use simple request/reply mode, instead of having multiple workers

tts/gradio_api.py

@@ -33,38 +33,44 @@ def forward_gpu(file_content, wav_path, latent_file, inp_text, time_step, p_w, t
     return wav_bytes
 
 def model_worker(input_queue, output_queue, device_id):
-    while True:
-        task = input_queue.get()
-        inp_audio_path, inp_npy_path, inp_text, infer_timestep, p_w, t_w = task
 
-        if inp_npy_path is None or inp_audio_path is None:
-            output_queue.put(None)
-            raise gr.Error("Please provide .wav and .npy file")
-        if (inp_audio_path.split('/')[-1][:-4] != inp_npy_path.split('/')[-1][:-4]):
-            output_queue.put(None)
-            raise gr.Error(".npy and .wav mismatch")
-        if len(inp_text) > 200:
-            output_queue.put(None)
-            raise gr.Error("input text is too long")
-        
-        try:
-            convert_to_wav(inp_audio_path)
-            wav_path = os.path.splitext(inp_audio_path)[0] + '.wav'
-            cut_wav(wav_path, max_len=24)
-            with open(wav_path, 'rb') as file:
-                file_content = file.read()
-            wav_bytes = forward_gpu(file_content, wav_path, inp_npy_path, inp_text, time_step=infer_timestep, p_w=p_w, t_w=t_w)
-            output_queue.put(wav_bytes)
-        except Exception as e:
-            traceback.print_exc()
-            print(task, str(e))
-            output_queue.put(None)
-            raise gr.Error("Generation failed")
+    task = input_queue.get()
+    inp_audio_path, inp_npy_path, inp_text, infer_timestep, p_w, t_w = task
+
+    if inp_npy_path is None or inp_audio_path is None:
+        output_queue.put(None)
+        raise gr.Error("Please provide .wav and .npy file")
+    if (inp_audio_path.split('/')[-1][:-4] != inp_npy_path.split('/')[-1][:-4]):
+        output_queue.put(None)
+        raise gr.Error(".npy and .wav mismatch")
+    if len(inp_text) > 200:
+        output_queue.put(None)
+        raise gr.Error("input text is too long")
+    
+    try:
+        convert_to_wav(inp_audio_path)
+        wav_path = os.path.splitext(inp_audio_path)[0] + '.wav'
+        cut_wav(wav_path, max_len=24)
+        with open(wav_path, 'rb') as file:
+            file_content = file.read()
+        wav_bytes = forward_gpu(file_content, wav_path, inp_npy_path, inp_text, time_step=infer_timestep, p_w=p_w, t_w=t_w)
+        output_queue.put(wav_bytes)
+    except Exception as e:
+        traceback.print_exc()
+        print(task, str(e))
+        output_queue.put(None)
+        raise gr.Error("Generation failed")
 
 
 def main(inp_audio, inp_npy, inp_text, infer_timestep, p_w, t_w, processes, input_queue, output_queue):
+    input_queue = mp_manager.Queue()
     print("Push task to the inp queue |", inp_audio, inp_npy, inp_text, infer_timestep, p_w, t_w)
     input_queue.put((inp_audio, inp_npy, inp_text, infer_timestep, p_w, t_w))
+
+    output_queue = mp_manager.Queue()
+
+    model_worker(input_queue, output_queue, 0)
+    
     res = output_queue.get()
     if res is not None:
         return res
@@ -78,16 +84,8 @@ if __name__ == '__main__':
     
     num_workers = 1
     devices = [0]
-    input_queue = mp_manager.Queue()
-    output_queue = mp_manager.Queue()
     processes = []
 
-    print("Start open workers")
-    for i in range(num_workers):
-        p = mp.Process(target=model_worker, args=(input_queue, output_queue, i % len(devices) if devices is not None else None))
-        p.start()
-        processes.append(p)
-
     api_interface = gr.Interface(fn=
                                 partial(main, processes=processes, input_queue=input_queue, 
                                         output_queue=output_queue),