Move mem_fraction_static adjustment for multimodal models to server_args.py & Fix session control & Other cleanups (#7748)

2025-07-04 16:33:33 -07:00
parent 975a5ec69c
commit 14229ccf8f
16 changed files with 339 additions and 137 deletions
--- a/test/srt/test_server_args.py
+++ b/test/srt/test_server_args.py
@@ -11,12 +11,14 @@ class TestPrepareServerArgs(CustomTestCase):
        server_args = prepare_server_args(
            [
                "--model-path",
-                "model_path",
+                "meta-llama/Meta-Llama-3.1-8B-Instruct",
                "--json-model-override-args",
                '{"rope_scaling": {"factor": 2.0, "rope_type": "linear"}}',
            ]
        )
-        self.assertEqual(server_args.model_path, "model_path")
+        self.assertEqual(
+            server_args.model_path, "meta-llama/Meta-Llama-3.1-8B-Instruct"
+        )
        self.assertEqual(
            json.loads(server_args.json_model_override_args),
            {"rope_scaling": {"factor": 2.0, "rope_type": "linear"}},