Memory usage #768
mammadmaheri7
started this conversation in
General
Replies: 1 comment 1 reply
-
This is a pretty big model in terms of num of constraints -- so I think the memory usage makes sense. I'd note that num of params is not always a good proxy for num of constraints (architectures will induce very big differences) |
Beta Was this translation helpful? Give feedback.
1 reply
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
I encountered an issue while attempting to verify the tiny_vit model, which contains approximately 5 million parameters. During the setup step, it is consuming a surprisingly large amount of RAM, exceeding even 1000GB (though the exact amount is unknown). This leads me to wonder whether there might be a bug in the process or if adjusting certain settings variables, such as
num_cols
,logrows
,tolerance
,etc., could potentially resolve the problem.The commands I executed are as follows (the onnx file, calibration data and settings after calibration have been attached in the zip file):
ezkl gen-settings -M output.onnx -O settings.json --input-visibility public --param-visibility fixed --output-visibility public
ezkl calibrate-settings -D cal_data.json -M output.onnx -O settings.json --target resource --lookup-safety-margin 1 --scales 0,5,10,15 --only-range-check-rebase
ezkl compile-circuit -M output.onnx -S settings.json --compiled-circuit network.ezkl
ezkl get-srs -S settings.json --srs-path kzg.srs
ezkl setup -M network.ezkl --srs-path=kzg.srs --vk-path=vk.key --pk-path=pk.key
Additionally, I would like to specify that I built the ezkl tool from the following commit:
Link to Commit
tiny_vit.zip
Beta Was this translation helpful? Give feedback.
All reactions