[Tokenizer] Add encode ordinary mode (no special token matching)#23830
Merged
jtuyls merged 1 commit intoiree-org:mainfrom Apr 3, 2026
Merged
[Tokenizer] Add encode ordinary mode (no special token matching)#23830jtuyls merged 1 commit intoiree-org:mainfrom
jtuyls merged 1 commit intoiree-org:mainfrom
Conversation
This was referenced Mar 19, 2026
benvanik
requested changes
Mar 25, 2026
benvanik
pushed a commit
that referenced
this pull request
Mar 25, 2026
#23852) ## Summary When text preceded back-to-back special tokens (e.g. <|user|>Hi<|end|><|assistant|>), the first was dropped because matching continued while a deferred token was pending, overwriting the single pending_special_token slot. This breaks every multi-turn chat prompt. Add the missing pending < 0 guard to skip matching until the deferred token is emitted. ## Test results ### HuggingFace smoketest (`huggingface_smoketest.py`) **1667/1667** tokenization comparisons pass across ~80 HuggingFace models (0 mismatches). 76 additional tests fail to **load** (not tokenization mismatches) — these are tiktoken models listed in the HF smoketest that aren't valid HuggingFace model identifiers. They're tested by the dedicated tiktoken smoketest instead. This is fixed in #23830. ### Tiktoken smoketest (`tiktoken_smoketest.py`) **72/76** tokenization comparisons pass across 4 tiktoken encodings (cl100k_base, o200k_base, r50k_base, p50k_base). 4 failures — **identical between upstream and this PR** (pre-existing, same as `fix-tokenizer-added-tokens` branch): | Encoding | Failing test | Cause | |----------|-------------|-------| | cl100k_base | `special_token_endoftext` | IREE matches `<\|endoftext\|>` as special token; tiktoken's `encode_ordinary` treats it as literal text | | o200k_base | `special_token_endoftext` | Same | | r50k_base | `special_token_endoftext` | Same | | p50k_base | `special_token_endoftext` | Same | Root cause: the IREE tokenizer has no "encode ordinary" mode (equivalent to tiktoken's `disallowed_special=()`). Fixed in: #23830. Signed-off-by: Jorn <jorn.tuyls@gmail.com> Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Add IREE_TOKENIZER_ENCODE_FLAG_NO_SPECIAL_TOKEN_MATCHING to skip special token recognition in input text. Equivalent to tiktoken's encode_ordinary(). Fixes 4/76 tiktoken smoketest failures. Also skips tiktoken models in the HuggingFace smoketest (tested by the dedicated tiktoken smoketest instead). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> Signed-off-by: Jorn <jorn.tuyls@gmail.com>
jtuyls
force-pushed
the
fix-tiktoken-encode-ordinary
branch
from
905f4a3 to
a5ec2c9
Compare
benvanik
approved these changes
Apr 3, 2026
jtuyls
added a commit
that referenced
this pull request
Apr 3, 2026
…23828) ## Summary Return RESOURCE_EXHAUSTED instead of silently truncating when the output buffer is too small, matching the documented API contract in tokenizer.h. ## Test results ### HuggingFace smoketest (`huggingface_smoketest.py`) **1667/1667** tokenization comparisons pass across ~80 HuggingFace models (0 mismatches). 76 additional tests fail to **load** (not tokenization mismatches) — these are tiktoken models listed in the HF smoketest that aren't valid HuggingFace model identifiers. They're tested by the dedicated tiktoken smoketest instead. These tests are removed from the HF smoketest in #23830. ### Tiktoken smoketest (`tiktoken_smoketest.py`) **72/76** tokenization comparisons pass across 4 tiktoken encodings (cl100k_base, o200k_base, r50k_base, p50k_base). 4 failures — **identical between upstream and this PR** (pre-existing, same as `fix-tokenizer-added-tokens` branch): | Encoding | Failing test | Cause | |----------|-------------|-------| | cl100k_base | `special_token_endoftext` | IREE matches `<\|endoftext\|>` as special token; tiktoken's `encode_ordinary` treats it as literal text | | o200k_base | `special_token_endoftext` | Same | | r50k_base | `special_token_endoftext` | Same | | p50k_base | `special_token_endoftext` | Same | Root cause: the IREE tokenizer has no "encode ordinary" mode (equivalent to tiktoken's `disallowed_special=()`). Fixed in #23830. Signed-off-by: Jorn <jorn.tuyls@gmail.com> Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
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.
Add IREE_TOKENIZER_ENCODE_FLAG_NO_SPECIAL_TOKEN_MATCHING to skip special token recognition in input text. Equivalent to tiktoken's encode_ordinary(). Fixes 4/76 tiktoken smoketest failures. Also skips tiktoken models in the
HuggingFace smoketest (tested by the dedicated tiktoken smoketest instead).
Test Results: