update

app.py

@@ -1,71 +1,97 @@
+import spaces
+import os
+import torch
+from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
+import gradio as gr
 
-from transformers import AutoTokenizer, AutoModelForCausalLM
 
-# Load the model and tokenizer
-model_name = "Qwen/Qwen1.5-0.5B-Chat"
-tokenizer = AutoTokenizer.from_pretrained(model_name)
-model = AutoModelForCausalLM.from_pretrained(model_name)
+huggingface_token = os.getenv("HUGGINGFACE_TOKEN")
+if not huggingface_token:
+    pass
+    print("no HUGGINGFACE_TOKEN if you need set secret ")
+    #raise ValueError("HUGGINGFACE_TOKEN environment variable is not set")
 
+model_id = "Qwen/Qwen1.5-0.5B-Chat"
 
-def generate_text(text):
-    # Tokenize the input text, including attention mas
-    #input_ids = tokenizer(text, return_tensors="pt", padding=True)
+device = "auto" # torch.device("cuda" if torch.cuda.is_available() else "cpu")
+dtype = torch.bfloat16
 
+tokenizer = AutoTokenizer.from_pretrained(model_id, token=huggingface_token)
 
+print(model_id,device,dtype)
+histories = []
+#model = None
 
-    messages = []
-    use_system_prompt = True
-    DEFAULT_SYSTEM_PROMPT = "you are helpfull assistant."
-    if use_system_prompt:
-        messages = [
-                {"role": "system", "content": DEFAULT_SYSTEM_PROMPT}
-            ]
 
-        user_messages = [
-            {"role": "user", "content": text}
-        ]
-        messages += user_messages
+def call_generate_text(prompt, system_message="You are a helpful assistant."):
+    if prompt =="":
+        print("empty prompt return")
+        return ""
 
-    prompt = tokenizer.apply_chat_template(
-            conversation=messages,
-            add_generation_prompt=True,
-            tokenize=False
-        )
+    global histories
+    #global model
+    #if model != None:# and model.is_cuda:
+    #    print("Model is alive")
+    #else:
+    #    model = AutoModelForCausalLM.from_pretrained(
+    #    model_id, token=huggingface_token ,torch_dtype=dtype,device_map=device
+    #)
+        
+    messages = [
+        {"role": "system", "content": system_message},
+    ]
     
-    input_datas = tokenizer(
-        prompt,
-        add_special_tokens=True,
-        return_tensors="pt"
-        )
-
-    # Generate text, passing the attention mask
-    generated_ids = model.generate(input_ids=input_datas.input_ids, attention_mask=input_datas.attention_mask,max_length=10000)
-    #generated_ids = model.generate(input_ids=input_ids, max_length=100)
+    messages += histories 
+
+    user_message = {"role": "user", "content": prompt}
+
+    messages += [user_message]
     
-    # Decode the generated tokens
-    generated_text = tokenizer.decode(generated_ids[0][input_datas.input_ids.size(1) :], skip_special_tokens=True)
-
-    # Print the generated text
-    #print(generated_text)
-    return generated_text
-
-from flask import Flask, request, jsonify
-
-app = Flask(__name__)
-#app.logger.disabled = True
-#log = logging.getLogger('werkzeug')
-#log.disabled = True
-
-@app.route('/')
-def predict():
-    param_value = request.args.get('param', '') 
-    # ここにモデルの推論ロジックを追加
-    #output = pipe(messages, **generation_args) 
-    #text = (output[0]['generated_text'])
-    #print("hello")
-    #result = {"prediction": "dummy_result"}
-    text = generate_text(param_value)
-    return f"{text}"
-
-if __name__ == '__main__':
-    app.run(host='0.0.0.0', port=7860)
+    try:
+        text = generate_text(messages)
+        histories += [user_message,{"role": "assistant", "content": text}]
+        #model.to("cpu")
+        return text
+    except RuntimeError  as e:
+        print(f"An unexpected error occurred: {e}")
+        #model = None
+
+    return ""
+
+iface = gr.Interface(
+    fn=call_generate_text,
+    inputs=[
+        gr.Textbox(lines=3, label="Input Prompt"),
+        gr.Textbox(lines=2, label="System Message", value="あなたは親切なアシスタントで常に日本語で返答します。"),
+    ],
+    outputs=gr.Textbox(label="Generated Text"),
+    title=f"{model_id}",
+    description=f"{model_id} CPU",
+)
+print("Initialized")
+
+@spaces.GPU(duration=120)
+def generate_text(messages):
+
+    model = AutoModelForCausalLM.from_pretrained(
+        model_id, token=huggingface_token ,torch_dtype=dtype,device_map=device
+    )
+        
+
+    text_generator = pipeline("text-generation", model=model, tokenizer=tokenizer,torch_dtype=dtype,device_map=device) #pipeline has not to(device)
+    result = text_generator(messages, max_new_tokens=256, do_sample=True, temperature=0.7)
+
+    generated_output = result[0]["generated_text"]
+    if isinstance(generated_output, list):
+        for message in reversed(generated_output):
+            if message.get("role") == "assistant":
+                content= message.get("content", "No content found.")
+                return content
+            
+        return "No assistant response found."
+    else:
+        return "Unexpected output format."
+
+if __name__ == "__main__":
+    print("Main")
+    iface.launch()

requirements.txt

@@ -1,5 +1,6 @@
-llama-cpp-python
-transformers
+numpy
 torch
+spaces
 accelerate
-flask
+transformers
+bitsandbytes