fix(proxy): keep a Vertex Live setup group's budget and pricing answer when the key may call it

Every name a setup frame matches runs the same deployment, so the first name the key may call is re-authorized and its denial stands instead of falling through to the bare model id, which no group's budget or pricing check covers

litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py

@@ -4095,23 +4095,17 @@ async def _first_model_denial(
     authorize_model: VertexLiveModelAuthorizer,
 ) -> ClientFrameRefusal | None:
     """
-    The first name the key may call that also passes re-authorization opens the session, and no later name is
-    checked. Otherwise the refusal is the first re-authorization denial, which says more than an access denial
+    Every name runs the same deployment, so the first name the key may call is re-authorized and that answer
+    stands, budget and pricing checks included. Only a key that may call none of them gets the first access denial
     """
     name: Final = model_names[0]
-    denial: Final = await _model_access_denial(name, valid_token, llm_router) or await _model_reauthorization_denial(
-        name, authorize_model
-    )
-    if denial is None or len(model_names) == 1:
-        return denial
+    access_denial: Final = await _model_access_denial(name, valid_token, llm_router)
+    if access_denial is None:
+        return await _model_reauthorization_denial(name, authorize_model)
+    if len(model_names) == 1:
+        return access_denial
     later_denial: Final = await _first_model_denial(model_names[1:], valid_token, llm_router, authorize_model)
-    if later_denial is None:
-        return None
-    return (
-        later_denial
-        if isinstance(later_denial, LoggedClientFrameRefusal) and isinstance(denial, ProxyException)
-        else denial
-    )
+    return access_denial if isinstance(later_denial, ProxyException) else later_denial
 
 
 async def _vertex_live_model_denial(

tests/unit/proxy/pass_through_endpoints/test_pass_through_endpoints.py

@@ -8541,6 +8541,33 @@ async def test_vertex_live_route_releases_the_budget_it_reserved_to_authorize_an
     assert valid_token.budget_reservation is not None and valid_token.budget_reservation["finalized"] is True
 
 
+@pytest.mark.asyncio
+async def test_vertex_live_setup_group_the_key_may_call_but_fails_budget_is_not_opened_under_its_bare_model_id():
+    """
+    A setup frame naming a group runs that group's deployment whichever name the key holds, so when the key may call
+    the group its budget and pricing checks decide, and the group's bare model id cannot open the session instead
+    """
+    group_denial: Final = ProxyException(
+        message=f"Budget has been exceeded for model {VERTEX_LIVE_GROUP}.",
+        type=ProxyErrorTypes.budget_exceeded,
+        param="model",
+        code=400,
+    )
+    valid_token: Final = UserAPIKeyAuth(token="hashed")
+
+    async def authorize_model(model: str) -> UserAPIKeyAuth:
+        if model == VERTEX_LIVE_GROUP:
+            raise group_denial
+        return valid_token
+
+    websocket, upstream_ws = await _run_vertex_live_gated_passthrough(
+        [_setup_frame(VERTEX_LIVE_GROUP)], valid_token, authorize_model=authorize_model
+    )
+
+    upstream_ws.send.assert_not_awaited()
+    assert _policy_close_reason(websocket) == group_denial.message
+
+
 @pytest.mark.asyncio
 async def test_vertex_live_route_keeps_the_key_allowlist_out_of_the_close_reason_and_in_the_proxy_log(caplog):
     """The client learns which model was refused, while only the proxy log names the models the key may reach"""