transformers
5b6f170a - Better and more extensive tests for RoPE (#46912)

Commit
30 days ago
Better and more extensive tests for RoPE (#46912) * fix * test * tmp * ! * fix * . * multimodal rope as well, just copy * expectations * missing device in init, should follow Llama for text rope! * Revert "fix" This reverts commit 7023cfa6ed835ad2c9319e0ce3246e247630cccd. * Revert "test" This reverts commit ab6e25ebc54965b4098f9fb2deaaae6b822b3f1d. * delete and move a lot of things * and this, should not skip * fix a bunch * small batch of fixes * skip a bunch of audio models, they need special/fixes audio sizes and can't really support scaled rope * fix copies * now fix? * check text config * dont skip, was fixed at some point magically * more comments * Add more nesting! add per layer config though we defi need to merge it with nested rope, that is too many fallbacks * kmp, per-layer config is not compatible with DS4 * wtf?!
Author
Parents
Loading