Fix maximum value check for token_batch_size to allow larger values #130
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
smarter
force-pushed
the
tbs
branch
4 times, most recently
from
8942c43 to
f677706
Compare
With a recent pytorch we get:
bergson/process_preconditioners.py:59:68 - error: Argument of type "ProcessGroup | Unknown | int" cannot be assigned to parameter "group" of type "ProcessGroup | None"
Type "ProcessGroup | Unknown | int" is not assignable to type "ProcessGroup | None"
Type "int" is not assignable to type "ProcessGroup | None"
"int" is not assignable to "ProcessGroup"
"int" is not assignable to "None" (reportArgumentType)
The non ProcessGroup case should only happen when the `ranks` argument is passed
explicitly to `new_group`
Also upgrade pyright in the CI although the problem is unrelated.
Previously, token_batch_size was limited by max_position_embeddings, but this is the maximum length of one sequence, and a batch can contain multiple sequences. As far as I can tell, the real limit is the default value of model_max_length in the tokenizer (some models have model_max_length == max_position_embeddings which causes this confusion). We also make the truncation logic more explicit by passing a max_length parameter instead of mutating model_max_len.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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Outdated suggestions cannot be applied.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.
Previously, token_batch_size was limited by max_position_embeddings, but this is the maximum length of one sequence, and a batch can contain multiple sequences. As far as I can tell, the real limit is the default value of model_max_length in the tokenizer (some models have model_max_length == max_position_embeddings which causes this confusion).
We also make the truncation logic more explicit by passing a max_length parameter instead of mutating model_max_len.