Reserve multimodal runtime allocations and keep padded inputs aligned (#34141)

Co-authored-by: Hanming Lu <69857889+hanming-lu@users.noreply.github.com>
Co-authored-by: Lianmin Zheng <lianminzheng@gmail.com>
Co-authored-by: wangwenchen0407 <wangwenchen@meta.com>
Co-authored-by: Hanming Lu <hanminglu@meta.com>

test/registered/vlm/test_vision_openai_server_a.py

@@ -8,6 +8,7 @@ import unittest
 
 import openai
 
+from sglang.srt.environ import envs
 from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.test_utils import (
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
@@ -46,6 +47,11 @@ class TestQwen3VLServer(ImageOpenAITestMixin, VideoOpenAITestMixin):
     model = "Qwen/Qwen3-VL-30B-A3B-Instruct"
     extra_args = ["--cuda-graph-max-bs-decode=4"]
 
+    @classmethod
+    def setUpClass(cls):
+        with envs.SGLANG_MM_FEATURE_CACHE_MB.override(512):
+            super().setUpClass()
+
 
 class TestQwen2VLContextLengthServer(CustomTestCase):
     # --context-length 300 is calibrated to this model's mm-token expansion: