--- /tmp/evalscope-sync/src-v191/evalscope/api/model/model.py	2026-07-21 03:18:18.000000000 +0000
+++ /data1/syy/evalscope/evalstone/evalscope/evalscope/api/model/model.py	2026-09-01 03:48:30.955904239 +0000
@@ -1,10 +1,12 @@
 import abc
 import asyncio
+from contextlib import contextmanager
 from functools import partial
 from pydantic_core import to_jsonable_python
 from typing import TYPE_CHECKING, Any, Dict, Generator, List, Literal, Optional, Sequence, Union
 
 from evalscope.api.messages import ChatMessage, ChatMessageAssistant, ChatMessageSystem, ChatMessageUser
+from evalscope.api.messages.request_stats import RequestStats
 from evalscope.api.registry import get_model_api
 from evalscope.api.tool import ToolChoice, ToolFunction, ToolInfo
 from evalscope.utils import get_logger, get_secret_value
@@ -43,6 +45,18 @@
         self.base_url = base_url
         self.api_key = api_key
         self.config = config
+        self.request_stats = RequestStats()
+
+    @contextmanager
+    def _track_logical_request(self):
+        """Count one logical generate() call after HTTP retries have finished."""
+        try:
+            yield
+        except Exception:
+            self.request_stats.record_logical(ok=False)
+            raise
+        else:
+            self.request_stats.record_logical(ok=True)
 
     @abc.abstractmethod
     def generate(
@@ -361,6 +375,10 @@
         if cached is not None:
             return cached
 
+    # TaskConfig defaults api_key to 'EMPTY'; treat it as unset so env vars can be used.
+    if api_key == 'EMPTY':
+        api_key = None
+
     logger.info(
         f'Creating model {model} with eval_type={eval_type} '
         f'base_url={base_url}, config={config.model_dump(exclude_none=True)}, model_args={model_args}'