Merge e3d7121f60 into 2a9fad611f

litellm/utils.py

@@ -7220,7 +7220,11 @@ def get_model_info_helper(
                         e,
                     )
 
-        if not catalog_only and custom_llm_provider == "huggingface":
+        if (
+            not catalog_only
+            and custom_llm_provider == "huggingface"
+            and not _is_potential_model_name_in_model_cost(potential_model_names)
+        ):
             max_tokens: Final = _get_max_position_embeddings(model_name=model)
             return ModelInfoBase(
                 key=model,

tests/unit/test_cost_calculator.py

@@ -1907,6 +1907,49 @@ def test_tiered_pricing_only_deployment_completion_cost_is_nonzero():
     assert cost > 0
 
 
+@pytest.mark.parametrize(
+    ("registered_model", "model"),
+    [
+        ("hf-tei-rerank-deployment", "hf-tei-rerank-deployment"),
+        ("huggingface/hf-tei-rerank-deployment", "huggingface/hf-tei-rerank-deployment"),
+        ("hf-tei-rerank-deployment", "HF-TEI-RERANK-DEPLOYMENT"),
+    ],
+)
+@pytest.mark.usefixtures("local_model_cost_map")
+def test_huggingface_rerank_uses_registered_input_cost_per_query(
+    registered_model: str, model: str, request: pytest.FixtureRequest
+):
+    """Registered per-query price must apply for huggingface/TEI rerank (#45975).
+
+    ``_get_model_info_helper`` used to return a synthetic zero-cost huggingface
+    description before consulting ``litellm.model_cost``, so a deployment priced
+    only via ``input_cost_per_query`` always billed $0 under that provider.
+    """
+    from litellm.cost_calculator import rerank_cost
+    from litellm.utils import cached_get_model_info_helper
+
+    cached_get_model_info_helper.cache_clear()
+    request.addfinalizer(cached_get_model_info_helper.cache_clear)
+    litellm.model_cost.pop(registered_model, None)
+    litellm.register_model({registered_model: {"input_cost_per_query": 0.001}})
+
+    cost = rerank_cost(
+        model=model,
+        custom_llm_provider="huggingface",
+        billed_units={"search_units": 1},
+    )
+    assert cost == (0.001, 0.0)
+
+    # Unregistered huggingface models still get the synthetic zero-cost fallback.
+    cached_get_model_info_helper.cache_clear()
+    info = litellm.get_model_info(
+        model="unregistered-hf-tei-model-xyz",
+        custom_llm_provider="huggingface",
+    )
+    assert info["input_cost_per_token"] == 0
+    assert info["output_cost_per_token"] == 0
+
+
 def test_per_query_priced_rerank_deployment_completion_cost_is_nonzero():
     """A rerank deployment priced only via ``input_cost_per_query`` must resolve
     cost against its ``router_model_id`` entry: the shared backend alias has