fix: second_per_grid_ts should be used to get mrope position (#3682)

test/srt/run_suite.py

@@ -68,7 +68,7 @@ suites = {
         TestFile("test_update_weights_from_tensor.py", 48),
         TestFile("test_vertex_endpoint.py", 31),
         TestFile("test_vision_chunked_prefill.py", 223),
-        TestFile("test_vision_llm.py", 18.4),
+        TestFile("test_vlm_accuracy.py", 60),
         TestFile("test_vision_openai_server.py", 344),
         TestFile("test_fim_completion.py", 120),
         TestFile("test_w8a8_quantization.py", 46),

test/srt/test_vision_openai_server.py

@@ -191,7 +191,7 @@ class TestOpenAIVisionServer(unittest.TestCase):
         # from transformers import AutoTokenizer
         from decord import VideoReader, cpu
 
-        max_frames_num = 12
+        max_frames_num = 20
         vr = VideoReader(video_path, ctx=cpu(0))
         total_frame_num = len(vr)
         uniform_sampled_frames = np.linspace(
@@ -226,6 +226,22 @@ class TestOpenAIVisionServer(unittest.TestCase):
 
         return messages
 
+    def prepare_video_messages_video_direct(self, video_path):
+        messages = [
+            {
+                "role": "user",
+                "content": [
+                    {
+                        "type": "image_url",
+                        "image_url": {"url": f"video:{video_path}"},
+                        "modalities": "video",
+                    },
+                    {"type": "text", "text": "Please describe the video in detail."},
+                ],
+            },
+        ]
+        return messages
+
     def test_video_chat_completion(self):
         url = "https://raw.githubusercontent.com/EvolvingLMMs-Lab/sglang/dev/onevision_local/assets/jobs.mp4"
         cache_dir = os.path.expanduser("~/.cache")
@@ -241,6 +257,7 @@ class TestOpenAIVisionServer(unittest.TestCase):
 
         client = openai.Client(api_key=self.api_key, base_url=self.base_url)
 
+        # messages = self.prepare_video_messages_video_direct(file_path)
         messages = self.prepare_video_messages(file_path)
 
         video_request = client.chat.completions.create(
@@ -266,6 +283,7 @@ class TestOpenAIVisionServer(unittest.TestCase):
             "man" in video_response
             or "person" in video_response
             or "individual" in video_response
+            or "speaker" in video_response
         ), video_response
         assert (
             "present" in video_response
@@ -368,7 +386,7 @@ class TestOpenAIVisionServer(unittest.TestCase):
             list(executor.map(self.run_decode_with_image, image_ids))
 
 
-class TestQWen2VLServer(TestOpenAIVisionServer):
+class TestQwen2VLServer(TestOpenAIVisionServer):
     @classmethod
     def setUpClass(cls):
         cls.model = "Qwen/Qwen2-VL-7B-Instruct"
@@ -382,14 +400,14 @@ class TestQWen2VLServer(TestOpenAIVisionServer):
             other_args=[
                 "--chat-template",
                 "qwen2-vl",
-                "--chunked-prefill-size",
-                "10000",
+                "--mem-fraction-static",
+                "0.4",
             ],
         )
         cls.base_url += "/v1"
 
 
-class TestQWen2_5_VLServer(TestOpenAIVisionServer):
+class TestQwen2_5_VLServer(TestOpenAIVisionServer):
     @classmethod
     def setUpClass(cls):
         cls.model = "Qwen/Qwen2.5-VL-7B-Instruct"
@@ -403,9 +421,6 @@ class TestQWen2_5_VLServer(TestOpenAIVisionServer):
             other_args=[
                 "--chat-template",
                 "qwen2-vl",
-                # FIXME: workaround to chunked prefill within image embeds
-                "--chunked-prefill-size",
-                "10000",
                 "--mem-fraction-static",
                 "0.4",
             ],
@@ -508,6 +523,8 @@ class TestMinicpmvServer(TestOpenAIVisionServer):
                 "--trust-remote-code",
                 "--chat-template",
                 "minicpmv",
+                "--mem-fraction-static",
+                "0.4",
             ],
         )
         cls.base_url += "/v1"

test/srt/test_vlm_accuracy.py

@@ -17,8 +17,6 @@ from sglang.srt.model_executor.model_runner import ModelRunner
 from sglang.srt.openai_api.protocol import ChatCompletionRequest
 from sglang.srt.server_args import ServerArgs
 
-MiniCPMV = "openbmb/MiniCPM-V-2_6"
-
 
 # Test the logits output between HF and SGLang
 class VisionLLMLogitsBase(unittest.IsolatedAsyncioTestCase):
@@ -155,7 +153,7 @@ class TestMiniCPMVLogits(VisionLLMLogitsBase):
     @classmethod
     def setUpClass(cls):
         super().setUpClass()
-        cls.model_path = MiniCPMV
+        cls.model_path = "openbmb/MiniCPM-V-2_6"
         cls.tokenizer = AutoTokenizer.from_pretrained(
             cls.model_path, trust_remote_code=True
         )