fix: bedrock-converse config-block placement, usage calculation, and stop filtering

litellm/llms/bedrock/chat/converse_transformation.py

@@ -891,12 +891,17 @@ class AmazonConverseConfig(BaseConfig):
                 optional_params["maxTokens"] = value
             if param == "stream":
                 optional_params["stream"] = value
-            if param == "stop":
-                if isinstance(value, str):
-                    if len(value) == 0:  # converse raises error for empty strings
-                        continue
-                    value = [value]
-                optional_params["stopSequences"] = value
+        if param == "stop":
+            if isinstance(value, str):
+                if len(value) == 0:  # converse raises error for empty strings
+                    continue
+                value = [value]
+            elif isinstance(value, list):
+                # Filter out empty strings from list (Bedrock rejects them)
+                value = [v for v in value if isinstance(v, str) and len(v) > 0]
+                if not value:  # Skip if empty list or all empty strings
+                    continue
+            optional_params["stopSequences"] = value
             if param == "temperature":
                 optional_params["temperature"] = value
             if param == "top_p":
@@ -1433,6 +1438,16 @@ class AmazonConverseConfig(BaseConfig):
             if tool_choice_values is not None:
                 bedrock_tool_config["toolChoice"] = tool_choice_values
 
+        # Extract config block params BEFORE calling _transform_inference_params
+        # to prevent them from being included in InferenceConfig(**inference_params)
+        # which would place them inside inferenceConfig instead of at the top level.
+        # Fixes: serviceTier, guardrailConfig, performanceConfig placement bug (#17336)
+        config_block_values: dict = {}
+        for config_name, config_class in self.get_config_blocks().items():
+            config_value = inference_params.pop(config_name, None)
+            if config_value is not None:
+                config_block_values[config_name] = config_class(**config_value)  # type: ignore
+
         data: CommonRequestObject = {
             "additionalModelRequestFields": additional_request_params,
             "system": system_content_blocks,
@@ -1441,11 +1456,9 @@ class AmazonConverseConfig(BaseConfig):
             ),
         }
 
-        # Handle all config blocks
-        for config_name, config_class in self.get_config_blocks().items():
-            config_value = inference_params.pop(config_name, None)
-            if config_value is not None:
-                data[config_name] = config_class(**config_value)  # type: ignore
+        # Apply config blocks at top-level (already extracted before _transform_inference_params)
+        for config_name, config_block in config_block_values.items():
+            data[config_name] = config_block  # type: ignore
 
         # Tool Config
         if bedrock_tool_config is not None:
@@ -1633,7 +1646,9 @@ class AmazonConverseConfig(BaseConfig):
         if "cacheWriteInputTokens" in usage:
             cache_creation_input_tokens = usage["cacheWriteInputTokens"]
             input_tokens += cache_creation_input_tokens
-
+        # Recalculate total_tokens to include cache tokens added to input_tokens
+        # The API's totalTokens doesn't include cache hits/writes, but prompt_tokens does
+        total_tokens = input_tokens + output_tokens
         prompt_tokens_details = PromptTokensDetailsWrapper(
             cached_tokens=cache_read_input_tokens
         )

tests/litellm/llms/bedrock/test_bedrock_converse_transformation.py

@@ -0,0 +1,66 @@
+import pytest
+from litellm.llms.bedrock.chat.converse_transformation import AmazonConverseConfig
+from litellm.types.llms.bedrock import ConverseTokenUsageBlock
+
+def test_transform_request_helper_config_blocks():
+    config = AmazonConverseConfig()
+    model = "anthropic.claude-3-haiku-20240307-v1:0"
+    
+    # Test serviceTier placement (Bug #17336)
+    optional_params = {
+        "serviceTier": {"type": "flex"},
+        "guardrailConfig": {"guardrailIdentifier": "test", "guardrailVersion": "1"}
+    }
+    
+    data = config._transform_request_helper(
+        model=model,
+        system_content_blocks=[],
+        optional_params=optional_params
+    )
+    
+    # Assert top-level placement
+    assert "serviceTier" in data
+    assert data["serviceTier"] == {"type": "flex"}
+    assert "guardrailConfig" in data
+    
+    # Assert NOT in inferenceConfig
+    assert "serviceTier" not in data["inferenceConfig"]
+    assert "guardrailConfig" not in data["inferenceConfig"]
+
+def test_transform_usage_recalculation():
+    config = AmazonConverseConfig()
+    usage_block = ConverseTokenUsageBlock(
+        inputTokens=100,
+        outputTokens=50,
+        totalTokens=150,
+        cacheReadInputTokens=20
+    )
+    
+    usage = config._transform_usage(usage_block)
+    
+    # Assert input_tokens includes cache tokens
+    assert usage.prompt_tokens == 120
+    # Assert total_tokens is recalculated (Bug 2)
+    assert usage.total_tokens == 170
+
+def test_map_openai_params_stop_filtering():
+    config = AmazonConverseConfig()
+    optional_params = {}
+    
+    # Test empty list (Bug 3)
+    config.map_openai_params(
+        non_default_params={"stop": []},
+        optional_params=optional_params,
+        model="bedrock/anthropic.claude-3",
+        drop_params=False
+    )
+    assert "stopSequences" not in optional_params
+    
+    # Test list with empty strings
+    config.map_openai_params(
+        non_default_params={"stop": ["", "valid"]},
+        optional_params=optional_params,
+        model="bedrock/anthropic.claude-3",
+        drop_params=False
+    )
+    assert optional_params["stopSequences"] == ["valid"]