From de59488d061c9f2b9ccdb3d22f00da7d09fabfd6 Mon Sep 17 00:00:00 2001 From: er-s-an <92594860+er-s-an@users.noreply.github.com> Date: Mon, 12 Oct 2026 09:21:03 +0800 Subject: [PATCH] fix(huggingface): respect registered model_cost for TEI rerank pricing Skip the synthetic zero-cost huggingface ModelInfo fallback when the model is already present in litellm.model_cost, so input_cost_per_query on a TEI rerank deployment is billed instead of always returning $0. Fixes #45975 --- litellm/utils.py | 2 +- tests/unit/test_cost_calculator.py | 31 ++++++++++++++++++++++++++++++ 2 files changed, 32 insertions(+), 1 deletion(-) diff --git a/litellm/utils.py b/litellm/utils.py index 781aa5f4fce..44ce3655eee 100644 --- a/litellm/utils.py +++ b/litellm/utils.py @@ -6645,7 +6645,7 @@ def get_model_info_helper( e, ) - if custom_llm_provider == "huggingface": + if custom_llm_provider == "huggingface" and model not in litellm.model_cost: max_tokens: Final = _get_max_position_embeddings(model_name=model) return ModelInfoBase( key=model, diff --git a/tests/unit/test_cost_calculator.py b/tests/unit/test_cost_calculator.py index 27b25c1f98e..ab18f265e3d 100644 --- a/tests/unit/test_cost_calculator.py +++ b/tests/unit/test_cost_calculator.py @@ -1559,6 +1559,37 @@ def test_tiered_pricing_only_deployment_completion_cost_is_nonzero(): assert cost > 0 +def test_huggingface_rerank_uses_registered_input_cost_per_query(): + """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() + litellm.model_cost.pop("hf-tei-rerank-deployment", None) + litellm.register_model({"hf-tei-rerank-deployment": {"input_cost_per_query": 0.001}}) + + cost = rerank_cost( + model="hf-tei-rerank-deployment", + 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