Merge b122988be0a4c47d81a731dead738a0a75233d37 into 13055fa56994e75a7152c176047c56c62bbeede4

GPT_SoVITS/TTS_infer_pack/TTS.py

@@ -287,8 +287,9 @@ class TTS_Config:
             configs: dict = self._load_configs(self.configs_path)
 
         assert isinstance(configs, dict)
-        version = configs.get("version", "v2").lower()
-        assert version in ["v1", "v2", "v3", "v4"]
+        version = "v2"
+        if "custom" in configs and configs["custom"]["version"].lower() in ["v1", "v2", "v3", "v4"]:
+            version = configs["custom"]["version"].lower()
         self.default_configs[version] = configs.get(version, self.default_configs[version])
         self.configs: dict = configs.get("custom", deepcopy(self.default_configs[version]))
 

GPT_SoVITS/inference_webui.py

@@ -91,6 +91,8 @@ infer_ttswebui = os.environ.get("infer_ttswebui", 9872)
 infer_ttswebui = int(infer_ttswebui)
 is_share = os.environ.get("is_share", "False")
 is_share = eval(is_share)
+local_mode = os.environ.get("local_mode", "False")
+local_mode = eval(local_mode)
 if "_CUDA_VISIBLE_DEVICES" in os.environ:
     os.environ["CUDA_VISIBLE_DEVICES"] = os.environ["_CUDA_VISIBLE_DEVICES"]
 is_half = eval(os.environ.get("is_half", "True")) and torch.cuda.is_available()
@@ -1273,7 +1275,7 @@ with gr.Blocks(title="GPT-SoVITS WebUI") as app:
 
 if __name__ == "__main__":
     app.queue().launch(  # concurrency_count=511, max_size=1022
-        server_name="0.0.0.0",
+        server_name="127.0.0.1" if local_mode else "0.0.0.0",
         inbrowser=True,
         share=is_share,
         server_port=infer_ttswebui,

api.py

@@ -176,9 +176,9 @@ import subprocess
 
 class DefaultRefer:
     def __init__(self, path, text, language):
-        self.path = args.default_refer_path
-        self.text = args.default_refer_text
-        self.language = args.default_refer_language
+        self.path = path
+        self.text = text
+        self.language = language
 
     def is_ready(self) -> bool:
         return is_full(self.path, self.text, self.language)
@@ -294,11 +294,13 @@ from process_ckpt import get_sovits_version_from_path_fast, load_sovits_new
 
 def get_sovits_weights(sovits_path):
     path_sovits_v3 = "GPT_SoVITS/pretrained_models/s2Gv3.pth"
-    is_exist_s2gv3 = os.path.exists(path_sovits_v3)
+    path_sovits_v4 = "GPT_SoVITS/pretrained_models/gsv-v4-pretrained/s2Gv4.pth"
 
     version, model_version, if_lora_v3 = get_sovits_version_from_path_fast(sovits_path)
-    if if_lora_v3 == True and is_exist_s2gv3 == False:
+    if if_lora_v3 == True and not os.path.exists(path_sovits_v3):
         logger.info("SoVITS V3 底模缺失，无法加载相应 LoRA 权重")
+    if model_version == "v4" and not os.path.exists(path_sovits_v4):
+        logger.info("SoVITS V4 底模缺失，无法加载相应 LoRA 权重")
 
     dict_s2 = load_sovits_new(sovits_path)
     hps = dict_s2["config"]
@@ -310,12 +312,13 @@ def get_sovits_weights(sovits_path):
         hps.model.version = "v1"
     else:
         hps.model.version = "v2"
-
     if model_version == "v3":
         hps.model.version = "v3"
+    if model_version == "v4":
+        hps.model.version = "v4"
 
     model_params_dict = vars(hps.model)
-    if model_version != "v3":
+    if model_version != "v3" and model_version != "v4":
         vq_model = SynthesizerTrn(
             hps.data.filter_length // 2 + 1,
             hps.train.segment_size // hps.data.hop_length,
@@ -342,10 +345,13 @@ def get_sovits_weights(sovits_path):
     else:
         vq_model = vq_model.to(device)
     vq_model.eval()
-    if if_lora_v3 == False:
+    if if_lora_v3 == False or model_version != "v4":
         vq_model.load_state_dict(dict_s2["weight"], strict=False)
     else:
-        vq_model.load_state_dict(load_sovits_new(path_sovits_v3)["weight"], strict=False)
+        if model_version == "v4":
+            vq_model.load_state_dict(load_sovits_new(path_sovits_v4)["weight"], strict=False)
+        else:
+            vq_model.load_state_dict(load_sovits_new(path_sovits_v3)["weight"], strict=False)
         lora_rank = dict_s2["lora_rank"]
         lora_config = LoraConfig(
             target_modules=["to_k", "to_q", "to_v", "to_out.0"],
@@ -394,13 +400,11 @@ def change_gpt_sovits_weights(gpt_path, sovits_path):
     try:
         gpt = get_gpt_weights(gpt_path)
         sovits = get_sovits_weights(sovits_path)
+        speaker_list["default"] = Speaker(name="default", gpt=gpt, sovits=sovits)
+        return JSONResponse({"code": 0, "message": "Success"}, status_code=200)
     except Exception as e:
         return JSONResponse({"code": 400, "message": str(e)}, status_code=400)
-
-    speaker_list["default"] = Speaker(name="default", gpt=gpt, sovits=sovits)
-    return JSONResponse({"code": 0, "message": "Success"}, status_code=200)
-
-
+    
 def get_bert_feature(text, word2ph):
     with torch.no_grad():
         inputs = tokenizer(text, return_tensors="pt")
@@ -759,7 +763,7 @@ def get_tts_wav(
         prompt_semantic = codes[0, 0]
         prompt = prompt_semantic.unsqueeze(0).to(device)
 
-        if version != "v3":
+        if version != "v3" and version != "v4":
             refers = []
             if inp_refs:
                 for path in inp_refs:
@@ -811,7 +815,7 @@ def get_tts_wav(
             pred_semantic = pred_semantic[:, -idx:].unsqueeze(0)
         t3 = ttime()
 
-        if version != "v3":
+        if version != "v3" and version != "v4":
             audio = (
                 vq_model.decode(pred_semantic, torch.LongTensor(phones2).to(device).unsqueeze(0), refers, speed=speed)
                 .detach()
@@ -880,7 +884,7 @@ def get_tts_wav(
         audio_opt = np.concatenate(audio_opt, 0)
         t4 = ttime()
 
-        sr = hps.data.sampling_rate if version != "v3" else 24000
+        sr = hps.data.sampling_rate if version != "v3" and version != "v4" else 24000
         if if_sr and sr == 24000:
             audio_opt = torch.from_numpy(audio_opt).float().to(device)
             audio_opt, sr = audio_sr(audio_opt.unsqueeze(0), sr)
@@ -901,7 +905,7 @@ def get_tts_wav(
     if not stream_mode == "normal":
         if media_type == "wav":
             sr = 48000 if if_sr else 24000
-            sr = hps.data.sampling_rate if version != "v3" else sr
+            sr = hps.data.sampling_rate if version != "v3" and version != "v4" else sr
             audio_bytes = pack_wav(audio_bytes, sr)
         yield audio_bytes.getvalue()
 

config.py

@@ -30,6 +30,9 @@ webui_port_subfix = 9871
 
 api_port = 9880
 
+# 设置为True可启用本地模式，该模式只允许本机访问，避免出现潜在安全问题。默认为False。
+local_mode = False
+
 if infer_device == "cuda":
     gpu_name = torch.cuda.get_device_name(0)
     if (

tools/subfix_webui.py

@@ -298,6 +298,7 @@ if __name__ == "__main__":
     parser.add_argument("--json_key_text", default="text", help="the text key name in json, Default: text")
     parser.add_argument("--json_key_path", default="wav_path", help="the path key name in json, Default: wav_path")
     parser.add_argument("--g_batch", default=10, help="max number g_batch wav to display, Default: 10")
+    parser.add_argument("--local_mode", action="store_true", help="enable local mode (bind to 127.0.0.1)")
 
     args = parser.parse_args()
 
@@ -407,7 +408,7 @@ if __name__ == "__main__":
         )
 
     demo.launch(
-        server_name="0.0.0.0",
+        server_name="127.0.0.1" if args.local_mode else "0.0.0.0",
         inbrowser=True,
         # quiet=True,
         share=eval(args.is_share),

tools/uvr5/webui.py

@@ -32,6 +32,7 @@ device = sys.argv[1]
 is_half = eval(sys.argv[2])
 webui_port_uvr5 = int(sys.argv[3])
 is_share = eval(sys.argv[4])
+local_mode = sys.argv[5].lower() == 'true' if len(sys.argv) > 5 else False
 
 
 def html_left(text, label="p"):
@@ -220,7 +221,7 @@ with gr.Blocks(title="UVR5 WebUI") as app:
                     api_name="uvr_convert",
                 )
 app.queue().launch(  # concurrency_count=511, max_size=1022
-    server_name="0.0.0.0",
+    server_name="127.0.0.1" if local_mode else "0.0.0.0",
     inbrowser=True,
     share=is_share,
     server_port=webui_port_uvr5,

webui.py

@@ -74,6 +74,7 @@ from config import (
     webui_port_main,
     webui_port_subfix,
     webui_port_uvr5,
+    local_mode,
 )
 from tools import my_utils
 from tools.i18n.i18n import I18nAuto, scan_language_list
@@ -388,6 +389,8 @@ def change_label(path_list):
             webui_port_subfix,
             is_share,
         )
+        if local_mode:
+            cmd += " --local_mode"
         yield (
             process_info(process_name_subfix, "opened"),
             {"__type__": "update", "visible": False},
@@ -412,6 +415,8 @@ def change_uvr5():
     global p_uvr5
     if p_uvr5 is None:
         cmd = '"%s" tools/uvr5/webui.py "%s" %s %s %s' % (python_exec, infer_device, is_half, webui_port_uvr5, is_share)
+        if local_mode:
+            cmd += " True"
         yield (
             process_info(process_name_uvr5, "opened"),
             {"__type__": "update", "visible": False},
@@ -450,6 +455,7 @@ def change_tts_inference(bert_path, cnhubert_base_path, gpu_number, gpt_path, so
         os.environ["is_half"] = str(is_half)
         os.environ["infer_ttswebui"] = str(webui_port_infer_tts)
         os.environ["is_share"] = str(is_share)
+        os.environ["local_mode"] = str(local_mode)
         yield (
             process_info(process_name_tts, "opened"),
             {"__type__": "update", "visible": False},
@@ -1955,7 +1961,7 @@ with gr.Blocks(title="GPT-SoVITS WebUI") as app:
             gr.Markdown(value=i18n("施工中，请静候佳音"))
 
     app.queue().launch(  # concurrency_count=511, max_size=1022
-        server_name="0.0.0.0",
+        server_name="127.0.0.1" if local_mode else "0.0.0.0",
         inbrowser=True,
         share=is_share,
         server_port=webui_port_main,