Fix cache-related tests#39676
Conversation
|
run-slow: kyutai_speech_to_text, musicgen_melody, qwen2_5_omni, rag, roformer, superglue |
|
This comment contains run-slow, running the specified jobs: models: ['models/kyutai_speech_to_text', 'models/musicgen_melody', 'models/qwen2_5_omni', 'models/rag', 'models/roformer', 'models/superglue'] |
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
|
run-slow: kyutai_speech_to_text, musicgen, musicgen_melody, rag, roformer, superglue |
|
This comment contains run-slow, running the specified jobs: models: ['models/kyutai_speech_to_text', 'models/musicgen', 'models/musicgen_melody', 'models/rag', 'models/roformer', 'models/superglue'] |
|
[For maintainers] Suggested jobs to run (before merge) run-slow: kyutai_speech_to_text, llava_next, llava_next_video, musicgen, musicgen_melody, qwen2_5_omni, qwen2_5_vl, qwen2_vl, rag, roformer, superglue |
ydshieh
left a comment
ydshieh
left a comment
There was a problem hiding this comment.
Not very familiar with some changes in src/transformers so I wll rely on CI and trust you. Just a few nit questions to get some idea.
Thank you!
Could you link the PR you mentioned in this PR description?
There was a problem hiding this comment.
do you know what causse this changes?
There was a problem hiding this comment.
yep, after #39374 we started using higher default max length. Text-only generation pipe already uses it, so it's fine
| logits_padded = res_padded.logits[inputs_dict["attention_mask"].bool()] | ||
| logits_padfree = res_padfree.logits[0] | ||
|
|
||
| torch.testing.assert_close(logits_padded.argmax(-1), logits_padfree.argmax(-1), rtol=0, atol=0) |
There was a problem hiding this comment.
ok, can't find it on the common test file, so fine.
There was a problem hiding this comment.
this one doesn't make sense because we want to check logits, not sampled argmax tokens. Even tiny diff in logits can give different tokens, and the next line check with torch.allclose is enough
| # Apply RoPE if self attention | ||
| if not is_cross_attention and sinusoidal_pos is not None: | ||
| if self.rotary_value: | ||
| query_layer, key_layer, value_layer = self.apply_rotary_position_embeddings( | ||
| sinusoidal_pos, query_layer, key_layer, value_layer | ||
| ) | ||
| else: | ||
| query_layer, key_layer = self.apply_rotary_position_embeddings( | ||
| sinusoidal_pos, query_layer, key_layer | ||
| ) |
There was a problem hiding this comment.
was it deleted at some point and here you just add it back?
There was a problem hiding this comment.
yep, accidentally deleted RoPE 🙈
zucchini-nlp
added
the
for patch
* fix * fix kyutai at last * fix unrelated tests and copies * update musicgen as well * revert tensor * fix old test failures * why it wasn't added?
* fix * fix kyutai at last * fix unrelated tests and copies * update musicgen as well * revert tensor * fix old test failures * why it wasn't added?
* fix * fix kyutai at last * fix unrelated tests and copies * update musicgen as well * revert tensor * fix old test failures * why it wasn't added?
* fix * fix kyutai at last * fix unrelated tests and copies * update musicgen as well * revert tensor * fix old test failures * why it wasn't added?
* fix * fix kyutai at last * fix unrelated tests and copies * update musicgen as well * revert tensor * fix old test failures * why it wasn't added?
* fix * fix kyutai at last * fix unrelated tests and copies * update musicgen as well * revert tensor * fix old test failures * why it wasn't added?
* fix * fix kyutai at last * fix unrelated tests and copies * update musicgen as well * revert tensor * fix old test failures * why it wasn't added?
* fix * fix kyutai at last * fix unrelated tests and copies * update musicgen as well * revert tensor * fix old test failures * why it wasn't added?
3 participants
What does this PR do?
As per title, failing after latest cache compatibility PR