3 ms·
Yeah, it seems to "work" in the sense that if you repeat wpe size (1024, 768) three times, so that it becomes (3072, 768), then the model can successfully gener
by sillysaurusx 7y ago
Yeah, it seems to "work" in the sense that if you repeat wpe size (1024, 768) three times, so that it becomes (3072, 768), then the model can successfully generate up to 1024 tokens.
Generating more tokens seems to work up to a point -- you can probably generate up to 1050 tokens with this technique. But at a certain point, more tokens = gibberish.
The cure is to train the new wpe layer the same way you'd train the smaller one. But this also means you don't have to start training from scratch.