add load model outside a function

app.py

@@ -20,6 +20,7 @@ vq_model = RQBottleneckTransformer.load_model(
         "whisper-vq-stoks-medium-en+pl-fixed.model"
     ).to(device)
 vq_model.ensure_whisper(device)
+
 @spaces.GPU
 def audio_to_sound_tokens_whisperspeech(audio_path):
     wav, sr = torchaudio.load(audio_path)
@@ -31,6 +32,7 @@ def audio_to_sound_tokens_whisperspeech(audio_path):
     
     result = ''.join(f'<|sound_{num:04d}|>' for num in codes)
     return f'<|sound_start|>{result}<|sound_end|>'
+
 @spaces.GPU
 def audio_to_sound_tokens_whisperspeech_transcribe(audio_path):
     wav, sr = torchaudio.load(audio_path)
@@ -42,45 +44,28 @@ def audio_to_sound_tokens_whisperspeech_transcribe(audio_path):
     
     result = ''.join(f'<|sound_{num:04d}|>' for num in codes)
     return f'<|reserved_special_token_69|><|sound_start|>{result}<|sound_end|>'
-def audio_to_sound_tokens(audio_path, target_bandwidth=1.5, device="cuda"):
-    model = EncodecModel.encodec_model_24khz()
-    model.set_target_bandwidth(target_bandwidth)
-    model.to(device)
-    
-    wav, sr = torchaudio.load(audio_path)
-    wav = convert_audio(wav, sr, model.sample_rate, model.channels)
-    wav = wav.unsqueeze(0).to(device)
-    
-    with torch.no_grad():
-        encoded_frames = model.encode(wav)
-        codes = torch.cat([encoded[0] for encoded in encoded_frames], dim=-1)
-    
-    audio_code1, audio_code2 = codes[0][0], codes[0][1]
-    flatten_tokens = torch.stack((audio_code1, audio_code2), dim=1).flatten().tolist()
-    result = ''.join(f'<|sound_{num:04d}|>' for num in flatten_tokens)
-    return f'<|sound_start|>{result}<|sound_end|>'
-@spaces.GPU
-def setup_pipeline(model_path, use_4bit=False, use_8bit=False):
-    tokenizer = AutoTokenizer.from_pretrained(model_path)
-    model_kwargs = {"device_map": "auto"}
-    if use_8bit:
-        model_kwargs["quantization_config"] = BitsAndBytesConfig(
-            load_in_8bit=True,
-            llm_int8_enable_fp32_cpu_offload=False,
-            llm_int8_has_fp16_weight=False,
-        )
-    else:
-        model_kwargs["torch_dtype"] = torch.bfloat16
-    model = AutoModelForCausalLM.from_pretrained(model_path, **model_kwargs)
-    return pipeline("text-generation", model=model, tokenizer=tokenizer)
 
 tts = TTSProcessor(device)
+use_8bit = False    
 llm_path = "homebrewltd/Llama3.1-s-instruct-2024-08-19-epoch-3"
-pipe = setup_pipeline(llm_path, use_8bit=False)
+tokenizer = AutoTokenizer.from_pretrained(llm_path)
+model_kwargs = {"device_map": "auto"}
+if use_8bit:
+    model_kwargs["quantization_config"] = BitsAndBytesConfig(
+        load_in_8bit=True,
+        llm_int8_enable_fp32_cpu_offload=False,
+        llm_int8_has_fp16_weight=False,
+    )
+else:
+    model_kwargs["torch_dtype"] = torch.bfloat16
+model = AutoModelForCausalLM.from_pretrained(llm_path, **model_kwargs)
+pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)
 tokenizer = pipe.tokenizer
 model = pipe.model
 # print(tokenizer.encode("<|sound_0001|>", add_special_tokens=False))# return the audio tensor
 # print(tokenizer.eos_token)
+
+
 @spaces.GPU
 def text_to_audio_file(text):
     # gen a random id for the audio file
@@ -96,6 +81,8 @@ def text_to_audio_file(text):
     # torchaudio.save(temp_file, audio.cpu(), sample_rate=24000)
     print(f"Saved audio to {temp_file}")
     return temp_file
+
+
 @spaces.GPU
 def process_input(input_type, text_input=None, audio_file=None):
     # if input_type == "text":
@@ -106,6 +93,7 @@ def process_input(input_type, text_input=None, audio_file=None):
     
     # if input_type == "text":
     #     os.remove(audio_file) 
+        
 @spaces.GPU
 def process_transcribe_input(input_type, text_input=None, audio_file=None):
     # if input_type == "text":
@@ -124,6 +112,7 @@ class StopOnTokens(StoppingCriteria):
             if input_ids[0][-1] == stop_id:
                 return True
         return False
+    
 @spaces.GPU
 def process_audio(audio_file, transcript=False):
     if audio_file is None: