support qwen3 emebedding (sgl-project#6990)

Author: Titan-p

python/sglang/srt/models/qwen3.py

@@ -333,6 +333,8 @@ def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
 
         params_dict = dict(self.named_parameters())
         for name, loaded_weight in weights:
+            if "Embedding" in self.config.name_or_path:
+                name = add_prefix(name, "model")
             layer_id = get_layer_id(name)
             if (
                 layer_id is not None

test/srt/models/test_embedding_models.py

@@ -26,6 +26,7 @@
     ("Alibaba-NLP/gte-Qwen2-1.5B-instruct", 1, 1e-5),
     ("intfloat/e5-mistral-7b-instruct", 1, 1e-5),
     ("marco/mcdse-2b-v1", 1, 1e-5),
+    ("Qwen/Qwen3-Embedding-8B", 1, 1e-5),
     # Temporarily disable before this model is fixed
     # ("jason9693/Qwen2.5-1.5B-apeach", 1, 1e-5),
 ]