Improve the output precision of embedding models (#19092)

2025-06-04 19:48:57 +08:00
parent 8711bc5e68
commit 35cf32df30
8 changed files with 69 additions and 28 deletions
--- a/tests/models/language/pooling/embed_utils.py
+++ b/tests/models/language/pooling/embed_utils.py
@@ -56,14 +56,10 @@ def correctness_test_embed_models(hf_runner,
                     max_model_len=None,
                     **vllm_extra_kwargs) as vllm_model:
        vllm_outputs = vllm_model.encode(example_prompts)
-        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 hf_runner(
            model_info.name,
-            dtype=model_dtype,
+            dtype="float32",
            is_sentence_transformer=True,
    ) as hf_model: