feat: use static kv cache

generation_utils.py

@@ -375,6 +375,7 @@ def process_batch(batch_items, tokenizer, model, spt, device, system_prompt, sta
         outputs = model.generate(
             input_ids=input_ids, 
             attention_mask=attention_mask,
+            cache_implementation="static",
         )
         print(f"Original outputs shape: {outputs.shape}")
         print(f"Start value: {start}")