[None][fix] Invalidate cached special-id set in slow convert_ids_to_tokens#16331
Draft
Shixiaowei02 wants to merge 1 commit into
Draft
Conversation
Collaborator
Author
|
/bot run --disable-fail-fast |
Collaborator
|
PR_Github #58990 [ run ] triggered by Bot. Commit: |
Collaborator
|
PR_Github #58990 [ run ] completed with state
|
Shixiaowei02
force-pushed
the
user/xiaoweis/fix-tokenizer-special-ids-cache-staleness
branch
from
July 16, 2026 06:18
4068de7 to
fcfa02f
Compare
Collaborator
Author
|
/bot run --disable-fail-fast |
Collaborator
|
PR_Github #59639 [ run ] triggered by Bot. Commit: |
Collaborator
|
PR_Github #59639 [ run ] completed with state
|
…okens The slow-tokenizer optimization in convert_ids_to_tokens snapshotted all_special_ids into a set at __init__ and reused it forever. If special tokens were (re)registered after the wrapper was constructed (e.g. add_special_tokens during chat-template / adapter setup), the snapshot went stale: the affected ids stopped being skipped and leaked into detokenized output when skip_special_tokens=True -- an accuracy regression. Recomputing all_special_ids on every call is correct but a severe perf regression: convert_tokens_to_ids(all_special_tokens) is expensive on slow tokenizers and runs once per streamed token (measured ~10^2-10^4x slower, superlinear in the number of special tokens). Instead keep the memoized set and rebuild it only when a cheap O(1) fingerprint of the tokenizer's special tokens changes: the count of registered added tokens (covers newly added specials, including additional_special_tokens) plus the ids of the named special tokens (covers reassigning/clearing bos/eos/pad/unk/sep/cls/mask). Steady-state streaming stays O(1) per call while the cache can never be stale. Add regression tests on a genuine slow tokenizer for a late-added special token and a reassigned named special token. Signed-off-by: xiaoweis <39303645+Shixiaowei02@users.noreply.github.com>
Shixiaowei02
force-pushed
the
user/xiaoweis/fix-tokenizer-special-ids-cache-staleness
branch
from
July 17, 2026 06:50
fcfa02f to
00925b2
Compare
Collaborator
Author
|
/bot run --disable-fail-fast |
Collaborator
|
PR_Github #59912 [ run ] triggered by Bot. Commit: |
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 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.
Description
This pull request improves the handling of special tokens in the
TransformersTokenizerwrapper, ensuring that late-registered or reassigned special tokens are correctly excluded during token detokenization, particularly for slow tokenizers. The changes address a bug where the set of special token IDs could become stale, causing incorrect output, and introduce tests to verify the fix.Special Token Handling Improvements
TransformersTokenizer.__init__with a dynamic, memoized cache that updates only when the set of special tokens changes, ensuring correct handling of late-registered or reassigned special tokens. ([1] [2])_special_token_idsmethod to efficiently and correctly compute the set of special token IDs, using a fingerprint to detect changes and avoid stale caches. (tensorrt_llm/tokenizer/tokenizer.pyL342-R376)Testing
_ToySlowTokenizer) and two regression tests to verify thatconvert_ids_to_tokenscorrectly reflects late-registered and reassigned special tokens, preventing stale caches from leaking special tokens into detokenized output. (tests/unittest/llmapi/test_tokenizer_decode.pyR20-R120)Test Coverage
PR Checklist
Please review the following before submitting your PR:
PR description clearly explains what and why. If using CodeRabbit's summary, please make sure it makes sense.
PR Follows TRT-LLM CODING GUIDELINES to the best of your knowledge.
Test cases are provided for new code paths (see test instructions)
If PR introduces API changes, an appropriate PR label is added - either
api-compatibleorapi-breaking. Forapi-breaking, includeBREAKINGin the PR title.Any new dependencies have been scanned for license and vulnerabilities
CODEOWNERS updated if ownership changes
Documentation updated as needed
Update tava architecture diagram if there is a significant design change in PR.
The reviewers assigned automatically/manually are appropriate for the PR.
Please check this after reviewing the above items as appropriate for this PR.
GitHub Bot Help
To see a list of available CI bot commands, please comment
/bot help.