feat: add Apple Silicon (MPS) support for macOS ARM64

Introduce a device abstraction layer (cosyvoice/utils/device.py) that
unifies CUDA, MPS, and CPU device management. Replace all hardcoded
CUDA-specific code paths in the inference pipeline with device-agnostic
alternatives, enabling CosyVoice to run natively on Apple Silicon Macs.

Key changes:
- Device abstraction: get_device(), get_stream_context(),
  get_autocast_context(), empty_cache()
- model.py: Replace CUDA device init, streams, AMP, and cache clearing
  across CosyVoiceModel, CosyVoice2Model, CosyVoice3Model
- cosyvoice.py: MPS-aware feature gates (TRT/vLLM require CUDA,
  JIT/fp16 require any GPU)
- frontend.py: CoreMLExecutionProvider support for ONNX Runtime
- common.py: Guard torch.cuda.manual_seed_all for non-CUDA environments
- requirements.txt: Remove CUDA-only index URLs, loosen PyTorch version
- setup_macos.sh: One-command setup script for Apple Silicon

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: 2 people

README.md

@@ -126,6 +126,32 @@ pip install --pre torch torchvision torchaudio --index-url https://download.pyto
 # その他の依存関係
 pip install -r requirements.txt
 
+### macOS Apple Silicon (M1/M2/M3/M4)
+
+For Apple Silicon Macs, use the dedicated setup script:
+
+``` sh
+git clone --recursive https://github.com/FunAudioLLM/CosyVoice.git
+cd CosyVoice
+bash setup_macos.sh
+```
+
+Or manually:
+
+``` sh
+conda create -n cosyvoice -y python=3.10
+conda activate cosyvoice
+conda install -c conda-forge pynini==2.1.5 -y
+pip install torch torchaudio
+pip install -r requirements.txt
+```
+
+**Apple Silicon notes:**
+- Inference runs on MPS (Metal Performance Shaders) — faster than CPU
+- TensorRT and vLLM are not available (CUDA-only)
+- Training with DeepSpeed/DDP is not supported
+- For CUDA environments (Linux), use `pip install -r requirements-cuda.txt` instead
+
 # Whisper（自動文字起こし用）
 pip install openai-whisper
 

cosyvoice/cli/cosyvoice.py

@@ -22,6 +22,7 @@
 from cosyvoice.cli.model import CosyVoiceModel, CosyVoice2Model, CosyVoice3Model
 from cosyvoice.utils.file_utils import logging
 from cosyvoice.utils.class_utils import get_model_type
+from cosyvoice.utils.device import is_cuda, is_gpu_available
 
 
 class CosyVoice:
@@ -44,9 +45,12 @@ def __init__(self, model_dir, load_jit=False, load_trt=False, fp16=False, trt_co
                                           '{}/spk2info.pt'.format(model_dir),
                                           configs['allowed_special'])
         self.sample_rate = configs['sample_rate']
-        if torch.cuda.is_available() is False and (load_jit is True or load_trt is True or fp16 is True):
-            load_jit, load_trt, fp16 = False, False, False
-            logging.warning('no cuda device, set load_jit/load_trt/fp16 to False')
+        if not is_cuda() and load_trt:
+            load_trt = False
+            logging.warning('TensorRT requires CUDA, disabling load_trt')
+        if not is_gpu_available() and (load_jit or fp16):
+            load_jit, fp16 = False, False
+            logging.warning('no GPU device, disabling load_jit/fp16')
         self.model = CosyVoiceModel(configs['llm'], configs['flow'], configs['hift'], fp16)
         self.model.load('{}/llm.pt'.format(model_dir),
                         '{}/flow.pt'.format(model_dir),
@@ -156,9 +160,16 @@ def __init__(self, model_dir, load_jit=False, load_trt=False, load_vllm=False, f
                                           '{}/spk2info.pt'.format(model_dir),
                                           configs['allowed_special'])
         self.sample_rate = configs['sample_rate']
-        if torch.cuda.is_available() is False and (load_jit is True or load_trt is True or load_vllm is True or fp16 is True):
-            load_jit, load_trt, load_vllm, fp16 = False, False, False, False
-            logging.warning('no cuda device, set load_jit/load_trt/load_vllm/fp16 to False')
+        if not is_cuda():
+            if load_trt:
+                load_trt = False
+                logging.warning('TensorRT requires CUDA, disabling load_trt')
+            if load_vllm:
+                load_vllm = False
+                logging.warning('vLLM requires CUDA, disabling load_vllm')
+        if not is_gpu_available() and (load_jit or fp16):
+            load_jit, fp16 = False, False
+            logging.warning('no GPU device, disabling load_jit/fp16')
         self.model = CosyVoice2Model(configs['llm'], configs['flow'], configs['hift'], fp16)
         self.model.load('{}/llm.pt'.format(model_dir),
                         '{}/flow.pt'.format(model_dir),
@@ -206,9 +217,12 @@ def __init__(self, model_dir, load_trt=False, load_vllm=False, fp16=False, trt_c
                                           '{}/spk2info.pt'.format(model_dir),
                                           configs['allowed_special'])
         self.sample_rate = configs['sample_rate']
-        if torch.cuda.is_available() is False and (load_trt is True or fp16 is True):
-            load_trt, fp16 = False, False
-            logging.warning('no cuda device, set load_trt/fp16 to False')
+        if not is_cuda() and load_trt:
+            load_trt = False
+            logging.warning('TensorRT requires CUDA, disabling load_trt')
+        if not is_gpu_available() and fp16:
+            fp16 = False
+            logging.warning('no GPU device, disabling fp16')
         self.model = CosyVoice3Model(configs['llm'], configs['flow'], configs['hift'], fp16)
         self.model.load('{}/llm.pt'.format(model_dir),
                         '{}/flow.pt'.format(model_dir),

cosyvoice/cli/frontend.py

@@ -25,6 +25,7 @@
 import inflect
 from cosyvoice.utils.file_utils import logging, load_wav
 from cosyvoice.utils.frontend_utils import contains_chinese, replace_blank, replace_corner_mark, remove_bracket, spell_out_number, split_paragraph, is_only_punctuation
+from cosyvoice.utils.device import get_device
 
 
 class CosyVoiceFrontEnd:
@@ -38,14 +39,19 @@ def __init__(self,
                  allowed_special: str = 'all'):
         self.tokenizer = get_tokenizer()
         self.feat_extractor = feat_extractor
-        self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+        self.device = get_device()
         option = onnxruntime.SessionOptions()
         option.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL
         option.intra_op_num_threads = 1
         self.campplus_session = onnxruntime.InferenceSession(campplus_model, sess_options=option, providers=["CPUExecutionProvider"])
+        if torch.cuda.is_available():
+            tokenizer_providers = ["CUDAExecutionProvider"]
+        elif "CoreMLExecutionProvider" in onnxruntime.get_available_providers():
+            tokenizer_providers = ["CoreMLExecutionProvider"]
+        else:
+            tokenizer_providers = ["CPUExecutionProvider"]
         self.speech_tokenizer_session = onnxruntime.InferenceSession(speech_tokenizer_model, sess_options=option,
-                                                                     providers=["CUDAExecutionProvider" if torch.cuda.is_available() else
-                                                                                "CPUExecutionProvider"])
+                                                                     providers=tokenizer_providers)
         if os.path.exists(spk2info):
             self.spk2info = torch.load(spk2info, map_location=self.device, weights_only=True)
         else:

cosyvoice/cli/model.py

@@ -24,6 +24,7 @@
 from cosyvoice.utils.common import fade_in_out
 from cosyvoice.utils.file_utils import convert_onnx_to_trt, export_cosyvoice2_vllm
 from cosyvoice.utils.common import TrtContextWrapper
+from cosyvoice.utils.device import get_device, get_stream_context, get_autocast_context, empty_cache
 
 
 class CosyVoiceModel:
@@ -33,7 +34,7 @@ def __init__(self,
                  flow: torch.nn.Module,
                  hift: torch.nn.Module,
                  fp16: bool = False):
-        self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+        self.device = get_device()
         self.llm = llm
         self.flow = flow
         self.hift = hift
@@ -52,7 +53,7 @@ def __init__(self,
         # rtf and decoding related
         self.stream_scale_factor = 1
         assert self.stream_scale_factor >= 1, 'stream_scale_factor should be greater than 1, change it according to your actual rtf'
-        self.llm_context = torch.cuda.stream(torch.cuda.Stream(self.device)) if torch.cuda.is_available() else nullcontext()
+        self.llm_context = get_stream_context(self.device)
         self.lock = threading.Lock()
         # dict used to store session related variable
         self.tts_speech_token_dict = {}
@@ -100,7 +101,7 @@ def get_trt_kwargs(self):
 
     def llm_job(self, text, prompt_text, llm_prompt_speech_token, llm_embedding, uuid):
         cur_silent_token_num, max_silent_token_num = 0, 5
-        with self.llm_context, torch.cuda.amp.autocast(self.fp16 is True and hasattr(self.llm, 'vllm') is False):
+        with self.llm_context, get_autocast_context(self.fp16 is True and hasattr(self.llm, 'vllm') is False, self.device):
             if isinstance(text, Generator):
                 assert (self.__class__.__name__ != 'CosyVoiceModel') and not hasattr(self.llm, 'vllm'), 'streaming input text is only implemented for CosyVoice2/3 and do not support vllm!'
                 token_generator = self.llm.inference_bistream(text=text,
@@ -133,7 +134,7 @@ def vc_job(self, source_speech_token, uuid):
         self.llm_end_dict[uuid] = True
 
     def token2wav(self, token, prompt_token, prompt_feat, embedding, uuid, finalize=False, speed=1.0):
-        with torch.cuda.amp.autocast(self.fp16):
+        with get_autocast_context(self.fp16, self.device):
             tts_mel, self.flow_cache_dict[uuid] = self.flow.inference(token=token.to(self.device, dtype=torch.int32),
                                                                       token_len=torch.tensor([token.shape[1]], dtype=torch.int32).to(self.device),
                                                                       prompt_token=prompt_token.to(self.device),
@@ -237,9 +238,7 @@ def tts(self, text=torch.zeros(1, 0, dtype=torch.int32), flow_embedding=torch.ze
             self.mel_overlap_dict.pop(this_uuid)
             self.hift_cache_dict.pop(this_uuid)
             self.flow_cache_dict.pop(this_uuid)
-        if torch.cuda.is_available():
-            torch.cuda.empty_cache()
-            torch.cuda.current_stream().synchronize()
+        empty_cache(self.device)
 
 
 class CosyVoice2Model(CosyVoiceModel):
@@ -249,7 +248,7 @@ def __init__(self,
                  flow: torch.nn.Module,
                  hift: torch.nn.Module,
                  fp16: bool = False):
-        self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+        self.device = get_device()
         self.llm = llm
         self.flow = flow
         self.hift = hift
@@ -266,7 +265,7 @@ def __init__(self,
         # speech fade in out
         self.speech_window = np.hamming(2 * self.source_cache_len)
         # rtf and decoding related
-        self.llm_context = torch.cuda.stream(torch.cuda.Stream(self.device)) if torch.cuda.is_available() else nullcontext()
+        self.llm_context = get_stream_context(self.device)
         self.lock = threading.Lock()
         # dict used to store session related variable
         self.tts_speech_token_dict = {}
@@ -290,7 +289,7 @@ def load_vllm(self, model_dir):
         del self.llm.llm.model.model.layers
 
     def token2wav(self, token, prompt_token, prompt_feat, embedding, token_offset, uuid, stream=False, finalize=False, speed=1.0):
-        with torch.cuda.amp.autocast(self.fp16):
+        with get_autocast_context(self.fp16, self.device):
             tts_mel, _ = self.flow.inference(token=token.to(self.device, dtype=torch.int32),
                                              token_len=torch.tensor([token.shape[1]], dtype=torch.int32).to(self.device),
                                              prompt_token=prompt_token.to(self.device),
@@ -389,9 +388,7 @@ def tts(self, text=torch.zeros(1, 0, dtype=torch.int32), flow_embedding=torch.ze
             self.tts_speech_token_dict.pop(this_uuid)
             self.llm_end_dict.pop(this_uuid)
             self.hift_cache_dict.pop(this_uuid)
-        if torch.cuda.is_available():
-            torch.cuda.empty_cache()
-            torch.cuda.current_stream().synchronize()
+        empty_cache(self.device)
 
 
 class CosyVoice3Model(CosyVoice2Model):
@@ -401,7 +398,7 @@ def __init__(self,
                  flow: torch.nn.Module,
                  hift: torch.nn.Module,
                  fp16: bool = False):
-        self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+        self.device = get_device()
         self.llm = llm
         self.flow = flow
         self.hift = hift
@@ -413,7 +410,7 @@ def __init__(self,
         self.stream_scale_factor = 2
         assert self.stream_scale_factor >= 1, 'stream_scale_factor should be greater than 1, change it according to your actual rtf'
         # rtf and decoding related
-        self.llm_context = torch.cuda.stream(torch.cuda.Stream(self.device)) if torch.cuda.is_available() else nullcontext()
+        self.llm_context = get_stream_context(self.device)
         self.lock = threading.Lock()
         # dict used to store session related variable
         self.tts_speech_token_dict = {}
@@ -423,7 +420,7 @@ def __init__(self,
         self.silent_tokens = [1, 2, 28, 29, 55, 248, 494, 2241, 2242, 2322, 2323]
 
     def token2wav(self, token, prompt_token, prompt_feat, embedding, token_offset, uuid, stream=False, finalize=False, speed=1.0):
-        with torch.cuda.amp.autocast(self.fp16):
+        with get_autocast_context(self.fp16, self.device):
             tts_mel, _ = self.flow.inference(token=token.to(self.device, dtype=torch.int32),
                                              token_len=torch.tensor([token.shape[1]], dtype=torch.int32).to(self.device),
                                              prompt_token=prompt_token.to(self.device),

cosyvoice/utils/common.py

@@ -182,7 +182,8 @@ def set_all_random_seed(seed):
     random.seed(seed)
     np.random.seed(seed)
     torch.manual_seed(seed)
-    torch.cuda.manual_seed_all(seed)
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed_all(seed)
 
 
 def mask_to_bias(mask: torch.Tensor, dtype: torch.dtype) -> torch.Tensor:

cosyvoice/utils/device.py

@@ -0,0 +1,80 @@
+# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Unified device management for CUDA, MPS (Apple Silicon), and CPU backends."""
+
+import random
+from contextlib import nullcontext
+
+import numpy as np
+import torch
+
+
+def get_device() -> torch.device:
+    """Return the best available device: cuda > mps > cpu."""
+    if torch.cuda.is_available():
+        return torch.device('cuda')
+    if hasattr(torch.backends, 'mps') and torch.backends.mps.is_available():
+        return torch.device('mps')
+    return torch.device('cpu')
+
+
+def is_cuda() -> bool:
+    return torch.cuda.is_available()
+
+
+def is_mps() -> bool:
+    return hasattr(torch.backends, 'mps') and torch.backends.mps.is_available()
+
+
+def is_gpu_available() -> bool:
+    return is_cuda() or is_mps()
+
+
+def get_stream_context(device: torch.device):
+    """Return a CUDA stream context or nullcontext for non-CUDA devices."""
+    if device.type == 'cuda':
+        return torch.cuda.stream(torch.cuda.Stream(device))
+    return nullcontext()
+
+
+def get_autocast_context(enabled: bool, device: torch.device):
+    """Return the appropriate autocast context for the device."""
+    if not enabled:
+        return nullcontext()
+    if device.type == 'cuda':
+        return torch.cuda.amp.autocast(enabled=True)
+    if device.type == 'mps':
+        return torch.autocast(device_type='mps', dtype=torch.float16)
+    return nullcontext()
+
+
+def empty_cache(device: torch.device):
+    """Clear device cache and synchronize."""
+    if device.type == 'cuda':
+        torch.cuda.empty_cache()
+        torch.cuda.current_stream().synchronize()
+    elif device.type == 'mps':
+        if hasattr(torch.mps, 'empty_cache'):
+            torch.mps.empty_cache()
+        if hasattr(torch.mps, 'synchronize'):
+            torch.mps.synchronize()
+
+
+def set_all_random_seed(seed: int):
+    """Set random seed across all available backends."""
+    random.seed(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed_all(seed)

requirements-cuda.txt

@@ -0,0 +1,5 @@
+# CUDA-specific requirements (Linux with NVIDIA GPU)
+# Install with: pip install -r requirements-cuda.txt
+--extra-index-url https://download.pytorch.org/whl/cu121
+--extra-index-url https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/onnxruntime-cuda-12/pypi/simple/
+-r requirements.txt

requirements.txt

@@ -1,5 +1,3 @@
---extra-index-url https://download.pytorch.org/whl/cu121
---extra-index-url https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/onnxruntime-cuda-12/pypi/simple/ # https://github.com/microsoft/onnxruntime/issues/21684
 conformer==0.3.2
 deepspeed==0.15.1; sys_platform == 'linux'
 diffusers==0.29.0
@@ -33,8 +31,8 @@ tensorboard==2.14.0
 tensorrt-cu12==10.13.3.9; sys_platform == 'linux'
 tensorrt-cu12-bindings==10.13.3.9; sys_platform == 'linux'
 tensorrt-cu12-libs==10.13.3.9; sys_platform == 'linux'
-torch==2.3.1
-torchaudio==2.3.1
+torch>=2.3.1
+torchaudio>=2.3.1
 transformers==4.51.3
 x-transformers==2.11.24
 uvicorn==0.30.0