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
Hi group,
Great to know TRT-LLM supports INT4 for Llama now! However, it's only for weights with activations in 16-bit. So, here's the question, compared to INT8 SQ where both W and A are in 8-bit, which is faster? Please help me out:)
Which quantization is faster W8A8 or W4A16?
INT4 (W4A16) because weights are in 4-bit and data moving is time consuming!
25%
INT8 (W8A8)
25%
Roughly the same...
0%
Hard to say as it depends on the kernel size, batch_size, seq_len, etc.
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
-
Hi group,
Great to know TRT-LLM supports INT4 for Llama now! However, it's only for weights with activations in 16-bit. So, here's the question, compared to INT8 SQ where both W and A are in 8-bit, which is faster? Please help me out:)
8 votes ·
Beta Was this translation helpful? Give feedback.
All reactions