[BugFix] fix wrong output when using lora and num_scheduler_steps=8 (#11161)

FIX issue https://github.com/vllm-project/vllm/issues/9688
https://github.com/vllm-project/vllm/issues/11086

 #12487

---------
Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>
Co-authored-by: weilong.yu <weilong.yu@shopee.com>
Co-authored-by: Jee Jee Li <pandaleefree@gmail.com>

vllm/worker/model_runner.py

@@ -1346,6 +1346,10 @@ class GPUModelRunnerBase(ModelRunnerBase[TModelInputForGPU]):
 
             self.execute_model(model_input, kv_caches, intermediate_tensors)
             torch.cuda.synchronize()
+            if self.lora_config:
+                # Remove dummy loras.
+                assert self.lora_manager is not None
+                self.remove_all_loras()
             return
 
     def remove_all_loras(self):

vllm/worker/worker.py

@@ -264,10 +264,7 @@ class Worker(LocalOrDistributedWorkerBase):
                f"{(available_kv_cache_memory / GiB_bytes):.2f}GiB.")
 
         logger.info(msg)
-
         # Final cleanup
-        if self.model_runner.lora_manager:
-            self.model_runner.remove_all_loras()
         gc.collect()
 
         return num_gpu_blocks, num_cpu_blocks