From 3201e85d65cfd1df64db05fc11ec7b32bb466ad1 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 1/2] 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 44dc75f45d7..e87e1e1fddc 100644
--- a/litellm/utils.py
+++ b/litellm/utils.py
@@ -6097,7 +6097,7 @@ 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 8dbef772472..d238a953c74 100644
--- a/tests/unit/test_cost_calculator.py
+++ b/tests/unit/test_cost_calculator.py
@@ -1049,6 +1049,37 @@ 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

From a6f9e518813aec7d6d4f65d547ac4cfa225edbdc Mon Sep 17 00:00:00 2001
From: er-s-an <92594860+er-s-an@users.noreply.github.com>
Date: Mon, 12 Oct 2026 17:16:10 +0800
Subject: [PATCH 2/2] fix(huggingface): use typed registered-model lookup

---
 litellm/utils.py                   |  2 +-
 tests/unit/test_cost_calculator.py | 20 ++++++++++++++++----
 2 files changed, 17 insertions(+), 5 deletions(-)

diff --git a/litellm/utils.py b/litellm/utils.py
index e87e1e1fddc..b2c62639363 100644
--- a/litellm/utils.py
+++ b/litellm/utils.py
@@ -6097,7 +6097,7 @@ def get_model_info_helper(
                         e,
                     )
 
-        if custom_llm_provider == "huggingface" and model not in litellm.model_cost:
+        if custom_llm_provider == "huggingface" and get_model_cost_key(model) is None:
             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 d238a953c74..f5c0949da94 100644
--- a/tests/unit/test_cost_calculator.py
+++ b/tests/unit/test_cost_calculator.py
@@ -1049,7 +1049,18 @@ def test_tiered_pricing_only_deployment_completion_cost_is_nonzero():
     assert cost > 0
 
 
-def test_huggingface_rerank_uses_registered_input_cost_per_query():
+@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
@@ -1060,11 +1071,12 @@ def test_huggingface_rerank_uses_registered_input_cost_per_query():
     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}})
+    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="hf-tei-rerank-deployment",
+        model=model,
         custom_llm_provider="huggingface",
         billed_units={"search_units": 1},
     )