[misc] Do not allow to use lora with chunked prefill. (vllm-project#5538

)

Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>

vllm/config.py

@@ -1124,6 +1124,8 @@ def verify_with_scheduler_config(self, scheduler_config: SchedulerConfig):
                 "Due to limitations of the custom LoRA CUDA kernel, "
                 "max_num_batched_tokens must be <= 65528 when "
                 "LoRA is enabled.")
+        if scheduler_config.chunked_prefill_enabled:
+            raise ValueError("LoRA is not supported with chunked prefill yet.")
 
 
 @dataclass