feat: support external custom models (#13429)

Co-authored-by: qiuxuan.lzw <qiuxuan.lzw@alibaba-inc.com>
Co-authored-by: Mick <mickjagger19@icloud.com>

test/srt/external_models/custom_qwen2_vl.py

@@ -0,0 +1,21 @@
+from sglang.srt.models.qwen2_vl import (
+    Qwen2VLForConditionalGeneration as OriginalQwen2VLForConditionalGeneration,
+)
+from sglang.srt.multimodal.processors.qwen_vl import QwenVLImageProcessor
+
+
+class Qwen2VLForConditionalGeneration(OriginalQwen2VLForConditionalGeneration):
+    def __init__(self, config, quant_config, prefix: str = "") -> None:
+        super().__init__(config, quant_config, prefix)
+        print("init custom model:", self.__class__.__name__)
+
+
+class CustomProcessor(QwenVLImageProcessor):
+    models = [Qwen2VLForConditionalGeneration]
+
+    def __init__(self, hf_config, server_args, _processor, *args, **kwargs):
+        super().__init__(hf_config, server_args, _processor, *args, **kwargs)
+        print("init custom processor:", self.__class__.__name__)
+
+
+EntryClass = Qwen2VLForConditionalGeneration

test/srt/run_suite.py

@@ -65,6 +65,7 @@ suites = {
         TestFile("test_eagle_infer_beta.py", 90),
         TestFile("test_constrained_decoding.py", 150),
         TestFile("test_eval_fp8_accuracy.py", 303),
+        TestFile("test_external_models.py", 155),
         TestFile("test_fa3.py", 420),
         TestFile("test_flashmla.py", 230),
         TestFile("test_fp8_utils.py", 5),

test/srt/test_external_models.py

@@ -0,0 +1,28 @@
+import os
+import unittest
+
+import sglang as sgl
+from sglang.test.test_utils import CustomTestCase
+
+
+class TestExternalModels(CustomTestCase):
+    def test_external_model(self):
+        os.environ["SGLANG_EXTERNAL_MODEL_PACKAGE"] = "external_models"
+        os.environ["SGLANG_EXTERNAL_MM_PROCESSOR_PACKAGE"] = "external_models"
+        prompt = "Today is a sunny day and I like"
+        model_path = "Qwen/Qwen2-VL-2B-Instruct"
+
+        engine = sgl.Engine(
+            model_path=model_path,
+            cuda_graph_max_bs=1,
+            max_total_tokens=64,
+            enable_multimodal=True,
+        )
+        out = engine.generate(prompt)["text"]
+        engine.shutdown()
+
+        self.assertGreater(len(out), 0)
+
+
+if __name__ == "__main__":
+    unittest.main()