Fix slow test_moshika_greedy_unconditional_fp16 #39251
Open
+13
−31
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Contributor
Author
|
run-slow: moshi |
Contributor
|
This comment contains run-slow, running the specified jobs: models: ['models/moshi'] |
|
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. |
eustlb
reviewed
Sep 2, 2025
Contributor
eustlb
left a comment
eustlb
left a comment
There was a problem hiding this comment.
Thanks for the update @manueldeprada, still some work to come to fix moshi...
Is this PR still relevant concerning changes to cache_utils.py and generation/utils.py ? If not can you solve merging conflicts?
Comment on lines
1863
to
-1886
| def _prepare_attention_mask_for_generation( | ||
| self, | ||
| input_ids: torch.LongTensor, | ||
| inputs_tensor: torch.Tensor, | ||
| generation_config: GenerationConfig, | ||
| kwargs: dict[str, Any], | ||
| model_kwargs: dict[str, Any], | ||
| ) -> torch.LongTensor: | ||
| pad_token_id = generation_config.pad_token_id | ||
| eos_token_id = generation_config.eos_token_id | ||
|
|
||
| default_attention_mask = torch.ones(input_ids.shape, dtype=torch.long, device=input_ids.device) | ||
| if pad_token_id is None: | ||
| return default_attention_mask | ||
|
|
||
| is_pad_token_in_inputs = (pad_token_id is not None) and torch.isin(input_ids, pad_token_id).any() | ||
| is_pad_token_not_equal_to_eos_token_id = (eos_token_id is None) or ~torch.isin( | ||
| eos_token_id, pad_token_id | ||
| ).any() | ||
| can_infer_attention_mask = is_pad_token_in_inputs * is_pad_token_not_equal_to_eos_token_id | ||
| attention_mask_from_padding = input_ids.ne(pad_token_id).long() | ||
|
|
||
| attention_mask = ( | ||
| attention_mask_from_padding * can_infer_attention_mask + default_attention_mask * ~can_infer_attention_mask | ||
| return super()._prepare_attention_mask_for_generation( | ||
| inputs_tensor=inputs_tensor, | ||
| generation_config=generation_config, | ||
| model_kwargs={}, | ||
| ) | ||
| return attention_mask |
Contributor
There was a problem hiding this comment.
can't we just remove _prepare_attention_mask_for_generation override here?
| slicing = torch.arange(max_cache_len, device=value_states.device) | ||
| current_seq_len = cache_position[-1] + 1 # Use last position to determine current length | ||
| to_shift = current_seq_len > max_cache_len | ||
| to_shift = current_seq_len >= max_cache_len |
Contributor
There was a problem hiding this comment.
this will break other models no?
| # - different models have a different cache name expected by the model (default = "past_key_values") | ||
| # - `max_length`, prepared above, is used to determine the maximum cache length | ||
| max_cache_length = generation_config.max_length - 1 | ||
| max_cache_length = generation_config.max_length |
Contributor
There was a problem hiding this comment.
this will break other models no?
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Fix #38725
Coming from #38725, previously, e18f233 attempted to fix the default attention mask issue that appeated with #34464, but it was still failing the slow test
tests/models/moshi/test_modeling_moshi.py::MoshiIntegrationTests::test_moshika_greedy_unconditional_fp16History from git bisect:
transformers/src/transformers/modeling_utils.py
Line 1413 in 84a6789
transformers/src/transformers/generation/utils.py
Line 2090 in 36bf1d2
transformers/src/transformers/cache_utils.py
Line 1740 in 1b22290
Setting cache_implementation="dynamic" makes the test pass, but the sliding window cache should not behave different. I believe this is due to the depth decoder being window 8 by default, but audio is confusing to me.
This PR is not a fix: the modeling code should be changed to accomodate what I highlight in the diff
cc @eustlb @ydshieh