Update handler: add random seed

handler.py

@@ -1,4 +1,5 @@
 import time
+import random
 from transformers import AutoTokenizer, AutoModelForCausalLM
 
 class EndpointHandler:
@@ -45,11 +46,15 @@ class EndpointHandler:
 
         model_inputs = self.tokenizer([text], return_tensors="pt").to("cuda")
 
+        torch.manual_seed(random.randint(0, 2 ** 32 - 1))
+
         time_start = time.time()
         generated_ids = self.model.generate(
             **model_inputs,
             max_new_tokens=max_new_tokens,
             temperature=1.0,
+            do_sample=True,
+            top_p=0.9,
             **parameters
         )
         time_end = time.time()