Bug Description
Issue
For small models and high concurrency numbers the following config snippet causes unexpected behavior:
settings:
runtime:
min_duration_ms: 600000
max_duration_ms: 3600000
Expected
The number of samples is calculated dynamically to satisfy min_duration_ms.
Observed
The experiment is run to produce a performance and/or an accuracy reports, without any time check.
We have seen experiments with Duration: 74.34 seconds in their report.txt files.
Workaround
Explicitly set n_samples_to_issue chosen so, that the minimum time constrain is satisfied:
settings:
runtime:
min_duration_ms: 600000
max_duration_ms: 3600000
n_samples_to_issue: 100000
Steps to Reproduce
- Create a config with big enough
min_duration_ms, say, 10h, but without n_samples_to_issue.
- Run an experiment with the config.
- Check the time taken (or look into
report.txt for the duration of the experiment).
Environment
inference-endpoints dad7d1a
Relevant Logs
Before submitting
Bug Description
Issue
For small models and high concurrency numbers the following config snippet causes unexpected behavior:
Expected
The number of samples is calculated dynamically to satisfy
min_duration_ms.Observed
The experiment is run to produce a performance and/or an accuracy reports, without any time check.
We have seen experiments with
Duration: 74.34 secondsin theirreport.txtfiles.Workaround
Explicitly set
n_samples_to_issuechosen so, that the minimum time constrain is satisfied:Steps to Reproduce
min_duration_ms, say, 10h, but withoutn_samples_to_issue.report.txtfor the duration of the experiment).Environment
inference-endpoints dad7d1a
Relevant Logs
Before submitting