[Model Runner V2] Don't use UVA buffer for prefill_len (#29713)

Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>

vllm/v1/worker/gpu/model_runner.py

@@ -410,6 +410,8 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
                 cu_num_new_blocks[i].append(x + len(block_ids))
                 new_block_ids[i].extend(block_ids)
             overwrite.append(True)
+        if scheduler_output.scheduled_new_reqs:
+            self.req_states.prefill_len.copy_to_gpu()
 
         # Add new blocks for the existing requests.
         cached_reqs = scheduler_output.scheduled_cached_reqs