--- /tmp/evalscope-sync/src-v191/evalscope/models/litellm_compatible.py	2026-07-21 03:18:18.000000000 +0000
+++ /data1/syy/evalscope/evalstone/evalscope/evalscope/models/litellm_compatible.py	2026-09-01 03:48:30.955904239 +0000
@@ -87,40 +87,42 @@
         if self.base_url:
             request['api_base'] = self.base_url
 
-        try:
-            t_start = time.monotonic()
-
-            response = retry_call(
-                litellm.completion,
-                retries=config.retries,
-                sleep_interval=config.retry_interval,
-                **request,
-            )
-
-            total_time = time.monotonic() - t_start
-            ttft: Optional[float] = None
-
-            if config.stream and not isinstance(response, ChatCompletion):
-                completion, ttft = collect_stream_response(response, request_start=t_start)
-            else:
-                completion = ChatCompletion(**response.model_dump())
-
-            choices = chat_choices_from_openai(completion, tools)
-            output = model_output_from_openai(completion, choices)
-
-            output.time = total_time
-            usage = output.usage
-            output.message.perf_metrics = PerformanceMetrics(
-                latency=total_time,
-                ttft=ttft,
-                input_tokens=usage.input_tokens if usage else 0,
-                output_tokens=usage.output_tokens if usage else 0,
-            )
-            return output
-
-        except Exception as ex:
-            logger.error(f'LiteLLM [{self.model_name}] error: {ex}')
-            raise
+        with self._track_logical_request():
+            try:
+                t_start = time.monotonic()
+
+                response = retry_call(
+                    litellm.completion,
+                    retries=config.retries,
+                    sleep_interval=config.retry_interval,
+                    on_attempt=self.request_stats.on_attempt,
+                    **request,
+                )
+
+                total_time = time.monotonic() - t_start
+                ttft: Optional[float] = None
+
+                if config.stream and not isinstance(response, ChatCompletion):
+                    completion, ttft = collect_stream_response(response, request_start=t_start)
+                else:
+                    completion = ChatCompletion(**response.model_dump())
+
+                choices = chat_choices_from_openai(completion, tools)
+                output = model_output_from_openai(completion, choices)
+
+                output.time = total_time
+                usage = output.usage
+                output.message.perf_metrics = PerformanceMetrics(
+                    latency=total_time,
+                    ttft=ttft,
+                    input_tokens=usage.input_tokens if usage else 0,
+                    output_tokens=usage.output_tokens if usage else 0,
+                )
+                return output
+
+            except Exception as ex:
+                logger.error(f'LiteLLM [{self.model_name}] error: {ex}')
+                raise
 
     async def generate_async(
         self,
@@ -157,38 +159,40 @@
         if self.base_url:
             request['api_base'] = self.base_url
 
-        try:
-            t_start = time.monotonic()
-
-            # Async generation with retry
-            response = await async_retry_call(
-                litellm.acompletion,
-                retries=config.retries,
-                sleep_interval=config.retry_interval,
-                **request,
-            )
-
-            total_time = time.monotonic() - t_start
-            ttft: Optional[float] = None
-
-            if config.stream and not isinstance(response, ChatCompletion):
-                completion, ttft = await async_collect_stream_response(response, request_start=t_start)
-            else:
-                completion = ChatCompletion(**response.model_dump())
-
-            choices = chat_choices_from_openai(completion, tools)
-            output = model_output_from_openai(completion, choices)
-
-            output.time = total_time
-            usage = output.usage
-            output.message.perf_metrics = PerformanceMetrics(
-                latency=total_time,
-                ttft=ttft,
-                input_tokens=usage.input_tokens if usage else 0,
-                output_tokens=usage.output_tokens if usage else 0,
-            )
-            return output
-
-        except Exception as ex:
-            logger.error(f'LiteLLM [{self.model_name}] async error: {ex}')
-            raise
+        with self._track_logical_request():
+            try:
+                t_start = time.monotonic()
+
+                # Async generation with retry
+                response = await async_retry_call(
+                    litellm.acompletion,
+                    retries=config.retries,
+                    sleep_interval=config.retry_interval,
+                    on_attempt=self.request_stats.on_attempt,
+                    **request,
+                )
+
+                total_time = time.monotonic() - t_start
+                ttft: Optional[float] = None
+
+                if config.stream and not isinstance(response, ChatCompletion):
+                    completion, ttft = await async_collect_stream_response(response, request_start=t_start)
+                else:
+                    completion = ChatCompletion(**response.model_dump())
+
+                choices = chat_choices_from_openai(completion, tools)
+                output = model_output_from_openai(completion, choices)
+
+                output.time = total_time
+                usage = output.usage
+                output.message.perf_metrics = PerformanceMetrics(
+                    latency=total_time,
+                    ttft=ttft,
+                    input_tokens=usage.input_tokens if usage else 0,
+                    output_tokens=usage.output_tokens if usage else 0,
+                )
+                return output
+
+            except Exception as ex:
+                logger.error(f'LiteLLM [{self.model_name}] async error: {ex}')
+                raise