fix(greptile): tie fireworks index cache to model_cost mutation generation

litellm/llms/fireworks_ai/chat/transformation.py

@@ -26,6 +26,7 @@ from litellm.types.utils import (
     ProviderSpecificModelInfo,
 )
 from litellm.utils import (
+    get_model_cost_mutation_generation,
     supports_function_calling,
     supports_reasoning,
     supports_tool_choice,
@@ -254,15 +255,16 @@ class FireworksAIConfig(OpenAIGPTConfig):
 
     # Cached index of fireworks_ai/* entries from litellm.model_cost. Building
     # this index requires a full scan of model_cost (tens of thousands of
-    # entries), so we memoize it and invalidate when the dict identity or size
-    # changes. The cached value is a list of (key_short, model_info) tuples
-    # restricted to fireworks_ai/* entries.
+    # entries), so we memoize it. The cache key is (id(model_cost),
+    # mutation_generation): the generation counter is bumped on every
+    # register_model / reload path, so add+remove or in-place value
+    # replacement (which can leave id and len unchanged) still invalidates.
     _fireworks_index_cache: Optional[Tuple[int, int, List[Tuple[str, dict]]]] = None
 
     @classmethod
     def _get_fireworks_index(cls) -> List[Tuple[str, dict]]:
         model_cost = litellm.model_cost
-        signature = (id(model_cost), len(model_cost))
+        signature = (id(model_cost), get_model_cost_mutation_generation())
         cached = cls._fireworks_index_cache
         if (
             cached is not None

litellm/utils.py

@@ -5387,6 +5387,16 @@ def _strip_model_name(model: str, custom_llm_provider: Optional[str]) -> str:
 # Global case-insensitive lookup map for model_cost (built eagerly at module import)
 _model_cost_lowercase_map: Optional[Dict[str, str]] = None
 
+# Monotonic counter bumped on every model_cost mutation. Consumers that
+# memoize derived state (e.g. provider-specific indices) can include this
+# value in their cache key so they invalidate even when key add+remove or
+# in-place value replacement leaves len/id unchanged.
+_model_cost_mutation_generation: int = 0
+
+
+def get_model_cost_mutation_generation() -> int:
+    return _model_cost_mutation_generation
+
 
 def _invalidate_model_cost_lowercase_map() -> None:
     """Invalidate the case-insensitive lookup map for model_cost.
@@ -5394,8 +5404,9 @@ def _invalidate_model_cost_lowercase_map() -> None:
     Call this whenever litellm.model_cost is modified to ensure the map is rebuilt.
     Also clears related LRU caches that depend on model_cost data.
     """
-    global _model_cost_lowercase_map
+    global _model_cost_lowercase_map, _model_cost_mutation_generation
     _model_cost_lowercase_map = None
+    _model_cost_mutation_generation += 1
 
     # Clear LRU caches that depend on model_cost data
     get_model_info.cache_clear()