diff --git a/litellm/proxy/hooks/parallel_request_limiter_v3.py b/litellm/proxy/hooks/parallel_request_limiter_v3.py
index 486e3cf88a4..89559ff72ae 100644
--- a/litellm/proxy/hooks/parallel_request_limiter_v3.py
+++ b/litellm/proxy/hooks/parallel_request_limiter_v3.py
@@ -1341,7 +1341,7 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger):
             else:
                 limit_value = rate_limit.get("tokens_per_unit")
                 inc_amount = int(increment_amounts.get("tokens", 0) or 0)
-            if limit_value is None:
+            if limit_value is None or inc_amount < 0:
                 continue
             counter_key = self.create_rate_limit_keys(descriptor_key, descriptor_value, rlt)
             # Counter-key TTL and window_size are conceptually distinct
diff --git a/tests/test_litellm/proxy/hooks/test_parallel_request_limiter_v3.py b/tests/test_litellm/proxy/hooks/test_parallel_request_limiter_v3.py
index 84e41711176..7b3f00a55a8 100644
--- a/tests/test_litellm/proxy/hooks/test_parallel_request_limiter_v3.py
+++ b/tests/test_litellm/proxy/hooks/test_parallel_request_limiter_v3.py
@@ -5055,3 +5055,17 @@ async def test_atomic_check_with_zero_increment_still_enforces_token_limit():
         )
         == 150
     )
+
+    negative_increment: Dict[str, int] = {"requests": -1, "tokens": -50}
+    refund_attempt = await handler.atomic_check_and_increment_by_n(
+        descriptors=[descriptor],
+        increments=[negative_increment],
+    )
+    assert refund_attempt["overall_code"] == "OK"
+    assert refund_attempt["statuses"] == []
+    assert (
+        await handler.internal_usage_cache.async_get_cache(
+            key=counter_key, litellm_parent_otel_span=None, local_only=True
+        )
+        == 150
+    )