diff --git a/litellm/proxy/spend_tracking/input_tokens.py b/litellm/proxy/spend_tracking/input_tokens.py index 5589875d86f..59c7fe2a41b 100644 --- a/litellm/proxy/spend_tracking/input_tokens.py +++ b/litellm/proxy/spend_tracking/input_tokens.py @@ -146,7 +146,8 @@ def _count_input_tokens_for_models( def models_counted_for(model: str, llm_router: Router | None, team_id: str | None) -> tuple[str, ...]: """The names to count a request for `model` from `team_id` with, keeping the largest count, one per deployment - routing would serve it from: the ones named `model`, else those of the one wildcard route that serves it. + routing would serve it from: the team's or else the shared ones named `model`, its alias target, or its routing + group's members, else those of the one wildcard route that serves it. An OpenAI deployment counts as its OpenAI model, so it gets that model's tokenizer and PDF page pricing, prefixed `openai/` when OpenAI itself serves it, so `openai_chat_pdf_file_tokens` can add what its chat completions bills. A @@ -160,12 +161,13 @@ def models_counted_for(model: str, llm_router: Router | None, team_id: str | Non def _routed_deployments(model: str, llm_router: Router, team_id: str | None) -> Sequence[object]: registry: Final = llm_router.registry() - named: Final = ( - *registry.deployments(model, source="all", team_id=team_id), - *registry.deployments(model, source="alias"), - *registry.deployments(model, source="routing_groups"), + resolved: Final = registry.resolve(model, kind="alias") or model + routing_group: Final = llm_router.get_routing_group(resolved) + members: Final = routing_group.models if routing_group is not None else (resolved,) + named: Final = tuple( + chain.from_iterable(registry.deployments(member, source="all", team_id=team_id) for member in members) ) - return named or registry.deployments(model, source="wildcard", team_id=team_id) + return named or registry.deployments(resolved, source="wildcard", team_id=team_id) def openai_chat_pdf_file_tokens(counted_model: str, pdf_files: int) -> int: diff --git a/tests/unit/proxy/spend_tracking/test_input_tokens.py b/tests/unit/proxy/spend_tracking/test_input_tokens.py index 82f1d021cb8..40aee7bb00d 100644 --- a/tests/unit/proxy/spend_tracking/test_input_tokens.py +++ b/tests/unit/proxy/spend_tracking/test_input_tokens.py @@ -159,6 +159,30 @@ def test_a_team_request_counts_with_the_deployment_its_team_is_served_by(): assert models_counted_for("reports", router, "team-b") == ("reports",) +@pytest.mark.parametrize("requested", ["reports", "latest-reports"], ids=["routing-group", "alias-of-routing-group"]) +def test_a_team_request_to_a_routing_group_counts_with_the_member_deployment_its_team_is_served_by(requested: str): + router: Final = Router( + model_list=[ + { + "model_name": "drafts", + "litellm_params": {"model": "azure/gpt-5.6-sol", "api_key": "x", "api_base": "https://x"}, + }, + { + "model_name": "drafts-team-a", + "litellm_params": {"model": "anthropic/claude-haiku-5-5", "api_key": "x"}, + "model_info": {"team_id": "team-a", "team_public_model_name": "drafts"}, + }, + ], + routing_groups=[{"group_name": "reports", "models": ["drafts"], "routing_strategy": "simple-shuffle"}], + model_group_alias={"latest-reports": "reports"}, + ) + + assert models_counted_for(requested, router, "team-a") == models_counted_for( + requested, _router((requested, "anthropic/claude-haiku-5-5")), None + ) + assert models_counted_for(requested, router, "team-b") == ("azure/gpt-5.6-sol",) + + def test_a_team_with_its_own_wildcard_counts_with_it_and_never_with_a_proxy_wide_wildcard(): router: Final = Router( model_list=[