From 341323a9272a2138d478f6020e3b7fee3c533610 Mon Sep 17 00:00:00 2001
From: Krrish Dholakia <krrishdholakia@berri.ai>
Date: Mon, 20 Jul 2026 23:01:41 +0000
Subject: [PATCH] fix(router): duration-only provider budget config no longer
 removes deployments

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---
 litellm/router_strategy/budget_limiter.py     | 27 ++++---
 .../router_strategy/test_budget_limiter.py    | 81 +++++++++++++++++++
 2 files changed, 95 insertions(+), 13 deletions(-)
 create mode 100644 tests/test_litellm/router_strategy/test_budget_limiter.py

diff --git a/litellm/router_strategy/budget_limiter.py b/litellm/router_strategy/budget_limiter.py
index 067f38ab11c..f723dc05ece 100644
--- a/litellm/router_strategy/budget_limiter.py
+++ b/litellm/router_strategy/budget_limiter.py
@@ -222,20 +222,21 @@ class RouterBudgetLimiting(CustomLogger):
                     provider = self._get_llm_provider_for_deployment(deployment)
                 if provider in provider_configs:
                     config = provider_configs[provider]
-                    if config.max_budget is None:
-                        continue
-                    current_spend = spend_map.get(f"provider_spend:{provider}:{config.budget_duration}", 0.0)
-                    self._track_provider_remaining_budget_prometheus(
-                        provider=provider,
-                        spend=current_spend,
-                        budget_limit=config.max_budget,
-                    )
+                    if config.max_budget is not None:
+                        current_spend = spend_map.get(f"provider_spend:{provider}:{config.budget_duration}", 0.0)
+                        self._track_provider_remaining_budget_prometheus(
+                            provider=provider,
+                            spend=current_spend,
+                            budget_limit=config.max_budget,
+                        )
 
-                    if config.max_budget and current_spend >= config.max_budget:
-                        debug_msg = f"Exceeded budget for provider {provider}: {current_spend} >= {config.max_budget}"
-                        deployment_above_budget_info += f"{debug_msg}\n"
-                        is_within_budget = False
-                        continue
+                        if current_spend >= config.max_budget:
+                            debug_msg = (
+                                f"Exceeded budget for provider {provider}: {current_spend} >= {config.max_budget}"
+                            )
+                            deployment_above_budget_info += f"{debug_msg}\n"
+                            is_within_budget = False
+                            continue
 
             # Check deployment budget
             if self.deployment_budget_config and is_within_budget:
diff --git a/tests/test_litellm/router_strategy/test_budget_limiter.py b/tests/test_litellm/router_strategy/test_budget_limiter.py
new file mode 100644
index 00000000000..32d3b473de0
--- /dev/null
+++ b/tests/test_litellm/router_strategy/test_budget_limiter.py
@@ -0,0 +1,81 @@
+"""Regression tests for RouterBudgetLimiting._filter_out_deployments_above_budget."""
+
+from litellm.router_strategy.budget_limiter import RouterBudgetLimiting
+from litellm.types.utils import BudgetConfig
+
+
+def _make_limiter(provider_budget_config):
+    limiter = RouterBudgetLimiting.__new__(RouterBudgetLimiting)
+    limiter.provider_budget_config = provider_budget_config
+    limiter.deployment_budget_config = None
+    limiter.tag_budget_config = None
+    return limiter
+
+
+def _openai_deployment():
+    return {
+        "model_name": "chat",
+        "litellm_params": {"model": "openai/gpt-4o-mini", "custom_llm_provider": "openai"},
+        "model_info": {"id": "deployment-1"},
+    }
+
+
+def test_duration_only_provider_config_keeps_deployment():
+    """A provider budget with budget_duration but no max_budget must NOT drop deployments.
+
+    Regression test for https://github.com/BerriAI/litellm/issues/33327
+    A missing max_budget means there is no numeric cap, not an exceeded budget.
+    """
+    limiter = _make_limiter({"openai": BudgetConfig(budget_duration="1d", max_budget=None)})
+
+    result, info = limiter._filter_out_deployments_above_budget(
+        potential_deployments=[],
+        healthy_deployments=[_openai_deployment()],
+        provider_configs={"openai": limiter.provider_budget_config["openai"]},
+        deployment_configs={},
+        deployment_providers=["openai"],
+        spend_map={},
+        request_tags=[],
+    )
+
+    assert len(result) == 1
+    assert info == ""
+
+
+def test_provider_over_max_budget_filters_deployment():
+    """A numeric max_budget that is exceeded must still drop the deployment.
+
+    Guards against the duration-only fix accidentally disabling budget enforcement.
+    """
+    limiter = _make_limiter({"openai": BudgetConfig(budget_duration="1d", max_budget=10.0)})
+
+    result, info = limiter._filter_out_deployments_above_budget(
+        potential_deployments=[],
+        healthy_deployments=[_openai_deployment()],
+        provider_configs={"openai": limiter.provider_budget_config["openai"]},
+        deployment_configs={},
+        deployment_providers=["openai"],
+        spend_map={"provider_spend:openai:1d": 25.0},
+        request_tags=[],
+    )
+
+    assert len(result) == 0
+    assert "Exceeded budget for provider openai" in info
+
+
+def test_provider_under_max_budget_keeps_deployment():
+    """A numeric max_budget that is not exceeded keeps the deployment eligible."""
+    limiter = _make_limiter({"openai": BudgetConfig(budget_duration="1d", max_budget=10.0)})
+
+    result, info = limiter._filter_out_deployments_above_budget(
+        potential_deployments=[],
+        healthy_deployments=[_openai_deployment()],
+        provider_configs={"openai": limiter.provider_budget_config["openai"]},
+        deployment_configs={},
+        deployment_providers=["openai"],
+        spend_map={"provider_spend:openai:1d": 3.0},
+        request_tags=[],
+    )
+
+    assert len(result) == 1
+    assert info == ""