Upload app.py with huggingface_hub

app.py

@@ -13,39 +13,52 @@ if not IS_DUPLICATE:
     print('DEBUG', kokoro.__version__, CUDA_AVAILABLE, misaki.__version__)
 
 CHAR_LIMIT = None if IS_DUPLICATE else 5000
-models = {gpu: KModel().to('cuda' if gpu else 'cpu').eval() for gpu in [False] + ([True] if CUDA_AVAILABLE else [])}
-pipelines = {lang_code: KPipeline(lang_code=lang_code, model=False) for lang_code in 'ab'}
-pipelines['a'].g2p.lexicon.golds['kokoro'] = 'kˈOkəɹO'
-pipelines['b'].g2p.lexicon.golds['kokoro'] = 'kˈQkəɹQ'
-
-@spaces.GPU(duration=30)
-def forward_gpu(ps, ref_s, speed):
-    return models[True](ps, ref_s, speed)
 
+_models = None
+_pipelines = None
 _loaded_voices = set()
+
+def _get_models():
+    global _models
+    if _models is None:
+        _models = {gpu: KModel().to('cuda' if gpu else 'cpu').eval() for gpu in [False] + ([True] if CUDA_AVAILABLE else [])}
+    return _models
+
+def _get_pipelines():
+    global _pipelines
+    if _pipelines is None:
+        _pipelines = {lang_code: KPipeline(lang_code=lang_code, model=False) for lang_code in 'ab'}
+        _pipelines['a'].g2p.lexicon.golds['kokoro'] = 'kˈOkəɹO'
+        _pipelines['b'].g2p.lexicon.golds['kokoro'] = 'kˈQkəɹQ'
+    return _pipelines
+
 def _ensure_voice(voice):
     if voice not in _loaded_voices:
-        pipelines[voice[0]].load_voice(voice)
+        _get_pipelines()[voice[0]].load_voice(voice)
         _loaded_voices.add(voice)
 
+@spaces.GPU(duration=30)
+def forward_gpu(ps, ref_s, speed):
+    return _get_models()[True](ps, ref_s, speed)
+
 def generate_first(text, voice='af_heart', speed=1, use_gpu=CUDA_AVAILABLE):
     text = text if CHAR_LIMIT is None else text.strip()[:CHAR_LIMIT]
-    pipeline = pipelines[voice[0]]
+    pipeline = _get_pipelines()
     _ensure_voice(voice)
-    pack = pipeline.load_voice(voice)
+    pack = pipeline[voice[0]].load_voice(voice)
     use_gpu = use_gpu and CUDA_AVAILABLE
-    for _, ps, _ in pipeline(text, voice, speed):
+    for _, ps, _ in pipeline[voice[0]](text, voice, speed):
         ref_s = pack[len(ps)-1]
         try:
             if use_gpu:
                 audio = forward_gpu(ps, ref_s, speed)
             else:
-                audio = models[False](ps, ref_s, speed)
+                audio = _get_models()[False](ps, ref_s, speed)
         except gr.exceptions.Error as e:
             if use_gpu:
                 gr.Warning(str(e))
                 gr.Info('Retrying with CPU. To avoid this error, change Hardware to CPU.')
-                audio = models[False](ps, ref_s, speed)
+                audio = _get_models()[False](ps, ref_s, speed)
             else:
                 raise gr.Error(e)
         return (24000, audio.numpy()), ps
@@ -55,30 +68,30 @@ def predict(text, voice='af_heart', speed=1):
     return generate_first(text, voice, speed, use_gpu=False)[0]
 
 def tokenize_first(text, voice='af_heart'):
-    pipeline = pipelines[voice[0]]
-    for _, ps, _ in pipeline(text, voice):
+    pipeline = _get_pipelines()
+    for _, ps, _ in pipeline[voice[0]](text, voice):
         return ps
     return ''
 
 def generate_all(text, voice='af_heart', speed=1, use_gpu=CUDA_AVAILABLE):
     text = text if CHAR_LIMIT is None else text.strip()[:CHAR_LIMIT]
-    pipeline = pipelines[voice[0]]
+    pipeline = _get_pipelines()
     _ensure_voice(voice)
-    pack = pipeline.load_voice(voice)
+    pack = pipeline[voice[0]].load_voice(voice)
     use_gpu = use_gpu and CUDA_AVAILABLE
     first = True
-    for _, ps, _ in pipeline(text, voice, speed):
+    for _, ps, _ in pipeline[voice[0]](text, voice, speed):
         ref_s = pack[len(ps)-1]
         try:
             if use_gpu:
                 audio = forward_gpu(ps, ref_s, speed)
             else:
-                audio = models[False](ps, ref_s, speed)
+                audio = _get_models()[False](ps, ref_s, speed)
         except gr.exceptions.Error as e:
             if use_gpu:
                 gr.Warning(str(e))
                 gr.Info('Switching to CPU')
-                audio = models[False](ps, ref_s, speed)
+                audio = _get_models()[False](ps, ref_s, speed)
             else:
                 raise gr.Error(e)
         yield 24000, audio.numpy()