-
Notifications
You must be signed in to change notification settings - Fork 247
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Use regular division inside Scale Estimation #3210
Use regular division inside Scale Estimation #3210
Conversation
tinyllama_scale_estimation_per_channel_backend_TORCH: | ||
metric_value: 0.81389 | ||
metric_value: 0.80873 | ||
num_int4: 188 | ||
num_int8: 124 | ||
atol: 0.006 # difference across devices: 0.80873 vs 0.81389 |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
A reference value for PT is changed to stay aligned with OV. Interestingly, the new value for OV (0.80873) is also the value that PT backend gets on some devices.
@nikita-savelyevv , please provide time compression data with these changes before we merge it |
Added |
Changes
Compute division inside SE algorithm always as
a/b
instead ofa*(1/b)
in some cases.Reason for changes
During implementation #2727 some choices were made regarding how division operation is computed in order for the changes to be completely aligned with the previous implementation. Namely, before #2727 some divisions were computed as
a*(1/b)
, and this is currently still the case.The way these divisions are computed originally was not intended. Now, all divisions are aligned to the
a/b
form.Compression time and memory are roughly the same.
*time column includes PT -> OV conversion time.
Related tickets
163286
Tests