[Model] Support Qwen2 embeddings and use tags to select model tests (#10184)

vllm/model_executor/models/registry.py

@@ -11,7 +11,8 @@ import tempfile
 from abc import ABC, abstractmethod
 from dataclasses import dataclass, field
 from functools import lru_cache
-from typing import Callable, Dict, List, Optional, Tuple, Type, TypeVar, Union
+from typing import (AbstractSet, Callable, Dict, List, Optional, Tuple, Type,
+                    TypeVar, Union)
 
 import cloudpickle
 import torch.nn as nn
@@ -110,6 +111,8 @@ _EMBEDDING_MODELS = {
     },
     "MistralModel": ("llama", "LlamaEmbeddingModel"),
     "Phi3ForCausalLM": ("phi3", "Phi3ForCausalLM"),
+    "Qwen2Model": ("qwen2", "Qwen2EmbeddingModel"),
+    "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
     "Qwen2ForRewardModel": ("qwen2_rm", "Qwen2ForRewardModel"),
     "Qwen2ForSequenceClassification": ("qwen2_cls", "Qwen2ForSequenceClassification"),  # noqa: E501
     # [Multimodal]
@@ -301,8 +304,8 @@ class _ModelRegistry:
     # Keyed by model_arch
     models: Dict[str, _BaseRegisteredModel] = field(default_factory=dict)
 
-    def get_supported_archs(self) -> List[str]:
-        return list(self.models.keys())
+    def get_supported_archs(self) -> AbstractSet[str]:
+        return self.models.keys()
 
     def register_model(
         self,