()[name = string("_inversed_417_y_0_to_fp16"), val = fp16(0x1.308p-2)];
+            tensor<fp16, [1, 188, 8, 128]> _inversed_417_cast_fp16 = mul(x = var_388_cast_fp16, y = _inversed_417_y_0_to_fp16)[name = string("_inversed_417_cast_fp16")];
+            int32 var_420_axis_0 = const()[name = string("op_420_axis_0"), val = int32(0)];
+            tensor<int32, [4]> transpose_32_perm_0 = const()[name = string("transpose_32_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [4]> transpose_33_perm_0 = const()[name = string("transpose_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [4]> transpose_34_perm_0 = const()[name = string("transpose_34_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [4]> transpose_35_perm_0 = const()[name = string("transpose_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [4]> transpose_36_perm_0 = const()[name = string("transpose_36_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [4]> transpose_37_perm_0 = const()[name = string("transpose_37_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [4]> transpose_38_perm_0 = const()[name = string("transpose_38_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [4]> transpose_39_perm_0 = const()[name = string("transpose_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [1, 8, 188, 128]> transpose_39 = transpose(perm = transpose_39_perm_0, x = _inversed_417_cast_fp16)[name = string("transpose_56")];
+            tensor<fp16, [1, 8, 188, 128]> transpose_38 = transpose(perm = transpose_38_perm_0, x = _inversed_373_cast_fp16)[name = string("transpose_57")];
+            tensor<fp16, [1, 8, 188, 128]> transpose_37 = transpose(perm = transpose_37_perm_0, x = _inversed_329_cast_fp16)[name = string("transpose_58")];
+            tensor<fp16, [1, 8, 188, 128]> transpose_36 = transpose(perm = transpose_36_perm_0, x = _inversed_285_cast_fp16)[name = string("transpose_59")];
+            tensor<fp16, [1, 8, 188, 128]> transpose_35 = transpose(perm = transpose_35_perm_0, x = _inversed_241_cast_fp16)[name = string("transpose_60")];
+            tensor<fp16, [1, 8, 188, 128]> transpose_34 = transpose(perm = transpose_34_perm_0, x = _inversed_197_cast_fp16)[name = string("transpose_61")];
+            tensor<fp16, [1, 8, 188, 128]> transpose_33 = transpose(perm = transpose_33_perm_0, x = _inversed_153_cast_fp16)[name = string("transpose_62")];
+            tensor<fp16, [1, 8, 188, 128]> transpose_32 = transpose(perm = transpose_32_perm_0, x = _inversed_109_cast_fp16)[name = string("transpose_63")];
+            tensor<fp16, [8, 1, 8, 188, 128]> var_420_cast_fp16 = stack(axis = var_420_axis_0, values = (transpose_32, transpose_33, transpose_34, transpose_35, transpose_36, transpose_37, transpose_38, transpose_39))[name = string("op_420_cast_fp16")];
+            string var_420_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_420_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            int32 var_423_axis_0 = const()[name = string("op_423_axis_0"), val = int32(0)];
+            tensor<int32, [4]> transpose_40_perm_0 = const()[name = string("transpose_40_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [4]> transpose_41_perm_0 = const()[name = string("transpose_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [4]> transpose_42_perm_0 = const()[name = string("transpose_42_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [4]> transpose_43_perm_0 = const()[name = string("transpose_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [4]> transpose_44_perm_0 = const()[name = string("transpose_44_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [4]> transpose_45_perm_0 = const()[name = string("transpose_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [4]> transpose_46_perm_0 = const()[name = string("transpose_46_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [4]> transpose_47_perm_0 = const()[name = string("transpose_47_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp16, [1, 8, 188, 128]> transpose_47 = transpose(perm = transpose_47_perm_0, x = var_409_cast_fp16)[name = string("transpose_48")];
+            tensor<fp16, [1, 8, 188, 128]> transpose_46 = transpose(perm = transpose_46_perm_0, x = var_365_cast_fp16)[name = string("transpose_49")];
+            tensor<fp16, [1, 8, 188, 128]> transpose_45 = transpose(perm = transpose_45_perm_0, x = var_321_cast_fp16)[name = string("transpose_50")];
+            tensor<fp16, [1, 8, 188, 128]> transpose_44 = transpose(perm = transpose_44_perm_0, x = var_277_cast_fp16)[name = string("transpose_51")];
+            tensor<fp16, [1, 8, 188, 128]> transpose_43 = transpose(perm = transpose_43_perm_0, x = var_233_cast_fp16)[name = string("transpose_52")];
+            tensor<fp16, [1, 8, 188, 128]> transpose_42 = transpose(perm = transpose_42_perm_0, x = var_189_cast_fp16)[name = string("transpose_53")];
+            tensor<fp16, [1, 8, 188, 128]> transpose_41 = transpose(perm = transpose_41_perm_0, x = var_145_cast_fp16)[name = string("transpose_54")];
+            tensor<fp16, [1, 8, 188, 128]> transpose_40 = transpose(perm = transpose_40_perm_0, x = var_101_cast_fp16)[name = string("transpose_55")];
+            tensor<fp16, [8, 1, 8, 188, 128]> var_423_cast_fp16 = stack(axis = var_423_axis_0, values = (transpose_40, transpose_41, transpose_42, transpose_43, transpose_44, transpose_45, transpose_46, transpose_47))[name = string("op_423_cast_fp16")];
+            string var_423_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_423_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<fp32, [8, 1, 8, 188, 128]> enc_v = cast(dtype = var_423_cast_fp16_to_fp32_dtype_0, x = var_423_cast_fp16)[name = string("cast_32")];
+            tensor<fp32, [8, 1, 8, 188, 128]> enc_k = cast(dtype = var_420_cast_fp16_to_fp32_dtype_0, x = var_420_cast_fp16)[name = string("cast_33")];
+        } -> (enc_k, enc_v);
+}

canary_cross_kv.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:02bf8060427056b229b8406434f4ffd00748a7ecf4c22b463ddb87f33de510d2
+size 33589312

canary_decoder_kv.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d986857aada35955d23c8451f035387b7aadcf7d1ef59b6fa40d4e042650457b
+size 243

canary_decoder_kv.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0d6b71c6182ec837f211caed7fa42ae60faf82cd30e55312fa48ef6fef24b141
+size 957

canary_decoder_kv.mlmodelc/metadata.json

@@ -0,0 +1,265 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "Canary-1b-v2 KV-cache transformer decoder step (8L, iOS18 stateful, single token)",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 16384)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 16384]",
+        "name" : "log_probs",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Float16",
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 9,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios18.expandDims" : 1,
+      "Ios18.softmax" : 17,
+      "Ios18.mul" : 56,
+      "Ios18.matmul" : 32,
+      "Ios18.equal" : 1,
+      "Ios18.greaterEqual" : 1,
+      "Select" : 1,
+      "Ios18.readState" : 16,
+      "Ios18.gather" : 2,
+      "Ios18.add" : 50,
+      "Ios18.layerNorm" : 26,
+      "Ios18.sliceUpdate" : 16,
+      "Ios18.writeState" : 16,
+      "Ios18.reshape" : 49,
+      "Ios18.linear" : 65,
+      "Ios18.relu" : 8,
+      "Ios18.cast" : 10,
+      "Ios18.sub" : 1,
+      "Ios18.transpose" : 48,
+      "Ios18.log" : 1,
+      "Ios18.sliceByIndex" : 16
+    },
+    "computePrecision" : "Mixed (Float16, Float32, Int16, Int32, UInt16)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 1 × 8 × 238 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 8, 238, 128]",
+        "name" : "k_0",
+        "type" : "State"
+      },
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 1 × 8 × 238 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 8, 238, 128]",
+        "name" : "k_1",
+        "type" : "State"
+      },
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 1 × 8 × 238 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 8, 238, 128]",
+        "name" : "k_2",
+        "type" : "State"
+      },
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 1 × 8 × 238 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 8, 238, 128]",
+        "name" : "k_3",
+        "type" : "State"
+      },
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 1 × 8 × 238 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 8, 238, 128]",
+        "name" : "k_4",
+        "type" : "State"
+      },
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 1 × 8 × 238 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 8, 238, 128]",
+        "name" : "k_5",
+        "type" : "State"
+      },
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 1 × 8 × 238 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 8, 238, 128]",
+        "name" : "k_6",
+        "type" : "State"
+      },
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 1 × 8 × 238 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 8, 238, 128]",
+        "name" : "k_7",
+        "type" : "State"
+      },
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 1 × 8 × 238 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 8, 238, 128]",
+        "name" : "v_0",
+        "type" : "State"
+      },
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 1 × 8 × 238 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 8, 238, 128]",
+        "name" : "v_1",
+        "type" : "State"
+      },
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 1 × 8 × 238 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 8, 238, 128]",
+        "name" : "v_2",
+        "type" : "State"
+      },
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 1 × 8 × 238 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 8, 238, 128]",
+        "name" : "v_3",
+        "type" : "State"
+      },
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 1 × 8 × 238 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 8, 238, 128]",
+        "name" : "v_4",
+        "type" : "State"
+      },
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 1 × 8 × 238 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 8, 238, 128]",
+        "name" : "v_5",
+        "type" : "State"
+      },
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 1 × 8 × 238 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 8, 238, 128]",
+        "name" : "v_6",
+        "type" : "State"
+      },
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 1 × 8 × 238 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 8, 238, 128]",
+        "name" : "v_7",
+        "type" : "State"
+      }
+    ],
+    "availability" : {
+      "macOS" : "15.0",
+      "tvOS" : "18.0",
+      "visionOS" : "2.0",
+      "watchOS" : "11.0",
+      "iOS" : "18.0",
+      "macCatalyst" : "18.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 1]",
+        "name" : "token",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "pos",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 8 × 1 × 8 × 188 × 128)",
+        "shortDescription" : "",
+        "shape" : "[8, 1, 8, 188, 128]",
+        "name" : "enc_k",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 8 × 1 × 8 × 188 × 128)",
+        "shortDescription" : "",
+        "shape" : "[8, 1, 8, 188, 128]",
+        "name" : "enc_v",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 1 × 238)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 1, 238]",
+        "name" : "self_mask",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.conversion_date" : "2026-07-01",
+      "com.github.apple.coremltools.source" : "torch==2.7.0",
+      "com.github.apple.coremltools.version" : "9.0b1",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "generatedClassName" : "canary_decoder_kv",
+    "method" : "predict"
+  }
+]

canary_decoder_kv.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b1e1ca6a08e0ba5c8bae40847faf728fe77920245a163fe30a52cdd9f9f7dd02
+size 270864448

canary_encoder.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dbfd16062a736f344edce2c16c2fcb84e9a55ce5979fb1d26192c8846a902b24
+size 243

canary_encoder.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4d912b07f00d4fd24bd9b577faa8692c2075a65e560bfebcc649d68b691f5151
+size 488

canary_encoder.mlmodelc/metadata.json

@@ -0,0 +1,104 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "Canary-1b-v2 FastConformer encoder (32L, d=1024, ss=8)",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 188 × 1024)",
+        "shortDescription" : "",
+        "shape" : "[1, 188, 1024]",
+        "name" : "enc_states",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "encoder_length",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Float16",
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios17.floor" : 3,
+      "Ios17.logicalAnd" : 2,
+      "Ios17.reshape" : 193,
+      "Ios16.softmax" : 32,
+      "Ios17.matmul" : 96,
+      "Ios17.transpose" : 227,
+      "Split" : 32,
+      "Select" : 96,
+      "Ios17.expandDims" : 5,
+      "Ios17.add" : 230,
+      "Tile" : 1,
+      "Ios17.sliceByIndex" : 64,
+      "Ios16.sigmoid" : 32,
+      "Pad" : 64,
+      "Ios17.logicalNot" : 2,
+      "Ios17.layerNorm" : 160,
+      "Ios16.silu" : 96,
+      "Ios17.less" : 1,
+      "Ios17.conv" : 101,
+      "Ios16.relu" : 3,
+      "Ios17.cast" : 4,
+      "Ios17.linear" : 257,
+      "Ios17.mul" : 131
+    },
+    "computePrecision" : "Mixed (Float16, Float32, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 128 × 1501)",
+        "shortDescription" : "",
+        "shape" : "[1, 128, 1501]",
+        "name" : "mel",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "mel_length",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.conversion_date" : "2026-06-25",
+      "com.github.apple.coremltools.source" : "torch==2.7.0",
+      "com.github.apple.coremltools.version" : "9.0b1",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "generatedClassName" : "canary_encoder",
+    "method" : "predict"
+  }
+]

canary_encoder.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a23ab46649b973c30598b5340f4740101dea8ec6aabfe7f3b336ad3e4c5d71c8
+size 1579377472

canary_preprocessor.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cc0282e7ce8491ed76924c26495686079d3affc0f6eeb6ed55fdf713fa3aacb4
+size 243

canary_preprocessor.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1edcc2319f81e09463ca63578d405a1642f70858ff1e010b65bb14ce6205a720
+size 485

canary_preprocessor.mlmodelc/metadata.json

@@ -0,0 +1,104 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "Canary-1b-v2 mel preprocessor (15s window, 128 mels)",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 128 × 1501)",
+        "shortDescription" : "",
+        "shape" : "[1, 128, 1501]",
+        "name" : "mel",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "mel_length",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Mixed (Float32, Int32)",
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 9,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios18.expandDims" : 8,
+      "Ios18.mul" : 1,
+      "Ios18.floorDiv" : 1,
+      "Ios18.matmul" : 1,
+      "Identity" : 1,
+      "Ios18.realDiv" : 3,
+      "Ios18.greaterEqual" : 1,
+      "Select" : 3,
+      "Tile" : 1,
+      "Ios18.add" : 4,
+      "Ios16.reduceSum" : 4,
+      "Ios18.reshape" : 2,
+      "Pad" : 1,
+      "Ios18.conv" : 2,
+      "Ios18.concat" : 1,
+      "Ios18.log" : 1,
+      "Ios18.sub" : 4,
+      "Ios18.cast" : 3,
+      "Stack" : 1,
+      "Ios18.pow" : 2,
+      "Ios18.less" : 1,
+      "Ios18.sqrt" : 1,
+      "Ios18.sliceByIndex" : 3
+    },
+    "computePrecision" : "Mixed (Float32, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "15.0",
+      "tvOS" : "18.0",
+      "visionOS" : "2.0",
+      "watchOS" : "11.0",
+      "iOS" : "18.0",
+      "macCatalyst" : "18.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 240000)",
+        "shortDescription" : "",
+        "shape" : "[1, 240000]",
+        "name" : "audio_signal",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "audio_length",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.conversion_date" : "2026-06-25",
+      "com.github.apple.coremltools.source" : "torch==2.7.0",
+      "com.github.apple.coremltools.version" : "9.0b1",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "generatedClassName" : "canary_preprocessor",
+    "method" : "predict"
+  }
+]

canary_preprocessor.mlmodelc/model.mil

@@ -0,0 +1,122 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3600.16.1"}, {"coremlc-version", "3600.22.1"}, {"coremltools-component-torch", "2.7.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0b1"}})]
+{
+    func main<ios18>(tensor<int32, [1]> audio_length, tensor<fp32, [1, 240000]> audio_signal) {
+            int32 var_9 = const()[name = string("op_9"), val = int32(1)];
+            int32 var_10 = const()[name = string("op_10"), val = int32(160)];
+            fp32 var_24 = const()[name = string("op_24"), val = fp32(0x0p+0)];
+            fp32 var_25 = const()[name = string("op_25"), val = fp32(0x1.4f8b58p-17)];
+            int32 var_34 = const()[name = string("op_34"), val = int32(512)];
+            tensor<int32, [1]> var_35 = add(x = audio_length, y = var_34)[name = string("op_35")];
+            int32 var_36 = const()[name = string("op_36"), val = int32(512)];
+            tensor<int32, [1]> var_37 = sub(x = var_35, y = var_36)[name = string("op_37")];
+            tensor<int32, [1]> floor_div_0 = floor_div(x = var_37, y = var_10)[name = string("floor_div_0")];
+            string var_38_dtype_0 = const()[name = string("op_38_dtype_0"), val = string("fp32")];
+            fp32 var_39_promoted = const()[name = string("op_39_promoted"), val = fp32(0x1p+0)];
+            tensor<fp32, [1]> var_38 = cast(dtype = var_38_dtype_0, x = floor_div_0)[name = string("cast_12")];
+            tensor<fp32, [1]> seq_len_1 = add(x = var_38, y = var_39_promoted)[name = string("seq_len_1")];
+            string seq_len_dtype_0 = const()[name = string("seq_len_dtype_0"), val = string("int32")];
+            tensor<int32, [2]> var_43_begin_0 = const()[name = string("op_43_begin_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [2]> var_43_end_0 = const()[name = string("op_43_end_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<bool, [2]> var_43_end_mask_0 = const()[name = string("op_43_end_mask_0"), val = tensor<bool, [2]>([true, false])];
+            tensor<bool, [2]> var_43_squeeze_mask_0 = const()[name = string("op_43_squeeze_mask_0"), val = tensor<bool, [2]>([false, true])];
+            tensor<fp32, [1]> var_43 = slice_by_index(begin = var_43_begin_0, end = var_43_end_0, end_mask = var_43_end_mask_0, squeeze_mask = var_43_squeeze_mask_0, x = audio_signal)[name = string("op_43")];
+            tensor<int32, [1]> var_44_axes_0 = const()[name = string("op_44_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1]> var_44 = expand_dims(axes = var_44_axes_0, x = var_43)[name = string("op_44")];
+            tensor<int32, [2]> var_46_begin_0 = const()[name = string("op_46_begin_0"), val = tensor<int32, [2]>([0, 1])];
+            tensor<int32, [2]> var_46_end_0 = const()[name = string("op_46_end_0"), val = tensor<int32, [2]>([1, 240000])];
+            tensor<bool, [2]> var_46_end_mask_0 = const()[name = string("op_46_end_mask_0"), val = tensor<bool, [2]>([true, true])];
+            tensor<fp32, [1, 239999]> var_46 = slice_by_index(begin = var_46_begin_0, end = var_46_end_0, end_mask = var_46_end_mask_0, x = audio_signal)[name = string("op_46")];
+            tensor<int32, [2]> var_48_begin_0 = const()[name = string("op_48_begin_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [2]> var_48_end_0 = const()[name = string("op_48_end_0"), val = tensor<int32, [2]>([1, 239999])];
+            tensor<bool, [2]> var_48_end_mask_0 = const()[name = string("op_48_end_mask_0"), val = tensor<bool, [2]>([true, false])];
+            tensor<fp32, [1, 239999]> var_48 = slice_by_index(begin = var_48_begin_0, end = var_48_end_0, end_mask = var_48_end_mask_0, x = audio_signal)[name = string("op_48")];
+            fp32 var_49 = const()[name = string("op_49"), val = fp32(0x1.f0a3d8p-1)];
+            tensor<fp32, [1, 239999]> var_50 = mul(x = var_48, y = var_49)[name = string("op_50")];
+            tensor<fp32, [1, 239999]> var_51 = sub(x = var_46, y = var_50)[name = string("op_51")];
+            bool input_1_interleave_0 = const()[name = string("input_1_interleave_0"), val = bool(false)];
+            tensor<fp32, [1, 240000]> input_1 = concat(axis = var_9, interleave = input_1_interleave_0, values = (var_44, var_51))[name = string("input_1")];
+            tensor<int32, [3]> var_57 = const()[name = string("op_57"), val = tensor<int32, [3]>([1, 1, 240000])];
+            tensor<fp32, [1, 1, 240000]> input_3 = reshape(shape = var_57, x = input_1)[name = string("input_3")];
+            fp32 const_3 = const()[name = string("const_3"), val = fp32(0x0p+0)];
+            tensor<int32, [6]> input_5_pad_0 = const()[name = string("input_5_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 256, 256])];
+            string input_5_mode_0 = const()[name = string("input_5_mode_0"), val = string("reflect")];
+            tensor<fp32, [1, 1, 240512]> input_5 = pad(constant_val = const_3, mode = input_5_mode_0, pad = input_5_pad_0, x = input_3)[name = string("input_5")];
+            tensor<int32, [2]> var_63 = const()[name = string("op_63"), val = tensor<int32, [2]>([1, 240512])];
+            tensor<fp32, [1, 240512]> input = reshape(shape = var_63, x = input_5)[name = string("input")];
+            tensor<fp32, [257, 1, 512]> expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor<fp32, [257, 1, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<fp32, [257, 1, 512]> expand_dims_4 = const()[name = string("expand_dims_4"), val = tensor<fp32, [257, 1, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526464)))];
+            tensor<int32, [1]> expand_dims_5 = const()[name = string("expand_dims_5"), val = tensor<int32, [1]>([160])];
+            tensor<int32, [1]> expand_dims_6_axes_0 = const()[name = string("expand_dims_6_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1, 240512]> expand_dims_6 = expand_dims(axes = expand_dims_6_axes_0, x = input)[name = string("expand_dims_6")];
+            string conv_0_pad_type_0 = const()[name = string("conv_0_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> conv_0_pad_0 = const()[name = string("conv_0_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> conv_0_dilations_0 = const()[name = string("conv_0_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 conv_0_groups_0 = const()[name = string("conv_0_groups_0"), val = int32(1)];
+            tensor<fp32, [1, 257, 1501]> conv_0 = conv(dilations = conv_0_dilations_0, groups = conv_0_groups_0, pad = conv_0_pad_0, pad_type = conv_0_pad_type_0, strides = expand_dims_5, weight = expand_dims_3, x = expand_dims_6)[name = string("conv_0")];
+            string conv_1_pad_type_0 = const()[name = string("conv_1_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> conv_1_pad_0 = const()[name = string("conv_1_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> conv_1_dilations_0 = const()[name = string("conv_1_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 conv_1_groups_0 = const()[name = string("conv_1_groups_0"), val = int32(1)];
+            tensor<fp32, [1, 257, 1501]> conv_1 = conv(dilations = conv_1_dilations_0, groups = conv_1_groups_0, pad = conv_1_pad_0, pad_type = conv_1_pad_type_0, strides = expand_dims_5, weight = expand_dims_4, x = expand_dims_6)[name = string("conv_1")];
+            int32 stack_0_axis_0 = const()[name = string("stack_0_axis_0"), val = int32(-1)];
+            tensor<fp32, [1, 257, 1501, 2]> stack_0 = stack(axis = stack_0_axis_0, values = (conv_0, conv_1))[name = string("stack_0")];
+            fp32 var_17_promoted = const()[name = string("op_17_promoted"), val = fp32(0x1p+1)];
+            tensor<fp32, [1, 257, 1501, 2]> var_67 = pow(x = stack_0, y = var_17_promoted)[name = string("op_67")];
+            tensor<int32, [1]> var_69_axes_0 = const()[name = string("op_69_axes_0"), val = tensor<int32, [1]>([-1])];
+            bool var_69_keep_dims_0 = const()[name = string("op_69_keep_dims_0"), val = bool(false)];
+            tensor<fp32, [1, 257, 1501]> var_69 = reduce_sum(axes = var_69_axes_0, keep_dims = var_69_keep_dims_0, x = var_67)[name = string("op_69")];
+            tensor<fp32, [1, 257, 1501]> x_9 = identity(x = var_69)[name = string("x_9")];
+            tensor<fp32, [1, 128, 257]> const_6 = const()[name = string("const_6"), val = tensor<fp32, [1, 128, 257]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1052864)))];
+            bool x_11_transpose_x_0 = const()[name = string("x_11_transpose_x_0"), val = bool(false)];
+            bool x_11_transpose_y_0 = const()[name = string("x_11_transpose_y_0"), val = bool(false)];
+            tensor<fp32, [1, 128, 1501]> x_11 = matmul(transpose_x = x_11_transpose_x_0, transpose_y = x_11_transpose_y_0, x = const_6, y = x_9)[name = string("x_11")];
+            fp32 var_76 = const()[name = string("op_76"), val = fp32(0x1p-24)];
+            tensor<fp32, [1, 128, 1501]> var_77 = add(x = x_11, y = var_76)[name = string("op_77")];
+            fp32 x_13_epsilon_0 = const()[name = string("x_13_epsilon_0"), val = fp32(0x1p-149)];
+            tensor<fp32, [1, 128, 1501]> x_13 = log(epsilon = x_13_epsilon_0, x = var_77)[name = string("x_13")];
+            tensor<int32, [1, 1501]> var_82 = const()[name = string("op_82"), val = tensor<int32, [1, 1501]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1184512)))];
+            tensor<int32, [1]> var_85_axes_0 = const()[name = string("op_85_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> mel_length = cast(dtype = seq_len_dtype_0, x = seq_len_1)[name = string("cast_11")];
+            tensor<int32, [1, 1]> var_85 = expand_dims(axes = var_85_axes_0, x = mel_length)[name = string("op_85")];
+            tensor<bool, [1, 1501]> valid_mask = less(x = var_82, y = var_85)[name = string("valid_mask")];
+            tensor<int32, [1]> var_87_axes_0 = const()[name = string("op_87_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, 1501]> var_87 = expand_dims(axes = var_87_axes_0, x = valid_mask)[name = string("op_87")];
+            tensor<int32, [3]> var_87_after_broadcast_reps_0 = const()[name = string("op_87_after_broadcast_reps_0"), val = tensor<int32, [3]>([1, 128, 1])];
+            tensor<bool, [1, 128, 1501]> var_87_after_broadcast = tile(reps = var_87_after_broadcast_reps_0, x = var_87)[name = string("op_87_after_broadcast")];
+            tensor<fp32, [1, 128, 1501]> var_24_after_broadcast = const()[name = string("op_24_after_broadcast"), val = tensor<fp32, [1, 128, 1501]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1190592)))];
+            tensor<fp32, [1, 128, 1501]> var_88 = select(a = x_13, b = var_24_after_broadcast, cond = var_87_after_broadcast)[name = string("op_88")];
+            tensor<int32, [1]> x_mean_numerator_axes_0 = const()[name = string("x_mean_numerator_axes_0"), val = tensor<int32, [1]>([2])];
+            bool x_mean_numerator_keep_dims_0 = const()[name = string("x_mean_numerator_keep_dims_0"), val = bool(false)];
+            tensor<fp32, [1, 128]> x_mean_numerator = reduce_sum(axes = x_mean_numerator_axes_0, keep_dims = x_mean_numerator_keep_dims_0, x = var_88)[name = string("x_mean_numerator")];
+            string cast_2_dtype_0 = const()[name = string("cast_2_dtype_0"), val = string("fp32")];
+            tensor<int32, [1]> x_mean_denominator_axes_0 = const()[name = string("x_mean_denominator_axes_0"), val = tensor<int32, [1]>([1])];
+            bool x_mean_denominator_keep_dims_0 = const()[name = string("x_mean_denominator_keep_dims_0"), val = bool(false)];
+            tensor<fp32, [1, 1501]> cast_2 = cast(dtype = cast_2_dtype_0, x = valid_mask)[name = string("cast_10")];
+            tensor<fp32, [1]> x_mean_denominator = reduce_sum(axes = x_mean_denominator_axes_0, keep_dims = x_mean_denominator_keep_dims_0, x = cast_2)[name = string("x_mean_denominator")];
+            tensor<int32, [1]> var_93_axes_0 = const()[name = string("op_93_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1]> var_93 = expand_dims(axes = var_93_axes_0, x = x_mean_denominator)[name = string("op_93")];
+            tensor<fp32, [1, 128]> x_mean = real_div(x = x_mean_numerator, y = var_93)[name = string("x_mean")];
+            tensor<int32, [1]> var_96_axes_0 = const()[name = string("op_96_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp32, [1, 128, 1]> var_96 = expand_dims(axes = var_96_axes_0, x = x_mean)[name = string("op_96")];
+            tensor<fp32, [1, 128, 1501]> var_97 = sub(x = x_13, y = var_96)[name = string("op_97")];
+            tensor<fp32, [1, 128, 1501]> var_98 = select(a = var_97, b = var_24_after_broadcast, cond = var_87_after_broadcast)[name = string("op_98")];
+            fp32 var_17_promoted_1 = const()[name = string("op_17_promoted_1"), val = fp32(0x1p+1)];
+            tensor<fp32, [1, 128, 1501]> var_99 = pow(x = var_98, y = var_17_promoted_1)[name = string("op_99")];
+            tensor<int32, [1]> var_101_axes_0 = const()[name = string("op_101_axes_0"), val = tensor<int32, [1]>([2])];
+            bool var_101_keep_dims_0 = const()[name = string("op_101_keep_dims_0"), val = bool(false)];
+            tensor<fp32, [1, 128]> var_101 = reduce_sum(axes = var_101_axes_0, keep_dims = var_101_keep_dims_0, x = var_99)[name = string("op_101")];
+            fp32 var_103 = const()[name = string("op_103"), val = fp32(0x1p+0)];
+            tensor<fp32, [1, 1]> var_104 = sub(x = var_93, y = var_103)[name = string("op_104")];
+            tensor<fp32, [1, 128]> var_105 = real_div(x = var_101, y = var_104)[name = string("op_105")];
+            tensor<fp32, [1, 128]> x_std_1 = sqrt(x = var_105)[name = string("x_std_1")];
+            tensor<fp32, [1, 128]> x_std = add(x = x_std_1, y = var_25)[name = string("x_std")];
+            tensor<int32, [1]> var_110_axes_0 = const()[name = string("op_110_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp32, [1, 128, 1]> var_110 = expand_dims(axes = var_110_axes_0, x = x_std)[name = string("op_110")];
+            tensor<fp32, [1, 128, 1501]> x = real_div(x = var_97, y = var_110)[name = string("x")];
+            tensor<bool, [1, 1501]> mask = greater_equal(x = var_82, y = var_85)[name = string("mask")];
+            tensor<int32, [1]> var_119_axes_0 = const()[name = string("op_119_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, 1501]> var_119 = expand_dims(axes = var_119_axes_0, x = mask)[name = string("op_119")];
+            tensor<fp32, [1, 128, 1501]> mel = select(a = var_24, b = x, cond = var_119)[name = string("processed_signal")];
+        } -> (mel, mel_length);
+}

canary_preprocessor.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e50ec10d5387fdcf4ef7905025c8df3a454d34a575300c8a852b734c08a82344
+size 1959168

metadata.json

@@ -0,0 +1,53 @@
+{
+  "model": "canary-1b-v2",
+  "window_samples": 240000,
+  "precision": "FP16",
+  "mel_shape": [
+    1,
+    128,
+    1501
+  ],
+  "enc_states_shape": [
+    1,
+    188,
+    1024
+  ],
+  "enc_len": 188,
+  "components": {
+    "preprocessor": "canary_preprocessor.mlpackage",
+    "encoder": "canary_encoder.mlpackage"
+  },
+  "optimize": {
+    "frontend": "preprocessor+encoder (copied from canary_coreml_fp16enc_sliced)",
+    "decoder": "kv-cache cross+step (copied from canary_coreml_kv)",
+    "deploy": "compute_units=cpu_and_ne",
+    "note": "Self-contained KV-cache build: front end (preprocessor + encoder) from canary_coreml_fp16enc_sliced + stateful KV-cache decoder (cross_kv + decoder_kv) from canary_coreml_kv. Run with validate_e2e_kv.py / the Swift KV host."
+  },
+  "decoder_kv": {
+    "cross_file": "canary_cross_kv.mlpackage",
+    "step_file": "canary_decoder_kv.mlpackage",
+    "l_dec": 238,
+    "t_enc": 188,
+    "n_layers": 8,
+    "n_heads": 8,
+    "head_size": 128,
+    "max_generation_delta": 50,
+    "vocab_size": 16384,
+    "pad": 2,
+    "eos": 3,
+    "seed": [
+      16053,
+      7,
+      4,
+      16,
+      64,
+      64,
+      5,
+      9,
+      11,
+      13
+    ],
+    "kv_cache": true,
+    "neg_inf": -10000.0
+  }
+}