ninja中文路径编译补丁支持：BigVGAN fused cuda kernel (#93)

* ninja支持中文路径编译补丁：BigVGAN fused cuda kernel

* 缓存参考音频的Mel

* ninja支持中文路径编译方案2：BigVGAN fused cuda kernel

indextts/BigVGAN/alias_free_activation/cuda/load.py

@@ -14,6 +14,37 @@ Set it to empty stringo avoid recompilation and assign arch flags explicity in e
 os.environ["TORCH_CUDA_ARCH_LIST"] = ""
 
 
+import re
+import shutil
+import tempfile
+
+# 补丁修复：sources 路径含中文字符时，生成 build.ninja 乱码导致编译失败
+# 使用临时目录来规避 ninja 编译失败（比如中文路径）
+def chinese_path_compile_support(sources, buildpath):
+    pattern = re.compile(r'[\u4e00-\u9fff]')  
+    if not bool(pattern.search(str(sources[0].resolve()))):
+        return buildpath # 检测非中文路径跳过
+    # Create build directory
+    resolves = [ item.name for item in sources]
+    ninja_compile_dir = os.path.join(tempfile.gettempdir(), "BigVGAN", "cuda")
+    os.makedirs(ninja_compile_dir, exist_ok=True)
+    new_buildpath = os.path.join(ninja_compile_dir, "build")
+    os.makedirs(new_buildpath, exist_ok=True)
+    print(f"ninja_buildpath: {new_buildpath}")
+    # Copy files to directory
+    sources.clear()
+    current_dir = os.path.dirname(__file__)
+    ALLOWED_EXTENSIONS = {'.py', '.cu', '.cpp', '.h'}
+    for filename in os.listdir(current_dir):
+        item = pathlib.Path(current_dir).joinpath(filename)
+        tar_path = pathlib.Path(ninja_compile_dir).joinpath(item.name)
+        if not item.suffix.lower() in ALLOWED_EXTENSIONS:continue
+        pathlib.Path(shutil.copy2(item, tar_path))
+        if tar_path.name in resolves:sources.append(tar_path)
+    return new_buildpath
+
+
+
 def load():
     # Check if cuda 11 is installed for compute capability 8.0
     cc_flag = []
@@ -58,6 +89,10 @@ def load():
         srcpath / "anti_alias_activation.cpp",
         srcpath / "anti_alias_activation_cuda.cu",
     ]
+    
+    # 兼容方案：ninja 特殊字符路径编译支持处理（比如中文路径）
+    buildpath = chinese_path_compile_support(sources, buildpath)
+    
     anti_alias_activation_cuda = _cpp_extention_load_helper(
         "anti_alias_activation_cuda", sources, extra_cuda_flags
     )

indextts/infer.py

@@ -110,6 +110,9 @@ class IndexTTS:
         self.normalizer = TextNormalizer()
         self.normalizer.load()
         print(">> TextNormalizer loaded")
+        # 缓存参考音频mel：
+        self.cache_audio_prompt = None
+        self.cache_cond_mel = None
 
     def preprocess_text(self, text):
         # chinese_punctuation = "，。！？；：“”‘’（）【】《》"
@@ -186,15 +189,26 @@ class IndexTTS:
         normalized_text = self.preprocess_text(text)
         print(f"normalized text:{normalized_text}")
 
-        audio, sr = torchaudio.load(audio_prompt)
-        audio = torch.mean(audio, dim=0, keepdim=True)
-        if audio.shape[0] > 1:
-            audio = audio[0].unsqueeze(0)
-        audio = torchaudio.transforms.Resample(sr, 24000)(audio)
-        cond_mel = MelSpectrogramFeatures()(audio).to(self.device)
-        cond_mel_frame = cond_mel.shape[-1]
-        if verbose:
-            print(f"cond_mel shape: {cond_mel.shape}", "dtype:", cond_mel.dtype)
+
+        # 如果参考音频改变了，才需要重新生成 cond_mel, 提升速度
+        if self.cache_cond_mel is None or self.cache_audio_prompt != audio_prompt:
+            audio, sr = torchaudio.load(audio_prompt)
+            audio = torch.mean(audio, dim=0, keepdim=True)
+            if audio.shape[0] > 1:
+                audio = audio[0].unsqueeze(0)
+            audio = torchaudio.transforms.Resample(sr, 24000)(audio)
+            cond_mel = MelSpectrogramFeatures()(audio).to(self.device)
+            cond_mel_frame = cond_mel.shape[-1]
+            if verbose:
+                print(f"cond_mel shape: {cond_mel.shape}", "dtype:", cond_mel.dtype)
+            
+            self.cache_audio_prompt = audio_prompt
+            self.cache_cond_mel = cond_mel
+        else:
+            cond_mel = self.cache_cond_mel
+            cond_mel_frame = cond_mel.shape[-1]
+            pass
+        
 
         auto_conditioning = cond_mel