[Model] Automatic conversion of classification and reward models (#11469)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>

tests/models/embedding/language/test_cls_models.py

@@ -1,7 +1,4 @@
-"""Compare the outputs of HF and vLLM when using greedy sampling.
-
-This test only tests small models. Big models such as 7B should be tested from
-test_big_models.py because it could use a larger instance to run tests.
+"""Compare the classification outputs of HF and vLLM models.
 
 Run `pytest tests/models/test_cls_models.py`.
 """

tests/models/embedding/language/test_scoring.py

@@ -1,6 +1,6 @@
-"""Compare the embedding outputs of HF and vLLM models.
+"""Compare the scoring outputs of HF and vLLM models.
 
-Run `pytest tests/models/embedding/language/test_embedding.py`.
+Run `pytest tests/models/embedding/language/test_scoring.py`.
 """
 import math
 

tests/models/test_registry.py

@@ -6,7 +6,9 @@ import torch.cuda
 from vllm.model_executor.models import (is_pooling_model,
                                         is_text_generation_model,
                                         supports_multimodal)
-from vllm.model_executor.models.adapters import as_embedding_model
+from vllm.model_executor.models.adapters import (as_classification_model,
+                                                 as_embedding_model,
+                                                 as_reward_model)
 from vllm.model_executor.models.registry import (_MULTIMODAL_MODELS,
                                                  _SPECULATIVE_DECODING_MODELS,
                                                  _TEXT_GENERATION_MODELS,
@@ -29,9 +31,10 @@ def test_registry_imports(model_arch):
             or model_arch in _MULTIMODAL_MODELS):
         assert is_text_generation_model(model_cls)
 
-    # All vLLM models should be convertible to an embedding model
-    embed_model = as_embedding_model(model_cls)
-    assert is_pooling_model(embed_model)
+    # All vLLM models should be convertible to a pooling model
+    assert is_pooling_model(as_classification_model(model_cls))
+    assert is_pooling_model(as_embedding_model(model_cls))
+    assert is_pooling_model(as_reward_model(model_cls))
 
     if model_arch in _MULTIMODAL_MODELS:
         assert supports_multimodal(model_cls)