fix(token_counter): support input_audio content blocks in _count_content_list (#38459)

litellm/litellm_core_utils/token_counter.py

@@ -738,6 +738,17 @@ def _count_content_list(
                 tool_name = str(c.get("tool_name") or "")
                 if tool_name:
                     num_tokens += count_function(tool_name)
+            elif c["type"] == "input_audio":
+                # OpenAI input_audio content block: {"type": "input_audio", "input_audio": {"data": "<b64>", "format": "wav"}}
+                # Estimate tokens based on payload size (or default floor)
+                input_audio = c.get("input_audio")
+                b64_data = input_audio.get("data") if isinstance(input_audio, dict) else None
+                if b64_data and isinstance(b64_data, str):
+                    decoded_bytes = len(b64_data) * 3 // 4
+                    # 1 token ~ 32 bytes of 24kHz/16-bit mono audio (approx standard across audio models)
+                    num_tokens += max(decoded_bytes // 32, 50)
+                else:
+                    num_tokens += 50
             else:
                 content_type = c.get("type", type(c).__name__) if isinstance(c, dict) else type(c).__name__
                 raise ValueError(

tests/test_litellm/test_input_audio_token_counter_38459.py

@@ -0,0 +1,32 @@
+import unittest
+from litellm.litellm_core_utils.token_counter import _count_content_list
+
+class TestInputAudioTokenCounter(unittest.TestCase):
+    def test_input_audio_block_does_not_raise(self):
+        """
+        Regression test for #38459:
+        _count_content_list should handle input_audio blocks without raising ValueError.
+        """
+        content_list = [
+            {"type": "text", "text": "Hello audio"},
+            {
+                "type": "input_audio",
+                "input_audio": {
+                    "data": "UklGRiQAAABXQVZFZm10IBAAAAABAAEAQB8AAIA+AAACABAAZGF0YQAAAAA=",
+                    "format": "wav"
+                }
+            }
+        ]
+        
+        # Simple character counter for testing
+        def mock_count_func(s):
+            return len(s)
+            
+        total_tokens = _count_content_list(
+            count_function=mock_count_func,
+            content_list=content_list,
+            use_default_image_token_count=True,
+            default_token_count=None
+        )
+        
+        self.assertGreater(total_tokens, len("Hello audio"))