[Core] Rework dtype resolution (#18751)

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

tests/models/language/pooling/mteb_utils.py

@@ -102,21 +102,18 @@ def mteb_test_embed_models(hf_runner,
         vllm_main_score = run_mteb_embed_task(VllmMtebEncoder(vllm_model),
                                               MTEB_EMBED_TASKS)
         vllm_dtype = vllm_model.model.llm_engine.model_config.dtype
-        model_dtype = getattr(
-            vllm_model.model.llm_engine.model_config.hf_config, "torch_dtype",
-            vllm_dtype)
 
-    with set_default_torch_dtype(model_dtype) and hf_runner(
+    with set_default_torch_dtype(vllm_dtype) and hf_runner(
             model_info.name, is_sentence_transformer=True,
-            dtype=model_dtype) as hf_model:
+            dtype=vllm_dtype) as hf_model:
 
         if hf_model_callback is not None:
             hf_model_callback(hf_model)
 
         st_main_score = run_mteb_embed_task(hf_model, MTEB_EMBED_TASKS)
 
-    print("VLLM:", vllm_dtype, vllm_main_score)
-    print("SentenceTransformer:", model_dtype, st_main_score)
+    print("VLLM:", vllm_main_score)
+    print("SentenceTransformers:", st_main_score)
     print("Difference:", st_main_score - vllm_main_score)
 
     assert st_main_score == pytest.approx(vllm_main_score, abs=MTEB_EMBED_TOL)

tests/models/language/pooling/test_classification.py

@@ -43,6 +43,6 @@ def test_models(
 
         # the tolerance value of 1e-2 is selected based on the
         # half datatype tests in
-        # tests/models/embedding/language/test_embedding.py
+        # tests/models/language/pooling/test_embedding.py
         assert torch.allclose(hf_output, vllm_output,
                               1e-3 if dtype == "float" else 1e-2)

tests/models/language/pooling/test_embedding.py

@@ -30,13 +30,11 @@ from ...utils import check_embeddings_close
         pytest.param("sentence-transformers/stsb-roberta-base-v2"),
     ],
 )
-@pytest.mark.parametrize("dtype", ["half"])
 def test_models(
     hf_runner,
     vllm_runner,
     example_prompts,
     model,
-    dtype: str,
     monkeypatch,
 ) -> None:
 
@@ -58,13 +56,11 @@ def test_models(
     # So we need to strip the input texts to avoid test failing.
     example_prompts = [str(s).strip() for s in example_prompts]
 
-    with hf_runner(model, dtype=dtype,
-                   is_sentence_transformer=True) as hf_model:
+    with hf_runner(model, is_sentence_transformer=True) as hf_model:
         hf_outputs = hf_model.encode(example_prompts)
 
     with vllm_runner(model,
                      task="embed",
-                     dtype=dtype,
                      max_model_len=None,
                      **vllm_extra_kwargs) as vllm_model:
         vllm_outputs = vllm_model.encode(example_prompts)