u7122029 / text_gen_calibration

0 stars 0 forks source link

Test that the output logits are the same as the last hidden state when passed through the lm head #18

Closed u7122029 closed 2 months ago

u7122029 commented 2 months ago

To ensure consistency and peace of mind, I want to check that the logits are not different between the following two methods:

  1. Getting logits from .generate() method via output_logits argument,
  2. Getting last hidden state from .generate() then computing logits via .lm_head() method.