Update handler.py

handler.py

@@ -16,36 +16,7 @@ class EndpointHandler():
         self.tokenizer = AutoTokenizer.from_pretrained(cfg['repo'])
 
     def __call__(self, data: Dict[str, Any]) -> List[Dict[str, Any]]:
-        inputs = data.pop("inputs", "")
-        temperature = data.pop("temperature", None)
-        if not temperature:
-            temperature = data.pop("temp", 0.33)
-        if temperature > 3 or temperature < 0:
-            return json.dumps({
-                "status": "error",
-                "reason": "invalid temperature ( 0.01 - 1.00 )"
-            })
-        top_p = data.pop("top-p", 0.85)
-        if top_p > 3 or top_p < 0:
-            return json.dumps({
-                "status": "error",
-                "reason": "invalid top percentage ( 0.01 - 1.00 )"
-            })
-        top_k = data.pop("top-k", 42)
-        if top_k > 100 or top_k < 0:
-            return json.dumps({
-                "status": "error",
-                "reason": "invalid top k ( 1 - 99 )"
-            })
-        system_prompt = data.pop("system-prompt", "You are a helpful assistant.")
-        fmat = data.pop("format", f"<|system|>\n{system_prompt} <|end|>\n<|user|>\n{inputs} <|end|>\n<|assistant|>")
-        try:
-            fmat = fmat.format(system_prompt = system_prompt, prompt = inputs)
-        except Exception as e:
-            return json.dumps({
-                "status": "error",
-                "reason": "invalid format"
-            })
+        inputs = data.pop("inputs", "Q: What is the chemical composition of common concrete in 2024?\nA: ")
         max_length = data.pop("max_length", 1024)
         try:
             max_length = int(max_length)
@@ -55,6 +26,10 @@ class EndpointHandler():
                 "reason": "max_length was passed as something that was absolutely not a plain old int"
             })
         
-        res = self.model(fmat, temperature=temperature, top_p=top_p, top_k=top_k, max_tokens=max_length)
+        res = self.model(f"""
+<|user|>
+{inputs} <|end|>
+<|assistant|>
+""", max_new_tokens=max_new_tokens, do_sample=False)
 
         return res