Merge pull request #38344 from ksk2023/fix-cost-alias-double-prefix

fix(cost_calculator): resolve real cost key when model_name alias contains '/'

litellm/cost_calculator.py

@@ -794,14 +794,27 @@ def _select_model_name_for_cost_calc(
         and custom_llm_provider is not None
         and not _model_contains_known_llm_provider(return_model)
     ):  # add provider prefix if not already present, to match model_cost
-        if region_name is not None:
-            return_model = f"{custom_llm_provider}/{region_name}/{return_model}"
-        else:
-            return_model = f"{custom_llm_provider}/{return_model}"
+        provider_prefix: Final = custom_llm_provider if region_name is None else f"{custom_llm_provider}/{region_name}"
+        return_model = _strip_unregistered_leading_segments(f"{provider_prefix}/{return_model}", region_name)
 
     return return_model
 
 
+def _strip_unregistered_leading_segments(model: str, region_name: str | None) -> str:
+    """Resolve a provider-prefixed slash alias like "vertex_ai/vertex/claude-opus-5" to the
+    registered cost key ("vertex_ai/claude-opus-5"), keeping the model unchanged when it already
+    resolves downstream (custom-priced router ids) or no stripped candidate is registered (#38069)."""
+    segments: Final = model.split("/")
+    if "/".join(segments[1:]) in litellm.model_cost:
+        return model
+    head_len: Final = 2 if region_name is not None and len(segments) > 2 and segments[1] == region_name else 1
+    head: Final = "/".join(segments[:head_len])
+    tail: Final = segments[head_len:]
+    strippable: Final = next((index for index, segment in enumerate(tail) if segment in LlmProvidersSet), len(tail))
+    candidates: Final = (f"{head}/{'/'.join(tail[start:])}" for start in range(min(strippable, len(tail) - 1) + 1))
+    return next((candidate for candidate in candidates if candidate in litellm.model_cost), model)
+
+
 @lru_cache(maxsize=DEFAULT_MAX_LRU_CACHE_SIZE)
 def _model_contains_known_llm_provider(model: str) -> bool:
     """

tests/test_litellm/test_cost_calculator.py

@@ -3909,6 +3909,167 @@ def test_completion_cost_prices_anthropic_shaped_cache_read_tokens(_local_model_
     assert cost == pytest.approx(3 * 4e-6 + 4014 * 4e-7 + 5 * 2e-5, rel=1e-9)
 
 
+def test_select_model_name_strips_unregistered_alias_prefix(_local_model_cost_map):
+    """A router-facing model_name alias containing "/" whose leading segment is NOT a
+    registered provider must not be double-prefixed into a non-existent cost key.
+
+    Regression test for #38069: alias "vertex/claude-opus-5" (real deployment
+    "vertex_ai/claude-opus-5") was re-prefixed into "vertex_ai/vertex/claude-opus-5",
+    silently pricing every streamed request at $0.
+    """
+
+    from litellm.cost_calculator import _select_model_name_for_cost_calc
+
+    response = litellm.ModelResponse(
+        id="x",
+        choices=[
+            {
+                "index": 0,
+                "message": {"role": "assistant", "content": "hi"},
+                "finish_reason": "stop",
+            }
+        ],
+        model="vertex/claude-opus-5",
+    )
+    response._hidden_params = {}
+
+    selected = _select_model_name_for_cost_calc(
+        model=None,
+        completion_response=response,
+        custom_llm_provider="vertex_ai",
+    )
+
+    assert selected == "vertex_ai/claude-opus-5"
+
+
+def test_select_model_name_strips_duplicated_region_segment(_local_model_cost_map):
+    """A "region/model" alias whose leading segment repeats the request's region must
+    resolve to the region-priced cost key instead of keeping the region segment twice."""
+
+    from litellm.cost_calculator import _select_model_name_for_cost_calc
+
+    response = litellm.ModelResponse(
+        id="x",
+        choices=[
+            {
+                "index": 0,
+                "message": {"role": "assistant", "content": "hi"},
+                "finish_reason": "stop",
+            }
+        ],
+        model="us-east-1/anthropic.claude-v2:1",
+    )
+    response._hidden_params = {"region_name": "us-east-1"}
+
+    selected = _select_model_name_for_cost_calc(
+        model=None,
+        completion_response=response,
+        custom_llm_provider="bedrock",
+    )
+
+    assert selected == "bedrock/us-east-1/anthropic.claude-v2:1"
+
+
+def test_completion_cost_nonzero_for_slash_alias_model_name(_local_model_cost_map):
+    """End-to-end cost through a "/"-containing alias must price above zero (#38069)."""
+
+    response = litellm.ModelResponse(
+        id="x",
+        choices=[
+            {
+                "index": 0,
+                "message": {"role": "assistant", "content": "hi"},
+                "finish_reason": "stop",
+            }
+        ],
+        model="vertex/claude-opus-5",
+    )
+    response._hidden_params = {"custom_llm_provider": "vertex_ai"}
+    response.usage = litellm.Usage(prompt_tokens=100, completion_tokens=50)
+
+    cost = litellm.completion_cost(
+        completion_response=response,
+        custom_llm_provider="vertex_ai",
+    )
+
+    assert cost == pytest.approx(100 * 5e-6 + 50 * 2.5e-5, rel=1e-9)
+
+
+def test_select_model_name_unresolvable_alias_unchanged(_local_model_cost_map):
+    """An alias that resolves to no known cost key keeps the legacy double-prefixed name."""
+
+    from litellm.cost_calculator import _select_model_name_for_cost_calc
+
+    response = litellm.ModelResponse(
+        id="x",
+        choices=[
+            {
+                "index": 0,
+                "message": {"role": "assistant", "content": "hi"},
+                "finish_reason": "stop",
+            }
+        ],
+        model="team/nonsense-model",
+    )
+    response._hidden_params = {}
+
+    selected = _select_model_name_for_cost_calc(
+        model=None,
+        completion_response=response,
+        custom_llm_provider="vertex_ai",
+    )
+
+    assert selected == "vertex_ai/team/nonsense-model"
+
+
+def test_completion_cost_keeps_custom_priced_slash_router_id(_local_model_cost_map):
+    """A custom-priced router id containing "/" keeps its custom pricing instead of being
+    rewritten to the built-in key its suffix happens to match."""
+
+    from litellm.cost_calculator import _select_model_name_for_cost_calc
+
+    litellm.register_model(
+        model_cost={
+            "vertex/claude-opus-5": {
+                "input_cost_per_token": 7e-6,
+                "output_cost_per_token": 8e-6,
+                "litellm_provider": "vertex_ai",
+            }
+        }
+    )
+
+    selected = _select_model_name_for_cost_calc(
+        model="vertex_ai/claude-opus-5",
+        completion_response=None,
+        custom_pricing=True,
+        custom_llm_provider="vertex_ai",
+        router_model_id="vertex/claude-opus-5",
+    )
+    assert selected == "vertex_ai/vertex/claude-opus-5"
+
+    response = litellm.ModelResponse(
+        id="x",
+        choices=[
+            {
+                "index": 0,
+                "message": {"role": "assistant", "content": "hi"},
+                "finish_reason": "stop",
+            }
+        ],
+        model="vertex/claude-opus-5",
+    )
+    response._hidden_params = {"custom_llm_provider": "vertex_ai"}
+    response.usage = litellm.Usage(prompt_tokens=100, completion_tokens=50)
+
+    cost = litellm.completion_cost(
+        completion_response=response,
+        custom_llm_provider="vertex_ai",
+        custom_pricing=True,
+        router_model_id="vertex/claude-opus-5",
+    )
+    assert cost == pytest.approx(100 * 7e-6 + 50 * 8e-6, rel=1e-9)
+
+
 @pytest.mark.parametrize(
     ("model", "expected_1hr_rate"),
     [("claude-3-haiku-20240307", 5e-07), ("claude-3-opus-20240229", 3e-05)],

type-discipline-budget.json

@@ -27,7 +27,7 @@
     "limit": 0
   },
   "LIT010": {
-    "limit": 16621
+    "limit": 16620
   },
   "LIT011": {
     "limit": 5585