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
We are currently experimenting with the 12.8 billion GPT NeoX model using deepspeed inference (init_inference) on an A100 device.
When inferred with the float16 type, the average per-token generation latency was measured at 24ms, whereas with the int8 type, it was measured at 54ms.
Why is the inference speed of a quantized model with int8 slower?
As far as I know, I understand that operation efficiency can increase due to bit reduction. However, is it not the case in reality?
reacted with thumbs up emoji reacted with thumbs down emoji reacted with laugh emoji reacted with hooray emoji reacted with confused emoji reacted with heart emoji reacted with rocket emoji reacted with eyes emoji
-
We are currently experimenting with the 12.8 billion GPT NeoX model using deepspeed inference (
init_inference) on an A100 device.When inferred with the float16 type, the average per-token generation latency was measured at 24ms, whereas with the int8 type, it was measured at 54ms.
Why is the inference speed of a quantized model with int8 slower?
As far as I know, I understand that operation efficiency can increase due to bit reduction. However, is it not the case in reality?
Beta Was this translation helpful? Give feedback.
All reactions