From 7d7066f495b74349b18ca54eb683df7dc3e56d51 Mon Sep 17 00:00:00 2001
From: Houyun Sun <18763177732@139.com>
Date: Tue, 27 Oct 2026 15:51:12 +0800
Subject: [PATCH] fix(router): register deployments whose cost-map row has a
 non-chat mode

Since #47419 a deployment with `litellm_params.model_id` and `model_info.base_model`
is priced from the backend model's cost-map row through
`model_info_with_pricing_overrides()`, which validates the merged row with
`TypeAdapter(_PricingModelInfo)`. `ModelInfoBase.mode` is a closed Literal that has no
`rerank`, so a Bedrock rerank deployment using an application inference profile failed
validation while the router was built, the deployment was skipped, and every call to
the model group returned "no healthy deployments".

Add the cost-map modes that were missing from the Literal (`rerank`,
`video_generation`, `image_edit`, `search`, `moderation`, `guardrail`,
`vector_store`), and make the pricing merge non-fatal: if the row still does not
validate, log a warning and keep the caller's explicit `model_info` instead of
dropping the deployment.

Fixes #48545
---
 litellm/types/utils.py                        |  7 +++
 litellm/utils.py                              | 15 ++++-
 tests/unit/test_router_rerank_pricing_mode.py | 59 +++++++++++++++++++
 3 files changed, 78 insertions(+), 3 deletions(-)
 create mode 100644 tests/unit/test_router_rerank_pricing_mode.py

diff --git a/litellm/types/utils.py b/litellm/types/utils.py
index 7e8645fd0dd..51546532f65 100644
--- a/litellm/types/utils.py
+++ b/litellm/types/utils.py
@@ -477,6 +477,13 @@ class ModelInfoBase(ProviderSpecificModelInfo, total=False):
             "extraction",
             "ocr",
             "realtime",
+            "rerank",
+            "video_generation",
+            "image_edit",
+            "search",
+            "moderation",
+            "guardrail",
+            "vector_store",
         ]
     ]
     supported_endpoints: list[str] | None
diff --git a/litellm/utils.py b/litellm/utils.py
index 781aa5f4fce..ded576211b8 100644
--- a/litellm/utils.py
+++ b/litellm/utils.py
@@ -3614,9 +3614,18 @@ def model_info_with_pricing_overrides(model_info: ModelInfo, registration: Mappi
     overrides: Final = {
         name: value for name, value in registration.items() if value is not None and is_server_derived_pricing_key(name)
     }
-    return _PRICING_MODEL_INFO.validate_python(
-        update_dictionary(dict(_without_overridden_price_variants(model_info, overrides)), overrides)
-    )
+    try:
+        return _PRICING_MODEL_INFO.validate_python(
+            update_dictionary(dict(_without_overridden_price_variants(model_info, overrides)), overrides)
+        )
+    except ValidationError as e:
+        # A cost-map row whose shape this version does not recognise (an unknown
+        # ``mode``, for example) must not stop a deployment from being registered.
+        # Keep the caller's explicit model_info instead of dropping the deployment.
+        verbose_logger.warning(
+            "litellm.utils: could not apply cost-map pricing to model_info, keeping the explicit model_info: %s", e
+        )
+        return model_info
 
 
 @dataclass(frozen=True, slots=True)
diff --git a/tests/unit/test_router_rerank_pricing_mode.py b/tests/unit/test_router_rerank_pricing_mode.py
new file mode 100644
index 00000000000..162f0770108
--- /dev/null
+++ b/tests/unit/test_router_rerank_pricing_mode.py
@@ -0,0 +1,59 @@
+"""A cost-map row with a non-chat ``mode`` must not drop the deployment.
+
+A Bedrock rerank deployment that uses an application inference profile and sets
+``model_info.base_model`` is priced from the backend model's cost-map row, which
+carries ``"mode": "rerank"``. ``_PricingModelInfo`` used to reject that literal,
+so the deployment was skipped while the router was being built and every request
+to the model group returned "no healthy deployments".
+"""
+
+import pytest
+
+from litellm import Router
+from litellm.utils import model_info_with_pricing_overrides
+
+RERANK_DEPLOYMENT = {
+    "model_name": "cohere-rerank-3-5",
+    "litellm_params": {
+        "model": "bedrock/arn:aws:bedrock:eu-central-1::foundation-model/cohere.rerank-v3-5:0",
+        "model_id": "arn:aws:bedrock:eu-central-1:123456789012:application-inference-profile/abc123",
+        "aws_region_name": "eu-central-1",
+        "aws_access_key_id": "fake-key",
+        "aws_secret_access_key": "fake-secret",
+    },
+    "model_info": {
+        "base_model": "cohere.rerank-v3-5:0",
+        "input_cost_per_query": 0.00175,
+    },
+}
+
+
+@pytest.fixture(autouse=True)
+def local_cost_map(monkeypatch):
+    monkeypatch.setenv("LITELLM_LOCAL_MODEL_COST_MAP", "True")
+
+
+def test_rerank_cost_map_mode_does_not_drop_the_deployment():
+    router = Router(model_list=[RERANK_DEPLOYMENT])
+
+    assert len(router.model_list) == 1
+    assert router.model_list[0]["model_info"]["input_cost_per_query"] == 0.00175
+
+
+def test_unknown_cost_map_mode_falls_back_to_the_explicit_model_info():
+    """An unrecognised mode must not raise while the router registers a deployment."""
+    model_info = {
+        "key": "some-model",
+        "max_tokens": 10,
+        "max_input_tokens": 10,
+        "max_output_tokens": 10,
+        "input_cost_per_token": 0.0,
+        "output_cost_per_token": 0.0,
+        "litellm_provider": "openai",
+        "mode": "a_mode_this_version_does_not_know",
+        "supported_openai_params": [],
+    }
+
+    priced = model_info_with_pricing_overrides(model_info, {})
+
+    assert priced["mode"] == "a_mode_this_version_does_not_know"