Update README.md

Co-authored-by: tzrain <tz_rain@foxmail.com>

.pre-commit-config.yaml

@@ -8,7 +8,7 @@ repos:
     # Run the linter.
     - id: ruff
       types_or: [ python, pyi ]
-      args: [ --fix ]
+      args: [ --fix , "--exit-zero" ]
     # Run the formatter.
     - id: ruff-format
       types_or: [ python, pyi ]

GPT_SoVITS/TTS_infer_pack/TTS.py

@@ -1407,7 +1407,10 @@ class TTS:
     ):
         prompt_semantic_tokens = self.prompt_cache["prompt_semantic"].unsqueeze(0).unsqueeze(0).to(self.configs.device)
         prompt_phones = torch.LongTensor(self.prompt_cache["phones"]).unsqueeze(0).to(self.configs.device)
-        refer_audio_spec = self.prompt_cache["refer_spec"][0].to(dtype=self.precision, device=self.configs.device)
+        raw_entry = self.prompt_cache["refer_spec"][0]
+        if isinstance(raw_entry, tuple):
+            raw_entry = raw_entry[0]
+        refer_audio_spec = raw_entry.to(dtype=self.precision,device=self.configs.device)
 
         fea_ref, ge = self.vits_model.decode_encp(prompt_semantic_tokens, prompt_phones, refer_audio_spec)
         ref_audio: torch.Tensor = self.prompt_cache["raw_audio"]
@@ -1474,7 +1477,10 @@ class TTS:
     ) -> List[torch.Tensor]:
         prompt_semantic_tokens = self.prompt_cache["prompt_semantic"].unsqueeze(0).unsqueeze(0).to(self.configs.device)
         prompt_phones = torch.LongTensor(self.prompt_cache["phones"]).unsqueeze(0).to(self.configs.device)
-        refer_audio_spec = self.prompt_cache["refer_spec"][0].to(dtype=self.precision, device=self.configs.device)
+        raw_entry = self.prompt_cache["refer_spec"][0]
+        if isinstance(raw_entry, tuple):
+            raw_entry = raw_entry[0]
+        refer_audio_spec = raw_entry.to(dtype=self.precision,device=self.configs.device)
 
         fea_ref, ge = self.vits_model.decode_encp(prompt_semantic_tokens, prompt_phones, refer_audio_spec)
         ref_audio: torch.Tensor = self.prompt_cache["raw_audio"]

GPT_SoVITS/inference_webui.py

@@ -202,16 +202,6 @@ if is_half == True:
 else:
     ssl_model = ssl_model.to(device)
 
-resample_transform_dict = {}
-
-
-def resample(audio_tensor, sr0, sr1):
-    global resample_transform_dict
-    key = "%s-%s" % (sr0, sr1)
-    if key not in resample_transform_dict:
-        resample_transform_dict[key] = torchaudio.transforms.Resample(sr0, sr1).to(device)
-    return resample_transform_dict[key](audio_tensor)
-
 
 ###todo:put them to process_ckpt and modify my_save func (save sovits weights), gpt save weights use my_save in process_ckpt
 # symbol_version-model_version-if_lora_v3
@@ -899,7 +889,7 @@ def get_tts_wav(
                 ref_audio = ref_audio.mean(0).unsqueeze(0)
             tgt_sr = 24000 if model_version == "v3" else 32000
             if sr != tgt_sr:
-                ref_audio = resample(ref_audio, sr, tgt_sr)
+                ref_audio = resample(ref_audio, sr, tgt_sr,device)
             # print("ref_audio",ref_audio.abs().mean())
             mel2 = mel_fn(ref_audio) if model_version == "v3" else mel_fn_v4(ref_audio)
             mel2 = norm_spec(mel2)

GPT_SoVITS/text/g2pw/onnx_api.py

@@ -23,8 +23,8 @@ from .utils import load_config
 onnxruntime.set_default_logger_severity(3)
 try:
     onnxruntime.preload_dlls()
-except:
-    traceback.print_exc()
+except:pass
+    #traceback.print_exc()
 warnings.filterwarnings("ignore")
 
 model_version = "1.1"

README.md

@@ -328,6 +328,23 @@ Use v4 from v1/v2/v3 environment:
 
 3. Download v4 pretrained models (gsv-v4-pretrained/s2v4.ckpt, and gsv-v4-pretrained/vocoder.pth) from [huggingface](https://huggingface.co/lj1995/GPT-SoVITS/tree/main) and put them into `GPT_SoVITS/pretrained_models`.
 
+## V2Pro Release Notes
+
+New Features:
+
+1. Slightly higher VRAM usage than v2, surpassing v4's performance, with v2's hardware cost and speed.
+   [more details](https://github.com/RVC-Boss/GPT-SoVITS/wiki/GPT%E2%80%90SoVITS%E2%80%90features-(%E5%90%84%E7%89%88%E6%9C%AC%E7%89%B9%E6%80%A7))
+
+2.v1/v2 and the v2Pro series share the same characteristics, while v3/v4 have similar features. For training sets with average audio quality, v1/v2/v2Pro can deliver decent results, but v3/v4 cannot. Additionally, the synthesized tone and timebre of v3/v4 lean more toward the reference audio rather than the overall training set.
+
+Use v2Pro from v1/v2/v3/v4 environment:
+
+1. `pip install -r requirements.txt` to update some packages
+
+2. Clone the latest codes from github.
+
+3. Download v2Pro pretrained models (v2Pro/s2Dv2Pro.pth, v2Pro/s2Gv2Pro.pth, v2Pro/s2Dv2ProPlus.pth, v2Pro/s2Gv2ProPlus.pth, and sv/pretrained_eres2netv2w24s4ep4.ckpt) from [huggingface](https://huggingface.co/lj1995/GPT-SoVITS/tree/main) and put them into `GPT_SoVITS/pretrained_models`.
+
 ## Todo List
 
 - [x] **High Priority:**

config.py

@@ -1,30 +1,32 @@
-import sys
 import os
+import re
+import sys
 
-import torch,re
+import torch
+
+from tools.i18n.i18n import I18nAuto
 
-from tools.i18n.i18n import I18nAuto, scan_language_list
 i18n = I18nAuto(language=os.environ.get("language", "Auto"))
 
 
 pretrained_sovits_name = {
-    "v1":"GPT_SoVITS/pretrained_models/s2G488k.pth",
-    "v2":"GPT_SoVITS/pretrained_models/gsv-v2final-pretrained/s2G2333k.pth",
-    "v3":"GPT_SoVITS/pretrained_models/s2Gv3.pth",###v3v4还要检查vocoder，算了。。。
-    "v4":"GPT_SoVITS/pretrained_models/gsv-v4-pretrained/s2Gv4.pth",
-    "v2Pro":"GPT_SoVITS/pretrained_models/v2Pro/s2Gv2Pro.pth",
-    "v2ProPlus":"GPT_SoVITS/pretrained_models/v2Pro/s2Gv2ProPlus.pth",
+    "v1": "GPT_SoVITS/pretrained_models/s2G488k.pth",
+    "v2": "GPT_SoVITS/pretrained_models/gsv-v2final-pretrained/s2G2333k.pth",
+    "v3": "GPT_SoVITS/pretrained_models/s2Gv3.pth",  ###v3v4还要检查vocoder，算了。。。
+    "v4": "GPT_SoVITS/pretrained_models/gsv-v4-pretrained/s2Gv4.pth",
+    "v2Pro": "GPT_SoVITS/pretrained_models/v2Pro/s2Gv2Pro.pth",
+    "v2ProPlus": "GPT_SoVITS/pretrained_models/v2Pro/s2Gv2ProPlus.pth",
 }
 
 pretrained_gpt_name = {
-    "v1":"GPT_SoVITS/pretrained_models/s1bert25hz-2kh-longer-epoch=68e-step=50232.ckpt",
-    "v2":"GPT_SoVITS/pretrained_models/gsv-v2final-pretrained/s1bert25hz-5kh-longer-epoch=12-step=369668.ckpt",
-    "v3":"GPT_SoVITS/pretrained_models/s1v3.ckpt",
-    "v4":"GPT_SoVITS/pretrained_models/s1v3.ckpt",
-    "v2Pro":"GPT_SoVITS/pretrained_models/s1v3.ckpt",
-    "v2ProPlus":"GPT_SoVITS/pretrained_models/s1v3.ckpt",
+    "v1": "GPT_SoVITS/pretrained_models/s1bert25hz-2kh-longer-epoch=68e-step=50232.ckpt",
+    "v2": "GPT_SoVITS/pretrained_models/gsv-v2final-pretrained/s1bert25hz-5kh-longer-epoch=12-step=369668.ckpt",
+    "v3": "GPT_SoVITS/pretrained_models/s1v3.ckpt",
+    "v4": "GPT_SoVITS/pretrained_models/s1v3.ckpt",
+    "v2Pro": "GPT_SoVITS/pretrained_models/s1v3.ckpt",
+    "v2ProPlus": "GPT_SoVITS/pretrained_models/s1v3.ckpt",
 }
-name2sovits_path={
+name2sovits_path = {
     # i18n("不训练直接推v1底模！"): "GPT_SoVITS/pretrained_models/s2G488k.pth",
     i18n("不训练直接推v2底模！"): "GPT_SoVITS/pretrained_models/gsv-v2final-pretrained/s2G2333k.pth",
     # i18n("不训练直接推v3底模！"): "GPT_SoVITS/pretrained_models/s2Gv3.pth",
@@ -32,29 +34,47 @@ name2sovits_path={
     i18n("不训练直接推v2Pro底模！"): "GPT_SoVITS/pretrained_models/v2Pro/s2Gv2Pro.pth",
     i18n("不训练直接推v2ProPlus底模！"): "GPT_SoVITS/pretrained_models/v2Pro/s2Gv2ProPlus.pth",
 }
-name2gpt_path={
+name2gpt_path = {
     # i18n("不训练直接推v1底模！"):"GPT_SoVITS/pretrained_models/s1bert25hz-2kh-longer-epoch=68e-step=50232.ckpt",
-    i18n("不训练直接推v2底模！"):"GPT_SoVITS/pretrained_models/gsv-v2final-pretrained/s1bert25hz-5kh-longer-epoch=12-step=369668.ckpt",
-    i18n("不训练直接推v3底模！"):"GPT_SoVITS/pretrained_models/s1v3.ckpt",
+    i18n(
+        "不训练直接推v2底模！"
+    ): "GPT_SoVITS/pretrained_models/gsv-v2final-pretrained/s1bert25hz-5kh-longer-epoch=12-step=369668.ckpt",
+    i18n("不训练直接推v3底模！"): "GPT_SoVITS/pretrained_models/s1v3.ckpt",
 }
-SoVITS_weight_root = ["SoVITS_weights", "SoVITS_weights_v2", "SoVITS_weights_v3", "SoVITS_weights_v4", "SoVITS_weights_v2Pro", "SoVITS_weights_v2ProPlus"]
-GPT_weight_root = ["GPT_weights", "GPT_weights_v2", "GPT_weights_v3", "GPT_weights_v4", "GPT_weights_v2Pro", "GPT_weights_v2ProPlus"]
-SoVITS_weight_version2root={
-    "v1":"SoVITS_weights",
-    "v2":"SoVITS_weights_v2",
-    "v3":"SoVITS_weights_v3",
-    "v4":"SoVITS_weights_v4",
-    "v2Pro":"SoVITS_weights_v2Pro",
-    "v2ProPlus":"SoVITS_weights_v2ProPlus",
+SoVITS_weight_root = [
+    "SoVITS_weights",
+    "SoVITS_weights_v2",
+    "SoVITS_weights_v3",
+    "SoVITS_weights_v4",
+    "SoVITS_weights_v2Pro",
+    "SoVITS_weights_v2ProPlus",
+]
+GPT_weight_root = [
+    "GPT_weights",
+    "GPT_weights_v2",
+    "GPT_weights_v3",
+    "GPT_weights_v4",
+    "GPT_weights_v2Pro",
+    "GPT_weights_v2ProPlus",
+]
+SoVITS_weight_version2root = {
+    "v1": "SoVITS_weights",
+    "v2": "SoVITS_weights_v2",
+    "v3": "SoVITS_weights_v3",
+    "v4": "SoVITS_weights_v4",
+    "v2Pro": "SoVITS_weights_v2Pro",
+    "v2ProPlus": "SoVITS_weights_v2ProPlus",
 }
-GPT_weight_version2root={
-    "v1":"GPT_weights",
-    "v2":"GPT_weights_v2",
-    "v3":"GPT_weights_v3",
-    "v4":"GPT_weights_v4",
-    "v2Pro":"GPT_weights_v2Pro",
-    "v2ProPlus":"GPT_weights_v2ProPlus",
+GPT_weight_version2root = {
+    "v1": "GPT_weights",
+    "v2": "GPT_weights_v2",
+    "v3": "GPT_weights_v3",
+    "v4": "GPT_weights_v4",
+    "v2Pro": "GPT_weights_v2Pro",
+    "v2ProPlus": "GPT_weights_v2ProPlus",
 }
+
+
 def custom_sort_key(s):
     # 使用正则表达式提取字符串中的数字部分和非数字部分
     parts = re.split("(\d+)", s)
@@ -62,27 +82,37 @@ def custom_sort_key(s):
     parts = [int(part) if part.isdigit() else part for part in parts]
     return parts
 
+
 def get_weights_names():
     SoVITS_names = []
     for key in name2sovits_path:
-        if os.path.exists(name2sovits_path[key]):SoVITS_names.append(key)
+        if os.path.exists(name2sovits_path[key]):
+            SoVITS_names.append(key)
     for path in SoVITS_weight_root:
-        if not os.path.exists(path):continue
+        if not os.path.exists(path):
+            continue
         for name in os.listdir(path):
             if name.endswith(".pth"):
                 SoVITS_names.append("%s/%s" % (path, name))
+    if not SoVITS_names:
+        SoVITS_names = [""]
     GPT_names = []
     for key in name2gpt_path:
-        if os.path.exists(name2gpt_path[key]):GPT_names.append(key)
+        if os.path.exists(name2gpt_path[key]):
+            GPT_names.append(key)
     for path in GPT_weight_root:
-        if not os.path.exists(path):continue
+        if not os.path.exists(path):
+            continue
         for name in os.listdir(path):
             if name.endswith(".ckpt"):
                 GPT_names.append("%s/%s" % (path, name))
-    SoVITS_names=sorted(SoVITS_names, key=custom_sort_key)
-    GPT_names=sorted(GPT_names, key=custom_sort_key)
+    SoVITS_names = sorted(SoVITS_names, key=custom_sort_key)
+    GPT_names = sorted(GPT_names, key=custom_sort_key)
+    if not GPT_names:
+        GPT_names = [""]
     return SoVITS_names, GPT_names
 
+
 def change_choices():
     SoVITS_names, GPT_names = get_weights_names()
     return {"choices": SoVITS_names, "__type__": "update"}, {
@@ -106,10 +136,6 @@ pretrained_gpt_path = "GPT_SoVITS/pretrained_models/s1bert25hz-2kh-longer-epoch=
 
 exp_root = "logs"
 python_exec = sys.executable or "python"
-if torch.cuda.is_available():
-    infer_device = "cuda"
-else:
-    infer_device = "cpu"
 
 webui_port_main = 9874
 webui_port_uvr5 = 9873
@@ -118,20 +144,55 @@ webui_port_subfix = 9871
 
 api_port = 9880
 
-if infer_device == "cuda":
-    gpu_name = torch.cuda.get_device_name(0)
-    if (
-        ("16" in gpu_name and "V100" not in gpu_name.upper())
-        or "P40" in gpu_name.upper()
-        or "P10" in gpu_name.upper()
-        or "1060" in gpu_name
-        or "1070" in gpu_name
-        or "1080" in gpu_name
-    ):
-        is_half = False
 
-if infer_device == "cpu":
-    is_half = False
+def get_device_dtype_sm(idx: int) -> tuple[torch.device, torch.dtype, float, float]:
+    cpu = torch.device("cpu")
+    cuda = torch.device(f"cuda:{idx}")
+    if not torch.cuda.is_available():
+        return cpu, torch.float32, 0.0, 0.0
+    device_idx = idx
+    capability = torch.cuda.get_device_capability(device_idx)
+    name = torch.cuda.get_device_name(device_idx)
+    mem_bytes = torch.cuda.get_device_properties(device_idx).total_memory
+    mem_gb = mem_bytes / (1024**3) + 0.4
+    major, minor = capability
+    sm_version = major + minor / 10.0
+    is_16_series = bool(re.search(r"16\d{2}", name))
+    if mem_gb < 4:
+        return cpu, torch.float32, 0.0, 0.0
+    if (sm_version >= 7.0 and sm_version != 7.5) or (5.3 <= sm_version <= 6.0):
+        if is_16_series and sm_version == 7.5:
+            return cuda, torch.float32, sm_version, mem_gb  # 16系卡除外
+        else:
+            return cuda, torch.float16, sm_version, mem_gb
+    return cpu, torch.float32, 0.0, 0.0
+
+
+IS_GPU = True
+GPU_INFOS: list[str] = []
+GPU_INDEX: set[int] = set()
+GPU_COUNT = torch.cuda.device_count()
+CPU_INFO: str = "0\tCPU " + i18n("CPU训练,较慢")
+tmp: list[tuple[torch.device, torch.dtype, float, float]] = []
+memset: set[float] = set()
+
+for i in range(max(GPU_COUNT, 1)):
+    tmp.append(get_device_dtype_sm(i))
+
+for j in tmp:
+    device = j[0]
+    memset.add(j[3])
+    if device.type != "cpu":
+        GPU_INFOS.append(f"{device.index}\t{torch.cuda.get_device_name(device.index)}")
+        GPU_INDEX.add(device.index)
+
+if not GPU_INFOS:
+    IS_GPU = False
+    GPU_INFOS.append(CPU_INFO)
+    GPU_INDEX.add(0)
+
+infer_device = max(tmp, key=lambda x: (x[2], x[3]))[0]
+is_half = any(dtype == torch.float16 for _, dtype, _, _ in tmp)
 
 
 class Config:

docs/cn/Changelog_CN.md

@@ -386,5 +386,11 @@
 - 2025.06.04 [Commit#b7c0c5ca](https://github.com/RVC-Boss/GPT-SoVITS/commit/b7c0c5ca878bcdd419fd86bf80dba431a6653356)~[Commit#298ebb03](https://github.com/RVC-Boss/GPT-SoVITS/commit/298ebb03c5a719388527ae6a586c7ea960344e70): **新增 GPT-SoVITS V2Pro 系列模型**.
   - 类型: 新功能
   - 提交: RVC-Boss
+- 2025.06.05 https://github.com/RVC-Boss/GPT-SoVITS/pull/2426: config/inference_webui初始化bug修复.
+  - 类型: 修复
+  - 提交: SapphireLab
+- 2025.06.05 https://github.com/RVC-Boss/GPT-SoVITS/pull/2427: 优化精度自动检测逻辑;给webui前端界面模块增加可收缩式支持.
+  - 类型: 新功能
+  - 提交: XXXXRT666