Need help in understaning "inf" value for --benchmark-grace-period

[Maze999]

Hi

--benchmark-grace-period
The grace period in seconds to wait for responses after benchmark duration ends. Only applies when --benchmark-duration is set. Responses received within this period are included in metrics. Use 'inf' to wait indefinitely for all responses.
Constraints: ≥ 0
Default: 30.0

does inf here mean that we can pass "inf" literally ?
i have a littly ambiguity because the constraints list it as >=0 and while passing "inf" is not being rejected it does result in an error

[ajcasagrande]

@Maze999
Yes, you should be able to pass a literal inf string, though it looks like you have uncovered a bug.

inf is the default for warmup I believe.

[ajcasagrande]

@Maze999 can you post your aiperf version and exact command line? I'm having trouble reproducing this.

[Maze999]

@ajcasagrande

aiperf --version

0.4.0

aiperf profile --model 'meta/llama-3.1-70b-instruct' \

--tokenizer 'meta-llama/Llama-3.1-70B-Instruct'
--endpoint-type 'chat' --streaming --url 'http://x.x.x.x' --random-seed 1234
--synthetic-input-tokens-mean 7000 --synthetic-input-tokens-stddev 0 --concurrency 512 --num-prompts 512 --output-tokens-mean 1000 --output-tokens-stddev 0 --extra-inputs 'max_completion_tokens:1000'
--extra-inputs 'ignore_eos:true' --ui 'dashboard' --benchmark-duration 900 --benchmark-grace-period inf --slice-duration 300 --no-server-metrics
--artifact-dir 'aiperf-artifacts4/meta_llama-3.1-70b-instruct-openai-chat-concurrency512' --profile-export-file 'llama-31-70b-fp8-tp2-pp1-latency-2nim-aiperfv04-run8_Summarization_512_7000_1000'

From log:

2026-01-28 06:20:58 - dataset_manager_e96daa1c - INFO - inputs.json file generated in 0.05 seconds
2026-01-28 06:20:58 - dataset_manager_e96daa1c - INFO - Dataset configured in 1.15 seconds
2026-01-28 06:20:58 - system_controller - INFO - All services configured in 2.59 seconds
2026-01-28 06:20:58 - system_controller - INFO - AIPerf System is CONFIGURED
2026-01-28 06:20:58 - timing_manager_de1d2dd9 - INFO - Credit issuing strategy for Request_Rate started
2026-01-28 06:20:58 - system_controller - INFO - All services started profiling successfully
2026-01-28 06:20:58 - system_controller - INFO - AIPerf System is PROFILING
2026-01-28 06:22:41 - asyncio - ERROR - Task exception was never retrieved
future: <Task finished name='Task-123' coro=<ZMQPullClient._process_message() done, defined at /usr/local/lib/python3.12/dist-packages/aiperf/zmq/pull_client.py:124> exception=OverflowError('cannot convert float infinity to integer')>
Traceback (most recent call last):
File "/usr/local/lib/python3.12/dist-packages/aiperf/zmq/pull_client.py", line 138, in _process_message
await self._pull_callbacksmessage.message_type
File "/usr/local/lib/python3.12/dist-packages/aiperf/records/records_manager.py", line 194, in _on_metric_records
should_include_request = self._should_include_request_by_duration(record_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/dist-packages/aiperf/records/records_manager.py", line 286, in _should_include_request_by_duration
duration_end_ns = self.start_time_ns + int(
^^^^
OverflowError: cannot convert float infinity to integer
2026-01-28 06:22:42 - asyncio - ERROR - Task exception was never retrieved
future: <Task finished name='Task-124' coro=<ZMQPullClient._process_message() done, defined at /usr/local/lib/python3.12/dist-packages/aiperf/zmq/pull_client.py:124> exception=OverflowError('cannot convert float infinity to integer')>
Traceback (most recent call last):
File "/usr/local/lib/python3.12/dist-packages/aiperf/zmq/pull_client.py", line 138, in _process_message
await self._pull_callbacksmessage.message_type
File "/usr/local/lib/python3.12/dist-packages/aiperf/records/records_manager.py", line 194, in _on_metric_records
should_include_request = self._should_include_request_by_duration(record_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/dist-packages/aiperf/records/records_manager.py", line 286, in _should_include_request_by_duration
duration_end_ns = self.start_time_ns + int(
^^^^
OverflowError: cannot convert float infinity to integer
2026-01-28 06:22:42 - asyncio - ERROR - Task exception was never retrieved
future: <Task finished name='Task-125' coro=<ZMQPullClient._process_message() done, defined at /usr/local/lib/python3.12/dist-packages/aiperf/zmq/pull_client.py:124> exception=OverflowError('cannot convert float infinity to integer')>
Traceback (most recent call last):
File "/usr/local/lib/python3.12/dist-packages/aiperf/zmq/pull_client.py", line 138, in _process_message
await self._pull_callbacksmessage.message_type
File "/usr/local/lib/python3.12/dist-packages/aiperf/records/records_manager.py", line 194, in _on_metric_records
should_include_request = self._should_include_request_by_duration(record_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/dist-packages/aiperf/records/records_manager.py", line 286, in _should_include_request_by_duration
duration_end_ns = self.start_time_ns + int(

[ajcasagrande]

Hi @Maze999 inf support was not tested on older versions. With my latest timing system refactors in main that are part of the 0.5.0 release, you likely wont see this issue anymore, as that specific code path has been completely removed.

[Maze999]

Thank you @ajcasagrande , As of now i am using the latest released 0.4.0
Do you have any tentative timeline for 0.5.0 release ?

[ajcasagrande]

@Maze999 I think officially Feb 11th. Its pretty much ready to go though.

https://github.com/ai-dynamo/aiperf/tree/release/0.5.0