[Bugfix][Frontend] Apply output_config and thinking in Anthropic count_tokens (#59180)

Signed-off-by: Kang Nahoon <ejrtks102020@gmail.com>
Co-authored-by: Flora Feng <4florafeng@gmail.com>

tests/entrypoints/anthropic/test_anthropic_messages_conversion.py

@@ -1987,6 +1987,32 @@ class TestThinkingConfig:
         assert result.include_reasoning is True
         assert result.thinking_token_budget is None
 
+    @pytest.mark.parametrize(
+        "config",
+        [
+            pytest.param({"output_config": {"effort": "low"}}, id="effort"),
+            pytest.param({"thinking": {"type": "disabled"}}, id="disabled"),
+            pytest.param(
+                {"thinking": {"type": "enabled", "budget_tokens": 2048}},
+                id="enabled",
+            ),
+        ],
+    )
+    def test_count_tokens_matches_messages(self, config):
+        """Effort and thinking change the rendered prompt, so count_tokens must
+        apply them the same way /v1/messages does."""
+        messages = [{"role": "user", "content": "Hello"}]
+        expected = _convert(
+            AnthropicMessagesRequest(
+                model="test-model", max_tokens=4096, messages=messages, **config
+            )
+        )
+        result = _convert(
+            AnthropicCountTokensRequest(model="test-model", messages=messages, **config)
+        )
+        assert result.reasoning_effort == expected.reasoning_effort
+        assert result.thinking_token_budget == expected.thinking_token_budget
+
 
 class TestProbeDisabledThinkingEffort:
     """``auto`` falls back to ``low`` when ``none`` cannot turn thinking off."""

vllm/entrypoints/anthropic/protocol.py

@@ -385,7 +385,9 @@ class AnthropicCountTokensRequest(BaseModel):
 
     model: str
     messages: list[AnthropicMessage]
+    output_config: AnthropicOutputConfig | None = None
     system: str | list[AnthropicContentBlock] | None = None
+    thinking: AnthropicThinkingConfig | None = None
     tool_choice: AnthropicToolChoice | None = None
     tools: list[AnthropicTool] | None = None
 

vllm/entrypoints/anthropic/serving.py

@@ -633,8 +633,6 @@ class AnthropicServingMessages(OpenAIServingChat):
         ``display`` is intentionally ignored: suppressing reasoning would mark
         it ended for structured outputs and drop it from multi-turn history.
         """
-        if isinstance(anthropic_request, AnthropicCountTokensRequest):
-            return
         thinking: AnthropicThinkingConfig | None = anthropic_request.thinking
         if thinking is None:
             return
@@ -655,8 +653,6 @@ class AnthropicServingMessages(OpenAIServingChat):
         anthropic_request: AnthropicMessagesRequest | AnthropicCountTokensRequest,
     ) -> None:
         """Handle output configuration such as output format and effort."""
-        if isinstance(anthropic_request, AnthropicCountTokensRequest):
-            return
         output_config: AnthropicOutputConfig | None = anthropic_request.output_config
         if output_config and output_config.format and output_config.format.json_schema:
             req.response_format = ResponseFormat(
@@ -1238,9 +1234,13 @@ class AnthropicServingMessages(OpenAIServingChat):
         raw_request: Request | None = None,
     ) -> AnthropicCountTokensResponse | ErrorResponse:
         """Implements Anthropic's messages.count_tokens endpoint."""
+        disabled_thinking_effort: AnthropicDisabledThinkingEffort = "none"
+        if request.thinking is not None and request.thinking.type == "disabled":
+            disabled_thinking_effort = await self._get_disabled_thinking_effort()
         chat_req = self.to_chat_completion_request(
             request,
             merge_inline_system=self._merge_inline_system,
+            disabled_thinking_effort=disabled_thinking_effort,
         )
         result = await self.render_chat_request(chat_req)
         if isinstance(result, ErrorResponse):