Update app.py

app.py

@@ -19,19 +19,18 @@ os.makedirs('./temp', exist_ok=True)
 
 print('\n\n\n')
 print('Loading model...')
-# cpu_pipe = transformers.pipeline(
-#     'text-generation', 
-#     model='dx2102/llama-midi',
-#     torch_dtype='float32',
-#     device='cpu',
-# )
+cpu_pipe = transformers.pipeline(
+    'text-generation', 
+    model='dx2102/llama-midi',
+    torch_dtype='float32',
+    device='cpu',
+)
 gpu_pipe = transformers.pipeline(
     'text-generation', 
     model='dx2102/llama-midi',
     torch_dtype='bfloat16',
     device='cuda:0',
 )
-cpu_pipe = gpu_pipe
 # print devices
 print(f"{gpu_pipe.device = }, {gpu_pipe.model.device = }")
 print(f"{cpu_pipe.device = }, {cpu_pipe.model.device = }")
@@ -58,6 +57,8 @@ example_prefix = '''pitch duration wait velocity instrument
 69 1970 0 20 0
 48 330 350 20 0
 '''
+print('cpu:', cpu_pipe(example_prefix, max_new_tokens=10)[0]['generated_text'])
+print('gpu:', gpu_pipe(example_prefix, max_new_tokens=10)[0]['generated_text'])