From cd8d4b9dfcefffe254c4e354e3a18c7a644c06bd Mon Sep 17 00:00:00 2001
From: Qiaolin Yu <qy254@cornell.edu>
Date: Thu, 15 May 2025 13:09:55 -0400
Subject: [PATCH] Fix lora bench (#6302)

---
 benchmark/lora/lora_bench.py           | 2 ++
 python/sglang/srt/lora/lora_manager.py | 4 +---
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/benchmark/lora/lora_bench.py b/benchmark/lora/lora_bench.py
index 1d9394261..199b3f3ab 100644
--- a/benchmark/lora/lora_bench.py
+++ b/benchmark/lora/lora_bench.py
@@ -170,6 +170,7 @@ async def benchmark(
         prompt_len=test_prompt_len,
         output_len=test_output_len,
         lora_name="dummy",  # the lora_name argument will not be used
+        image_data=None,
         extra_request_body=extra_request_body,
     )
     test_output = await request_func(request_func_input=test_input)
@@ -194,6 +195,7 @@ async def benchmark(
             prompt_len=prompt_len,
             output_len=output_len,
             lora_name="dummy",
+            image_data=None,
             extra_request_body=extra_request_body,
         )
         tasks.append(
diff --git a/python/sglang/srt/lora/lora_manager.py b/python/sglang/srt/lora/lora_manager.py
index 70e6ca838..68b2a3621 100644
--- a/python/sglang/srt/lora/lora_manager.py
+++ b/python/sglang/srt/lora/lora_manager.py
@@ -170,9 +170,7 @@ class LoRAManager:
                 dim=0,
                 out=self.cuda_graph_batch_info.seg_indptr[1 : bs + 1],
             )
-            self.cuda_graph_batch_info.max_len = int(
-                torch.max(self.cuda_graph_batch_info.seg_lens[:bs])
-            )
+            self.cuda_graph_batch_info.max_len = 1
 
             for i, lora_path in enumerate(forward_batch.lora_paths):
                 self.cuda_graph_batch_info.weight_indices[i] = (