You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
From my discussion with Harsh here's our observation related to CUDA 10.1:
Increasing the netty readtimeout to 120 second in UT resolves the issues, however it slows down the build process a bit.
On P3.8xlarge it takes 5m + time for installation.
On my local MacBook it takes around ~(3 mins 20 sec) as compared to ~(1 min 50 sec) without this change
As for "Add a sample warmup code in UT” : we could add a sample script in or before UT which will execute a sample model on GPU (outside TorchServe) which will warmup the GPU before actually starting the UT execution. As per our observation it takes around 2 minutes time for GPU warmup with cudatoolkit 10.1 subsequent uninstall and re-install on GPU machine with timeout changes takes ~3 mins (as GPUs have warmed up in the first install) , which is same as that on CPU.
The text was updated successfully, but these errors were encountered:
From my discussion with Harsh here's our observation related to CUDA 10.1:
Increasing the netty readtimeout to 120 second in UT resolves the issues, however it slows down the build process a bit.
On P3.8xlarge it takes 5m + time for installation.
On my local MacBook it takes around ~(3 mins 20 sec) as compared to ~(1 min 50 sec) without this change
As for "Add a sample warmup code in UT” : we could add a sample script in or before UT which will execute a sample model on GPU (outside TorchServe) which will warmup the GPU before actually starting the UT execution. As per our observation it takes around 2 minutes time for GPU warmup with cudatoolkit 10.1 subsequent uninstall and re-install on GPU machine with timeout changes takes ~3 mins (as GPUs have warmed up in the first install) , which is same as that on CPU.
The text was updated successfully, but these errors were encountered: