Upload 14 files

Flow-N250-fp16.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:284d3e102c55d68802a44985f987e70546706cc3331506a7919610672a0c02e9
+size 243

Flow-N250-fp16.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:26bb56ba0063c0938eb7329bb264d50ce9a23a88a54ddf7644bc6688d636ad8c
+size 491

Flow-N250-fp16.mlmodelc/metadata.json

@@ -0,0 +1,131 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Float16",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 80 × 500)",
+        "shortDescription" : "",
+        "shape" : "[1, 80, 500]",
+        "name" : "mel",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "num_prompt_mel",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios17.mul" : 2481,
+      "Ios17.sub" : 10,
+      "Ios16.softplus" : 20,
+      "Ios17.reshape" : 1762,
+      "Ios16.softmax" : 220,
+      "Ios17.matmul" : 440,
+      "Ios17.transpose" : 907,
+      "Ios17.maximum" : 1,
+      "Split" : 680,
+      "Select" : 1,
+      "Ios17.expandDims" : 1341,
+      "Ios17.add" : 1802,
+      "Tile" : 3,
+      "Ios17.sliceByIndex" : 880,
+      "Ios17.squeeze" : 880,
+      "Ios17.gather" : 1,
+      "Pad" : 22,
+      "Ios17.layerNorm" : 450,
+      "Ios16.silu" : 20,
+      "Ios17.conv" : 22,
+      "Ios17.clip" : 1,
+      "Ios17.cast" : 447,
+      "Ios17.realDiv" : 1,
+      "Ios17.linear" : 1581,
+      "Ios17.leakyRelu" : 1,
+      "Ios17.concat" : 462,
+      "Ios17.greaterEqual" : 1,
+      "Ios17.tanh" : 20,
+      "Stack" : 440,
+      "Ios16.gelu" : 220,
+      "Ios16.reduceL2Norm" : 1
+    },
+    "computePrecision" : "Mixed (Float16, Float32, Int16, 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"
+    },
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.conversion_date" : "2026-04-21",
+      "com.github.apple.coremltools.source" : "torch==2.11.0",
+      "com.github.apple.coremltools.version" : "9.0",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1 × 250)",
+        "shortDescription" : "",
+        "shape" : "[1, 250]",
+        "name" : "token_total",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "num_prompt_tokens",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 500 × 80)",
+        "shortDescription" : "",
+        "shape" : "[1, 500, 80]",
+        "name" : "prompt_feat",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 192)",
+        "shortDescription" : "",
+        "shape" : "[1, 192]",
+        "name" : "embedding",
+        "type" : "MultiArray"
+      }
+    ],
+    "generatedClassName" : "Flow_N250_fp16",
+    "method" : "predict"
+  }
+]

Flow-N250-fp16.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8ff4749144aecc55b175d7c1dcfa6f1d719c2b9a5b70fea76a1d3e7691b1ef30
+size 664579200

Flow-N250-fp16.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2fd2cb06bffd826546757e496e2b4800b77af50bd61bc42c81634361eec4a793
+size 4628237

Flow-N250-fp16.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8ff4749144aecc55b175d7c1dcfa6f1d719c2b9a5b70fea76a1d3e7691b1ef30
+size 664579200

Flow-N250-fp16.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "12143FB2-1600-4F9F-943E-3D1360C0B1D5": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        },
+        "FA7ED383-45E5-444D-B341-92A0A4649791": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        }
+    },
+    "rootModelIdentifier": "12143FB2-1600-4F9F-943E-3D1360C0B1D5"
+}

README.md

@@ -35,7 +35,7 @@ load either; `.mlmodelc` skips the one-time compile step on first use
 |---|---|---|---|
 | `LLM-Prefill-T256-M768-fp16` | CPU + ANE | Qwen2-0.5B prefill, 256-token context, 768-slot KV cache | fp16 |
 | `LLM-Decode-M768-fp16` | CPU + ANE | Single-step AR decode, 768-slot KV cache, 24 layers × 2 KV heads × 64 dim | fp16 |
-| `Flow-N250-fp32` | CPU only | Speech-token → mel (80-bin, 24 kHz), N_total=250 | fp32 (fp16 NaNs on fused LayerNorm) |
+| `Flow-N250-fp16` | CPU + GPU | Speech-token → mel (80-bin, 24 kHz), N_total=250 | fp16 (pure CPU overflows fused LayerNorm → NaN; ANE refuses to compile; GPU path uses fp32 accumulators internally and is stable) |
 | `HiFT-T500-fp16` | CPU + ANE | Mel → 24 kHz PCM, T=500 frames | fp16 |
 
 Total disk footprint (`.mlmodelc` + `.mlpackage` + runtime tables): ~6.6 GB on
@@ -51,11 +51,14 @@ disk. If you only need one format, delete the other after download.
 - `speech_embedding-fp16.safetensors` — 12 MB. CosyVoice3 `speech_embedding`
   table (6761 × 896 fp16); row-lookup per decoded speech token.
 
-`voices/`
-- `cosyvoice3-default-zh.safetensors` + `.json` — default zero-shot voice
-  bundle extracted from CosyVoice upstream `zero_shot_prompt.wav`
-  (utterance: "希望你以后能够做的比我还好呦。", N_speech = 87).
-  Schema documented in the voices repo README.
+`voices/` — 11 zero-shot voice bundles (~1 MB total)
+- `cosyvoice3-default-zh.safetensors` — default voice from CosyVoice upstream
+  `zero_shot_prompt.wav` (female, 希望你以后能够做的比我还好呦。, N_speech = 87).
+- `aishell3-zh-SSB*.safetensors` — 10 AISHELL-3 speakers bootstrapped via
+  `verify/bootstrap_aishell3_voices.py` (5 female + 5 male, north + south
+  accents). See `aishell3-bootstrap.json` for per-voice provenance.
+- Each `.safetensors` ships with a `.json` prompt-text sidecar and follows the
+  schema documented in the companion `cosyvoice3-voices-zh` repo.
 
 `tokenizer/`
 - `vocab.json` + `merges.txt` + `tokenizer_config.json` — stock Qwen2 BPE

manifest.json

@@ -10,7 +10,7 @@
     "target_platform": "Apple Silicon (M-series)",
     "min_os": "macOS 14 / iOS 17",
     "neural_engine": ["LLM-Prefill", "LLM-Decode", "HiFT"],
-    "cpu_only": ["Flow"]
+    "cpu_and_gpu": ["Flow"]
   },
   "model_graph": {
     "llm_hidden_dim": 896,
@@ -72,25 +72,24 @@
       }
     },
     {
-      "name": "Flow-N250-fp32",
+      "name": "Flow-N250-fp16",
       "paths": {
-        "mlpackage": "Flow-N250-fp32.mlpackage",
-        "mlmodelc":  "Flow-N250-fp32.mlmodelc"
+        "mlpackage": "Flow-N250-fp16.mlpackage",
+        "mlmodelc":  "Flow-N250-fp16.mlmodelc"
       },
-      "dtype": "fp32",
-      "compute_units": "cpuOnly",
-      "purpose": "Speech tokens -> 80-bin log-mel @ 24 kHz. fp16 produces NaNs on fused LayerNorm.",
-      "size_bytes": 1333084160,
+      "dtype": "fp16",
+      "compute_units": "cpuAndGPU",
+      "purpose": "Speech tokens -> 80-bin log-mel @ 24 kHz. Must run with cpuAndGPU: pure CPU overflows the fused LayerNorm and produces NaNs; ANE refuses to compile this graph (ANECCompile fails). GPU path uses fp32 accumulators internally and is stable + ~3x faster than the previous fp32/cpuOnly shipping config.",
+      "size_bytes": 669208054,
       "inputs": {
         "token_total": "[1, 250] int32 (prompt_ids || new_ids, right-padded)",
         "num_prompt_tokens": "[1] int32",
-        "num_new_tokens": "[1] int32",
         "prompt_feat": "[1, 500, 80] fp32 (right-padded)",
-        "num_prompt_mel": "[1] int32",
         "embedding": "[1, 192] fp32 (CAMPPlus speaker embedding)"
       },
       "outputs": {
-        "mel": "[1, 80, 500] fp32 (full buffer; slice to num_prompt_mel..num_prompt_mel+2*N_new)"
+        "mel": "[1, 80, 500] fp32 (full buffer; slice to num_prompt_mel..num_prompt_mel+2*N_new)",
+        "num_prompt_mel": "[1] int32"
       }
     },
     {