Keep K3 suite selection and report-schema scoring in bash, merge K3/vision dataset_args into dpv4 yamls, and pin EvalScope at 735d920ee911 with local patches. Co-authored-by: Cursor <cursoragent@cursor.com>
105 lines
4.4 KiB
Diff
105 lines
4.4 KiB
Diff
--- /tmp/evalscope-sync/src-v191/evalscope/models/openai_responses.py 2026-07-21 03:18:18.000000000 +0000
|
|
+++ /data1/syy/evalscope/evalstone/evalscope/evalscope/models/openai_responses.py 2026-09-01 03:48:30.959904179 +0000
|
|
@@ -74,27 +74,29 @@
|
|
) -> ModelOutput:
|
|
request, tools, config = self._build_request(input, tools, tool_choice, config)
|
|
|
|
- try:
|
|
- t_start = time.monotonic()
|
|
- ttft: Optional[float] = None
|
|
-
|
|
- response = retry_call(
|
|
- self.client.responses.create,
|
|
- retries=config.retries,
|
|
- sleep_interval=config.retry_interval,
|
|
- **request,
|
|
- )
|
|
- if not self._is_response_object(response):
|
|
- response, ttft = collect_response_stream(response, request_start=t_start)
|
|
-
|
|
- total_time = time.monotonic() - t_start
|
|
- return self._build_output(response, tools, total_time, ttft)
|
|
-
|
|
- except (BadRequestError, UnprocessableEntityError, PermissionDeniedError) as ex:
|
|
- return self.handle_bad_request(ex)
|
|
- except ValueError as ex:
|
|
- logger.error(f'Model [{self.model_name}] returned an invalid response: {ex}')
|
|
- raise
|
|
+ with self._track_logical_request():
|
|
+ try:
|
|
+ t_start = time.monotonic()
|
|
+ ttft: Optional[float] = None
|
|
+
|
|
+ response = retry_call(
|
|
+ self.client.responses.create,
|
|
+ retries=config.retries,
|
|
+ sleep_interval=config.retry_interval,
|
|
+ on_attempt=self.request_stats.on_attempt,
|
|
+ **request,
|
|
+ )
|
|
+ if not self._is_response_object(response):
|
|
+ response, ttft = collect_response_stream(response, request_start=t_start)
|
|
+
|
|
+ total_time = time.monotonic() - t_start
|
|
+ return self._build_output(response, tools, total_time, ttft)
|
|
+
|
|
+ except (BadRequestError, UnprocessableEntityError, PermissionDeniedError) as ex:
|
|
+ return self.handle_bad_request(ex)
|
|
+ except ValueError as ex:
|
|
+ logger.error(f'Model [{self.model_name}] returned an invalid response: {ex}')
|
|
+ raise
|
|
|
|
async def generate_async(
|
|
self,
|
|
@@ -105,27 +107,29 @@
|
|
) -> ModelOutput:
|
|
request, tools, config = self._build_request(input, tools, tool_choice, config)
|
|
|
|
- try:
|
|
- t_start = time.monotonic()
|
|
- ttft: Optional[float] = None
|
|
-
|
|
- response = await async_retry_call(
|
|
- self.async_client.responses.create,
|
|
- retries=config.retries,
|
|
- sleep_interval=config.retry_interval,
|
|
- **request,
|
|
- )
|
|
- if not self._is_response_object(response):
|
|
- response, ttft = await async_collect_response_stream(response, request_start=t_start)
|
|
-
|
|
- total_time = time.monotonic() - t_start
|
|
- return self._build_output(response, tools, total_time, ttft)
|
|
-
|
|
- except (BadRequestError, UnprocessableEntityError, PermissionDeniedError) as ex:
|
|
- return self.handle_bad_request(ex)
|
|
- except ValueError as ex:
|
|
- logger.error(f'Model [{self.model_name}] returned an invalid response: {ex}')
|
|
- raise
|
|
+ with self._track_logical_request():
|
|
+ try:
|
|
+ t_start = time.monotonic()
|
|
+ ttft: Optional[float] = None
|
|
+
|
|
+ response = await async_retry_call(
|
|
+ self.async_client.responses.create,
|
|
+ retries=config.retries,
|
|
+ sleep_interval=config.retry_interval,
|
|
+ on_attempt=self.request_stats.on_attempt,
|
|
+ **request,
|
|
+ )
|
|
+ if not self._is_response_object(response):
|
|
+ response, ttft = await async_collect_response_stream(response, request_start=t_start)
|
|
+
|
|
+ total_time = time.monotonic() - t_start
|
|
+ return self._build_output(response, tools, total_time, ttft)
|
|
+
|
|
+ except (BadRequestError, UnprocessableEntityError, PermissionDeniedError) as ex:
|
|
+ return self.handle_bad_request(ex)
|
|
+ except ValueError as ex:
|
|
+ logger.error(f'Model [{self.model_name}] returned an invalid response: {ex}')
|
|
+ raise
|
|
|
|
def _build_request(
|
|
self,
|