Review suggestion + adds one more test

Signed-off-by: Flavia Beo <[email protected]>

tests/test_config.py

@@ -122,6 +122,23 @@ def test_get_pooling_config():
     assert minilm_pooling_config.pooling_type == PoolingType.MEAN
 
 
+def test_get_bert_sentence_transformer_config():
+    bge_model_config = ModelConfig(
+        model="BAAI/bge-base-en-v1.5",
+        tokenizer="BAAI/bge-base-en-v1.5",
+        tokenizer_mode="auto",
+        trust_remote_code=False,
+        seed=0,
+        dtype="float16",
+        revision=None,
+    )
+
+    bert_bge_model_config = bge_model_config._get_bert_config()
+
+    assert bert_bge_model_config["max_seq_length"] == 512
+    assert bert_bge_model_config["do_lower_case"]
+
+
 def test_rope_customization():
     TEST_ROPE_SCALING = {"rope_type": "dynamic", "factor": 2.0}
     TEST_ROPE_THETA = 16_000_000.0

vllm/config.py

@@ -1818,6 +1818,9 @@ def _get_and_verify_max_len(
                     "original_max_position_embeddings"]
             derived_max_model_len *= scaling_factor
 
+    if bert_config and "max_seq_lenght" in bert_config:
+        derived_max_model_len = bert_config["max_seq_length"]
+
     # If the user specified a max length, make sure it is smaller than the
     # derived length from the HF model config.
     if max_model_len is None:
@@ -1851,8 +1854,6 @@ def _get_and_verify_max_len(
                     f"{msg} To allow overriding this maximum, set "
                     "the env var VLLM_ALLOW_LONG_MAX_MODEL_LEN=1")
 
-    if bert_config:
-        max_model_len = bert_config["max_seq_length"]
     return int(max_model_len)