what's skip edges
is better or worse?
is better or worse?
Should be better because the edge layers are unquantized.
is better or worse?
The difference is simply which blocks receive INT8 ConvRot treatment. In the fully treated INT8 version, all eligible transformer blocks are quantized, including the edge blocks. In the skip-edges version, the first and last blocks are left at their original higher precision while the main body of the model still uses INT8 ConvRot. This slightly increases VRAM/model size, but reduces the chance of quantization error being introduced at the model's sensitive input and output boundaries, so skip-edges is the more conservative quality-focused variant.
Does skip edges have the same speed as the non-skip?
Does skip edges have the same speed as the non-skip?
Should be exactly the same.
Does skip edges have the same speed as the non-skip?
Should be exactly the same.
Thanks for confirming this. All this technical jargon is really hard to follow at times.