Upload app.py

app.py

@@ -131,14 +131,17 @@ def call_llm(messages: List[dict], model: str, logs: List[str]) -> dict:
         sys_txt = messages[0].get("content", "") if messages else ""
         usr_txt = messages[1].get("content", "") if len(messages) > 1 else ""
         extra_rules = "\n\n請務必只輸出單一 JSON 物件，不得包含任何 JSON 之外的文字或符號。"
-
+        print('準備 chat prompt（加上 JSON 輸出約束）')
         chat = [
             {"role": "system", "content": sys_txt},
             {"role": "user", "content": usr_txt + extra_rules}
         ]
+        print(f"user content:{usr_txt + extra_rules}")
+        
         prompt = _hf_tok.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)
 
         inputs = _hf_tok(prompt, return_tensors="pt").to(_hf_model.device)
+        print("inputs")
         with torch.no_grad():
             out_ids = _hf_model.generate(
                 **inputs,
@@ -148,10 +151,13 @@ def call_llm(messages: List[dict], model: str, logs: List[str]) -> dict:
                 eos_token_id=_hf_tok.eos_token_id,
                 pad_token_id=_hf_tok.eos_token_id
             )
+        print("torch.no_grad")
         full = _hf_tok.decode(out_ids[0], skip_special_tokens=True)
         gen = full[len(prompt):] if full.startswith(prompt) else full
+        print("gen")
         logs.append(f"[LOCAL LLM] Gen chars={len(gen)}")
-
+        print(gen)
+        logs.append(gen)
         # 嘗試解析 JSON
         try:
             data = json.loads(gen)