Fix external_models import path and migrate model loading tests (#16458)

python/sglang/srt/layers/moe/moe_runner/deep_gemm.py

@@ -45,7 +45,7 @@ _DEEPGEMM_ON_H20 = get_bool_env_var("SGLANG_DEEPGEMM_ON_H20")
 
 # TODO(kaixih@nvidia): ideally we should merge this logic into
 # `fill_gateup_input_triton_kernel` to directly generate e8m0 scale.
-@torch.compile
+@torch.compile(disable=_is_hip or _is_npu)
 def _cast_to_e8m0_with_rounding_up(x: torch.Tensor) -> torch.Tensor:
     temp = x.to(torch.float32).view(torch.int32)
     exp = torch.bitwise_right_shift(temp, 23)

python/sglang/test/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