4 ms·
The two ways are equivalent (it's always next token prediction) but the latter is way more efficient as it computes the loss for N tokens in a single forward pa
by blackbear_ 1y ago
The two ways are equivalent (it's always next token prediction) but the latter is way more efficient as it computes the loss for N tokens in a single forward pass.