Merge pull request #183 from Lion-Wu/mps

Add MPS Backend Support for macOS

GPT_SoVITS/AR/data/bucket_sampler.py

@@ -41,12 +41,13 @@ class DistributedBucketSampler(Sampler[T_co]):
         if num_replicas is None:
             if not dist.is_available():
                 raise RuntimeError("Requires distributed package to be available")
-            num_replicas = dist.get_world_size()
+            num_replicas = dist.get_world_size() if torch.cuda.is_available() else 1
         if rank is None:
             if not dist.is_available():
                 raise RuntimeError("Requires distributed package to be available")
-            rank = dist.get_rank()
+            rank = dist.get_rank() if torch.cuda.is_available() else 0
-            torch.cuda.set_device(rank)
+            if torch.cuda.is_available():
+                torch.cuda.set_device(rank)
         if rank >= num_replicas or rank < 0:
             raise ValueError(
                 "Invalid rank {}, rank should be in the interval"

GPT_SoVITS/inference_webui.py

@@ -40,7 +40,15 @@ from my_utils import load_audio
 from tools.i18n.i18n import I18nAuto
 i18n = I18nAuto()
 
-device = "cuda"
+os.environ['PYTORCH_ENABLE_MPS_FALLBACK'] = '1' # 确保直接启动推理UI时也能够设置。
+
+if torch.cuda.is_available():
+    device = "cuda"
+elif torch.backends.mps.is_available():
+    device = "mps"
+else:
+    device = "cpu"
+
 tokenizer = AutoTokenizer.from_pretrained(bert_path)
 bert_model = AutoModelForMaskedLM.from_pretrained(bert_path)
 if is_half == True:

GPT_SoVITS/prepare_datasets/1-get-text.py

@@ -46,7 +46,7 @@ if os.path.exists(txt_path) == False:
     bert_dir = "%s/3-bert" % (opt_dir)
     os.makedirs(opt_dir, exist_ok=True)
     os.makedirs(bert_dir, exist_ok=True)
-    device = "cuda:0"
+    device = "cuda:0" if torch.cuda.is_available() else "mps"
     tokenizer = AutoTokenizer.from_pretrained(bert_pretrained_dir)
     bert_model = AutoModelForMaskedLM.from_pretrained(bert_pretrained_dir)
     if is_half == True:

GPT_SoVITS/prepare_datasets/2-get-hubert-wav32k.py

@@ -47,7 +47,7 @@ os.makedirs(wav32dir,exist_ok=True)
 
 maxx=0.95
 alpha=0.5
-device="cuda:0"
+device="cuda:0" if torch.cuda.is_available() else "mps"
 model=cnhubert.get_model()
 # is_half=False
 if(is_half==True):

GPT_SoVITS/prepare_datasets/3-get-semantic.py

@@ -38,7 +38,7 @@ semantic_path = "%s/6-name2semantic-%s.tsv" % (opt_dir, i_part)
 if os.path.exists(semantic_path) == False:
     os.makedirs(opt_dir, exist_ok=True)
 
-    device = "cuda:0"
+    device = "cuda:0" if torch.cuda.is_available() else "mps"
     hps = utils.get_hparams_from_file(s2config_path)
     vq_model = SynthesizerTrn(
         hps.data.filter_length // 2 + 1,

GPT_SoVITS/s1_train.py

@@ -116,9 +116,9 @@ def main(args):
         devices=-1,
         benchmark=False,
         fast_dev_run=False,
-        strategy=DDPStrategy(
+        strategy = "auto" if torch.backends.mps.is_available() else DDPStrategy(
             process_group_backend="nccl" if platform.system() != "Windows" else "gloo"
-        ),
+        ),  # mps 不支持多节点训练
         precision=config["train"]["precision"],
         logger=logger,
         num_sanity_val_steps=0,

GPT_SoVITS/s2_train.py

@@ -44,9 +44,12 @@ global_step = 0
 
 def main():
     """Assume Single Node Multi GPUs Training Only"""
-    assert torch.cuda.is_available(), "CPU training is not allowed."
+    assert torch.cuda.is_available() or torch.backends.mps.is_available(), "Only GPU training is allowed."
 
-    n_gpus = torch.cuda.device_count()
+    if torch.backends.mps.is_available():
+        n_gpus = 1
+    else:
+        n_gpus = torch.cuda.device_count()
     os.environ["MASTER_ADDR"] = "localhost"
     os.environ["MASTER_PORT"] = str(randint(20000, 55555))
 
@@ -70,13 +73,14 @@ def run(rank, n_gpus, hps):
         writer_eval = SummaryWriter(log_dir=os.path.join(hps.s2_ckpt_dir, "eval"))
 
     dist.init_process_group(
-        backend="gloo" if os.name == "nt" else "nccl",
+        backend = "gloo" if os.name == "nt" or torch.backends.mps.is_available() else "nccl",
         init_method="env://",
         world_size=n_gpus,
         rank=rank,
     )
     torch.manual_seed(hps.train.seed)
-    torch.cuda.set_device(rank)
+    if torch.cuda.is_available():
+        torch.cuda.set_device(rank)
 
     train_dataset = TextAudioSpeakerLoader(hps.data)  ########
     train_sampler = DistributedBucketSampler(
@@ -128,9 +132,14 @@ def run(rank, n_gpus, hps):
         hps.train.segment_size // hps.data.hop_length,
         n_speakers=hps.data.n_speakers,
         **hps.model,
-    ).cuda(rank)
+    ).cuda(rank) if torch.cuda.is_available() else SynthesizerTrn(
+        hps.data.filter_length // 2 + 1,
+        hps.train.segment_size // hps.data.hop_length,
+        n_speakers=hps.data.n_speakers,
+        **hps.model,
+    ).to("mps")
 
-    net_d = MultiPeriodDiscriminator(hps.model.use_spectral_norm).cuda(rank)
+    net_d = MultiPeriodDiscriminator(hps.model.use_spectral_norm).cuda(rank) if torch.cuda.is_available() else MultiPeriodDiscriminator(hps.model.use_spectral_norm).to("mps")
     for name, param in net_g.named_parameters():
         if not param.requires_grad:
             print(name, "not requires_grad")
@@ -174,8 +183,12 @@ def run(rank, n_gpus, hps):
         betas=hps.train.betas,
         eps=hps.train.eps,
     )
-    net_g = DDP(net_g, device_ids=[rank], find_unused_parameters=True)
+    if torch.cuda.is_available():
-    net_d = DDP(net_d, device_ids=[rank], find_unused_parameters=True)
+        net_g = DDP(net_g, device_ids=[rank], find_unused_parameters=True)
+        net_d = DDP(net_d, device_ids=[rank], find_unused_parameters=True)
+    else:
+        net_g = net_g.to("mps")
+        net_d = net_d.to("mps")
 
     try:  # 如果能加载自动resume
         _, _, _, epoch_str = utils.load_checkpoint(
@@ -205,6 +218,9 @@ def run(rank, n_gpus, hps):
                 net_g.module.load_state_dict(
                     torch.load(hps.train.pretrained_s2G, map_location="cpu")["weight"],
                     strict=False,
+                ) if torch.cuda.is_available() else net_g.load_state_dict(
+                    torch.load(hps.train.pretrained_s2G, map_location="cpu")["weight"],
+                    strict=False,
                 )
             )  ##测试不加载优化器
         if hps.train.pretrained_s2D != "":
@@ -213,6 +229,8 @@ def run(rank, n_gpus, hps):
             print(
                 net_d.module.load_state_dict(
                     torch.load(hps.train.pretrained_s2D, map_location="cpu")["weight"]
+                ) if torch.cuda.is_available() else net_d.load_state_dict(
+                    torch.load(hps.train.pretrained_s2D, map_location="cpu")["weight"]
                 )
             )
 
@@ -288,18 +306,26 @@ def train_and_evaluate(
         text,
         text_lengths,
     ) in tqdm(enumerate(train_loader)):
-        spec, spec_lengths = spec.cuda(rank, non_blocking=True), spec_lengths.cuda(
+        if torch.cuda.is_available():
-            rank, non_blocking=True
+            spec, spec_lengths = spec.cuda(rank, non_blocking=True), spec_lengths.cuda(
-        )
+                rank, non_blocking=True
-        y, y_lengths = y.cuda(rank, non_blocking=True), y_lengths.cuda(
+            )
-            rank, non_blocking=True
+            y, y_lengths = y.cuda(rank, non_blocking=True), y_lengths.cuda(
-        )
+                rank, non_blocking=True
-        ssl = ssl.cuda(rank, non_blocking=True)
+            )
-        ssl.requires_grad = False
+            ssl = ssl.cuda(rank, non_blocking=True)
-        # ssl_lengths = ssl_lengths.cuda(rank, non_blocking=True)
+            ssl.requires_grad = False
-        text, text_lengths = text.cuda(rank, non_blocking=True), text_lengths.cuda(
+            # ssl_lengths = ssl_lengths.cuda(rank, non_blocking=True)
-            rank, non_blocking=True
+            text, text_lengths = text.cuda(rank, non_blocking=True), text_lengths.cuda(
-        )
+                rank, non_blocking=True
+            )
+        else:
+            spec, spec_lengths = spec.to("mps"), spec_lengths.to("mps")
+            y, y_lengths = y.to("mps"), y_lengths.to("mps")
+            ssl = ssl.to("mps")
+            ssl.requires_grad = False
+            # ssl_lengths = ssl_lengths.cuda(rank, non_blocking=True)
+            text, text_lengths = text.to("mps"), text_lengths.to("mps")
 
         with autocast(enabled=hps.train.fp16_run):
             (
@@ -500,13 +526,21 @@ def evaluate(hps, generator, eval_loader, writer_eval):
             text_lengths,
         ) in enumerate(eval_loader):
             print(111)
-            spec, spec_lengths = spec.cuda(), spec_lengths.cuda()
+            if torch.cuda.is_available():
-            y, y_lengths = y.cuda(), y_lengths.cuda()
+                spec, spec_lengths = spec.cuda(), spec_lengths.cuda()
-            ssl = ssl.cuda()
+                y, y_lengths = y.cuda(), y_lengths.cuda()
-            text, text_lengths = text.cuda(), text_lengths.cuda()
+                ssl = ssl.cuda()
+                text, text_lengths = text.cuda(), text_lengths.cuda()
+            else:
+                spec, spec_lengths = spec.to("mps"), spec_lengths.to("mps")
+                y, y_lengths = y.to("mps"), y_lengths.to("mps")
+                ssl = ssl.to("mps")
+                text, text_lengths = text.to("mps"), text_lengths.to("mps")
             for test in [0, 1]:
                 y_hat, mask, *_ = generator.module.infer(
                     ssl, spec, spec_lengths, text, text_lengths, test=test
+                ) if torch.cuda.is_available() else generator.infer(
+                    ssl, spec, spec_lengths, text, text_lengths, test=test
                 )
                 y_hat_lengths = mask.sum([1, 2]).long() * hps.data.hop_length
 

README.md

@@ -43,9 +43,24 @@ If you are a Windows user (tested with win>=10) you can install directly via the
 
 - Python 3.9, PyTorch 2.0.1, CUDA 11
 - Python 3.10.13, PyTorch 2.1.2, CUDA 12.3
+- Python 3.9, PyTorch 2.3.0.dev20240122, macOS 14.3 (Apple Silicon, MPS)
 
 _Note: numba==0.56.4 require py<3.11_
 
+### For Mac Users
+If you are a Mac user, please install by using the following commands:
+#### Create  Environment
+```bash
+conda create -n GPTSoVits python=3.9
+conda activate GPTSoVits
+```
+#### Install Requirements
+```bash
+pip install -r requirements.txt
+pip uninstall torch torchaudio
+pip3 install --pre torch torchaudio --index-url https://download.pytorch.org/whl/nightly/cpu
+```
+_Note: For preprocessing with UVR5, it is recommended to [download the original project GUI](https://github.com/Anjok07/ultimatevocalremovergui) and select GPU for operation. Additionally, there may be memory leak issues when using Mac for inference, restarting the inference webUI can release the memory._
 ### Quick Install with Conda
 
 ```bash
@@ -58,16 +73,9 @@ bash install.sh
 #### Pip Packages
 
 ```bash
-pip install torch numpy scipy tensorboard librosa==0.9.2 numba==0.56.4 pytorch-lightning gradio==3.14.0 ffmpeg-python onnxruntime tqdm cn2an pypinyin pyopenjtalk g2p_en chardet transformers jieba_fast
+pip install -r requirements.txt
 ```
 
-#### Additional Requirements
-
-If you need Chinese ASR (supported by FunASR), install:
-
-```bash
-pip install modelscope torchaudio sentencepiece funasr
-```
 
 #### FFmpeg
 

api.py

@@ -30,25 +30,25 @@ endpoint: `/`
 
 使用执行参数指定的参考音频:
 GET:
-    `http://127.0.0.1:9880?text=你所热爱的，就是你的生活。&text_language=zh`
+    `http://127.0.0.1:9880?text=先帝创业未半而中道崩殂，今天下三分，益州疲弊，此诚危急存亡之秋也。&text_language=zh`
 POST:
 ```json
 {
-    "text": "你所热爱的，就是你的生活。",
+    "text": "先帝创业未半而中道崩殂，今天下三分，益州疲弊，此诚危急存亡之秋也。",
     "text_language": "zh"
 }
 ```
 
 手动指定当次推理所使用的参考音频:
 GET:
-    `http://127.0.0.1:9880?refer_wav_path=123.wav&prompt_text=一二三。&prompt_language=zh&text=你所热爱的，就是你的生活。&text_language=zh`
+    `http://127.0.0.1:9880?refer_wav_path=123.wav&prompt_text=一二三。&prompt_language=zh&text=先帝创业未半而中道崩殂，今天下三分，益州疲弊，此诚危急存亡之秋也。&text_language=zh`
 POST:
 ```json
 {
     "refer_wav_path": "123.wav",
     "prompt_text": "一二三。",
     "prompt_language": "zh",
-    "text": "你所热爱的，就是你的生活。",
+    "text": "先帝创业未半而中道崩殂，今天下三分，益州疲弊，此诚危急存亡之秋也。",
     "text_language": "zh"
 }
 ```
@@ -138,7 +138,8 @@ parser.add_argument("-dr", "--default_refer_path", type=str, default="", help="
 parser.add_argument("-dt", "--default_refer_text", type=str, default="", help="默认参考音频文本")
 parser.add_argument("-dl", "--default_refer_language", type=str, default="", help="默认参考音频语种")
 
-parser.add_argument("-d", "--device", type=str, default=g_config.infer_device, help="cuda / cpu")
+parser.add_argument("-d", "--device", type=str, default=g_config.infer_device, help="cuda / cpu / mps")
+parser.add_argument("-p", "--port", type=int, default=g_config.api_port, help="default: 9880")
 parser.add_argument("-a", "--bind_addr", type=str, default="127.0.0.1", help="default: 127.0.0.1")
 parser.add_argument("-p", "--port", type=int, default=g_config.api_port, help="default: 9880")
 parser.add_argument("-fp", "--full_precision", action="store_true", default=False, help="覆盖config.is_half为False, 使用全精度")
@@ -278,7 +279,7 @@ vq_model.eval()
 print(vq_model.load_state_dict(dict_s2["weight"], strict=False))
 hz = 50
 max_sec = config['data']['max_sec']
-t2s_model = Text2SemanticLightningModule(config, "ojbk", is_train=False)
+t2s_model = Text2SemanticLightningModule(config, "****", is_train=False)
 t2s_model.load_state_dict(dict_s1["weight"])
 if is_half:
     t2s_model = t2s_model.half()
@@ -439,6 +440,7 @@ def handle(refer_wav_path, prompt_text, prompt_language, text, text_language):
     wav.seek(0)
 
     torch.cuda.empty_cache()
+    torch.mps.empty_cache()
     return StreamingResponse(wav, media_type="audio/wav")
 
 

config.py

@@ -1,5 +1,6 @@
 import sys,os
 
+import torch
 
 # 推理用的指定模型
 sovits_path = ""
@@ -14,7 +15,12 @@ pretrained_gpt_path = "GPT_SoVITS/pretrained_models/s1bert25hz-2kh-longer-epoch=
 
 exp_root = "logs"
 python_exec = sys.executable or "python"
-infer_device = "cuda"
+if torch.cuda.is_available():
+    infer_device = "cuda"
+elif torch.backends.mps.is_available():
+    infer_device = "mps"
+else:
+    infer_device = "cpu"
 
 webui_port_main = 9874
 webui_port_uvr5 = 9873

docs/cn/README.md

@@ -38,10 +38,29 @@ https://github.com/RVC-Boss/GPT-SoVITS/assets/129054828/05bee1fa-bdd8-4d85-9350-
 
 如果你是Windows用户（已在win>=10上测试），可以直接通过预打包文件安装。只需下载[预打包文件](https://huggingface.co/lj1995/GPT-SoVITS-windows-package/resolve/main/GPT-SoVITS-beta.7z?download=true)，解压后双击go-webui.bat即可启动GPT-SoVITS-WebUI。
 
-### Python和PyTorch版本
 
-已在Python 3.9、PyTorch 2.0.1和CUDA 11上测试。
+### 测试通过的Python和PyTorch版本
 
+- Python 3.9、PyTorch 2.0.1和CUDA 11
+- Python 3.10.13, PyTorch 2.1.2和CUDA 12.3
+- Python 3.9、Pytorch 2.3.0.dev20240122和macOS 14.3（Apple 芯片，MPS）
+
+_注意: numba==0.56.4 需要 python<3.11_
+
+### Mac 用户
+如果你是Mac用户，请使用以下命令安装：
+#### 创建环境
+```bash
+conda create -n GPTSoVits python=3.9
+conda activate GPTSoVits
+```
+#### 安装依赖
+```bash
+pip install -r requirements.txt
+pip uninstall torch torchaudio
+pip3 install --pre torch torchaudio --index-url https://download.pytorch.org/whl/nightly/cpu
+```
+_注意：如需使用UVR5进行预处理，建议[下载原项目GUI](https://github.com/Anjok07/ultimatevocalremovergui)，勾选GPU运行。另外，使用Mac推理时可能存在内存泄漏问题，重启推理UI即可释放内存。_
 ### 使用Conda快速安装
 
 ```bash
@@ -53,15 +72,7 @@ bash install.sh
 #### Pip包
 
 ```bash
-pip install torch numpy scipy tensorboard librosa==0.9.2 numba==0.56.4 pytorch-lightning gradio==3.14.0 ffmpeg-python onnxruntime tqdm cn2an pypinyin pyopenjtalk g2p_en chardet transformers
+pip install -r requirements.txt
-```
-
-#### 额外要求
-
-如果你需要中文自动语音识别（由FunASR支持），请安装：
-
-```bash
-pip install modelscope torchaudio sentencepiece funasr
 ```
 
 #### FFmpeg

docs/ja/README.md

@@ -37,9 +37,26 @@ https://github.com/RVC-Boss/GPT-SoVITS/assets/129054828/05bee1fa-bdd8-4d85-9350-
 Windows ユーザーであれば（win>=10 にてテスト済み）、prezip 経由で直接インストールできます。[prezip](https://huggingface.co/lj1995/GPT-SoVITS-windows-package/resolve/main/GPT-SoVITS-beta.7z?download=true) をダウンロードして解凍し、go-webui.bat をダブルクリックするだけで GPT-SoVITS-WebUI が起動します。
 
 ### Python と PyTorch のバージョン
+- Python 3.9, PyTorch 2.0.1, CUDA 11
+- Python 3.10.13, PyTorch 2.1.2, CUDA 12.3
+- Python 3.9, PyTorch 2.3.0.dev20240122, macOS 14.3 (Apple Silicon, MPS)
 
-Python 3.9、PyTorch 2.0.1、CUDA 11でテスト済。
+_注記: numba==0.56.4 は py<3.11 が必要です_
 
+### Macユーザーへ
+Macユーザーの方は、以下のコマンドを使用してインストールしてください。
+#### 環境作成
+```bash
+conda create -n GPTSoVits python=3.9
+conda activate GPTSoVits
+```
+#### Pip パッケージ
+```bash
+pip install -r requirements.txt
+pip uninstall torch torchaudio
+pip3 install --pre torch torchaudio --index-url https://download.pytorch.org/whl/nightly/cpu
+```
+_注記: UVR5を使用した前処理には、[元のプロジェクトGUIをダウンロード](https://github.com/Anjok07/ultimatevocalremovergui)して、操作にGPUを選択することを推奨します。さらに、Macを使用して推論する際にメモリリークの問題が発生する可能性がありますが、推論のwebUIを再起動することでメモリを解放できます。_
 ### Conda によるクイックインストール
 
 ```bash
@@ -52,15 +69,7 @@ bash install.sh
 #### Pip パッケージ
 
 ```bash
-pip install torch numpy scipy tensorboard librosa==0.9.2 numba==0.56.4 pytorch-lightning gradio==3.14.0 ffmpeg-python onnxruntime tqdm cn2an pypinyin pyopenjtalk g2p_en chardet transformers
+pip install -r requirementx.txt
-```
-
-#### 追加要件
-
-中国語の ASR（FunASR がサポート）が必要な場合は、以下をインストールしてください:
-
-```bash
-pip install modelscope torchaudio sentencepiece funasr
 ```
 
 #### FFmpeg

webui.py

@@ -45,14 +45,17 @@ i18n = I18nAuto()
 from scipy.io import wavfile
 from tools.my_utils import load_audio
 from multiprocessing import cpu_count
+
+os.environ['PYTORCH_ENABLE_MPS_FALLBACK'] = '1' # 当遇到mps不支持的步骤时使用cpu
+
 n_cpu=cpu_count()
            
-# 判断是否有能用来训练和加速推理的N卡
 ngpu = torch.cuda.device_count()
 gpu_infos = []
 mem = []
 if_gpu_ok = False
 
+# 判断是否有能用来训练和加速推理的N卡
 if torch.cuda.is_available() or ngpu != 0:
     for i in range(ngpu):
         gpu_name = torch.cuda.get_device_name(i)
@@ -61,6 +64,12 @@ if torch.cuda.is_available() or ngpu != 0:
             if_gpu_ok = True  # 至少有一张能用的N卡
             gpu_infos.append("%s\t%s" % (i, gpu_name))
             mem.append(int(torch.cuda.get_device_properties(i).total_memory/ 1024/ 1024/ 1024+ 0.4))
+# 判断是否支持mps加速
+if torch.backends.mps.is_available():
+    if_gpu_ok = True
+    gpu_infos.append("%s\t%s" % ("0", "Apple GPU"))
+    mem.append(psutil.virtual_memory().total/ 1024 / 1024 / 1024) # 实测使用系统内存作为显存不会爆显存
+
 if if_gpu_ok and len(gpu_infos) > 0:
     gpu_info = "\n".join(gpu_infos)
     default_batch_size = min(mem) // 2