Reserve more memory for DeepSeekOCR model and adjust server start timeout for DeepGEMM to reduce flakiness (#15277)

test/srt/test_flashmla.py

@@ -35,10 +35,11 @@ class TestFlashMLAAttnBackend(unittest.TestCase):
                     "flashmla",
                 ]
             )
+        # Use longer timeout for DeepGEMM JIT compilation which can take 10-20 minutes
         cls.process = popen_launch_server(
             cls.model,
             cls.base_url,
-            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH * 2,
             other_args=other_args,
         )
 
@@ -91,10 +92,11 @@ class TestFlashMLAMTP(CustomTestCase):
                     "flashmla",
                 ]
             )
+        # Use longer timeout for DeepGEMM JIT compilation which can take 10-20 minutes
         cls.process = popen_launch_server(
             cls.model,
             cls.base_url,
-            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH * 2,
             other_args=other_args,
         )
 

test/srt/test_vision_openai_server_a.py

@@ -162,6 +162,10 @@ class TestQwen2AudioServer(AudioOpenAITestMixin):
 class TestDeepseekOCRServer(TestOpenAIMLLMServerBase):
     model = "deepseek-ai/DeepSeek-OCR"
     trust_remote_code = False
+    extra_args = [
+        "--mem-fraction-static=0.70",
+        "--cuda-graph-max-bs=4",
+    ]
 
     def verify_single_image_response_for_ocr(self, response):
         """Verify DeepSeek-OCR grounding output with coordinates"""

test/srt/test_vision_openai_server_common.py

@@ -38,16 +38,23 @@ class TestOpenAIMLLMServerBase(CustomTestCase):
     def setUpClass(cls):
         cls.base_url = DEFAULT_URL_FOR_TEST
         cls.api_key = "sk-123456"
+
+        # Build other_args: always include extra_args, conditionally include fixed_args
+        other_args = list(cls.extra_args)
+        if cls.trust_remote_code:
+            other_args.extend(cls.fixed_args)
+        else:
+            # Exclude --trust-remote-code but keep other fixed args like --enable-multimodal
+            other_args.extend(
+                arg for arg in cls.fixed_args if arg != "--trust-remote-code"
+            )
+
         cls.process = popen_launch_server(
             cls.model,
             cls.base_url,
             timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
             api_key=cls.api_key,
-            other_args=(
-                cls.extra_args + cls.fixed_args + ["--trust-remote-code"]
-                if cls.trust_remote_code
-                else []
-            ),
+            other_args=other_args,
         )
         cls.base_url += "/v1"