Fix nans in attention, add pad in tokenizer if missing #269

gsarti · 2024-04-29T10:13:46Z

Description

Fix the issue in the attention implementation from Why sometimes the target attribution can be 'nan'? #268 where non-terminal position in the tensor were set to nan if they were 0s
Fix the pad token in cases where it is not specified by default in the loaded model

Fix nans in attn, add pad if missing

2581de5

gsarti mentioned this pull request Apr 29, 2024

Why sometimes the target attribution can be 'nan'? #268

Closed

1 task

gsarti merged commit 8de5e29 into main Apr 29, 2024
3 checks passed

LuukSuurmeijer pushed a commit to LuukSuurmeijer/inseq that referenced this pull request May 10, 2024

Fix nans in attn, add pad if missing (inseq-team#269)

9168e39

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Fix nans in attention, add pad in tokenizer if missing #269

Fix nans in attention, add pad in tokenizer if missing #269

gsarti commented Apr 29, 2024 •

edited

Fix nans in attention, add pad in tokenizer if missing #269

Fix nans in attention, add pad in tokenizer if missing #269

Conversation

gsarti commented Apr 29, 2024 • edited

Description

gsarti commented Apr 29, 2024 •

edited