BPE merge by Unicode scalar, not grapheme cluster (#352, Bug 4)#355
Open
apocryphx wants to merge 1 commit into
Open
BPE merge by Unicode scalar, not grapheme cluster (#352, Bug 4)#355apocryphx wants to merge 1 commit into
apocryphx wants to merge 1 commit into
Conversation
, Bug 4) `BPETokenizer.bpe(token:)` decomposed the input into initial BPE symbols via `Array(token).map { String($0) }`, which iterates Swift `Character` (extended grapheme clusters). Non-spacing combining marks that the BPE vocab and merge table treat as standalone scalars — e.g. Thai vowel mark U+0E31, Devanagari halant U+094D, the variation selector + combining keycap behind emoji keycaps — therefore fused with their base character into a single symbol, preventing the merge loop from ever considering the combining-mark scalar as its own atom. Llama-7B's `byte_fallback: true` then byte-encoded the unmatched fused symbol, even though both halves were direct vocab entries. There were actually two grapheme-cluster traps in the BPE path: 1. The initial symbol decomposition at the top of `bpe(token:)` — switched to `token.unicodeScalars.map { String($0) }`, plus the early-return guard for trivial inputs counts scalars rather than `Character`s. 2. The downstream re-split of the BPE result. `bpe(token:)` returned its pieces joined by ASCII space, and the caller re-split with `bpe(token: text).split(separator: " ").map { String($0) }`. But Swift's `split` operates on grapheme clusters: if a piece begins with a non- spacing mark, the preceding ASCII space fuses with the mark into a single grapheme and `split` silently swallows the boundary. The merged substring (including the literal space) then byte-fallbacks. To remove the round-trip entirely, `bpe(token:)` now returns `[String]` directly. Adds a regression test using huggyllama/llama-7b over the Thai input "สวัส". Pre-fix: byte-fallback on `ว` and `ั`; post-fix: 4 direct vocab matches plus the leading `▁`, matching HF Python. The benchmark test in Tests/Benchmarks/BPETokenizerBenchmarkTests.swift calls `bpe(token:)` via `_ = model.bpe(token: encoded)`, so the return-type change does not affect it. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
pcuenca
approved these changes
May 15, 2026
Member
pcuenca
left a comment
pcuenca
left a comment
There was a problem hiding this comment.
Thank you! Suggesting a couple of minor nits.
Comment on lines
+246
to
+252
| // Iterate Unicode scalars rather than grapheme clusters. Combining marks | ||
| // (e.g. Thai vowel U+0E31, Devanagari halant U+094D, the variation selector | ||
| // and combining keycap U+FE0F + U+20E3) form a single Swift `Character` | ||
| // with their base, but the BPE vocab and merge table are scalar-indexed — | ||
| // grouping them prevents merges and forces spurious byte-fallback even | ||
| // when both scalars are direct vocab entries. | ||
| // Reference: https://github.com/huggingface/swift-transformers/issues/352 |
Member
There was a problem hiding this comment.
Suggested change
| // Iterate Unicode scalars rather than grapheme clusters. Combining marks | |
| // (e.g. Thai vowel U+0E31, Devanagari halant U+094D, the variation selector | |
| // and combining keycap U+FE0F + U+20E3) form a single Swift `Character` | |
| // with their base, but the BPE vocab and merge table are scalar-indexed — | |
| // grouping them prevents merges and forces spurious byte-fallback even | |
| // when both scalars are direct vocab entries. | |
| // Reference: https://github.com/huggingface/swift-transformers/issues/352 |
| // Reference: https://github.com/huggingface/swift-transformers/issues/352 | ||
| let initialSymbols = token.unicodeScalars.map { String($0) } | ||
| if initialSymbols.count <= 1 { | ||
| return [token] |
Member
There was a problem hiding this comment.
Suggested change
| return [token] | |
| return symbols.isEmpty ? [] : [token] |
Technically, if the input is "" I think we should return an empty array of pieces rather than an array containing one empty piece. Both will result in the same thing, but the former skips the attempt to tokenize the empty piece.
Using symbols instead of initialSymbols as per the related comment.
| var symbols = Array(token).map { String($0) } | ||
| // Initial symbols: one entry per Unicode scalar of `token`. We keep these | ||
| // as a doubly linked list embedded in parallel arrays of indices. | ||
| var symbols = initialSymbols |
Member
There was a problem hiding this comment.
initialSymbols is only used for the isEmpty test and immediately reassigned. I'd go for var symbols = token.unicodeScalars.map { String($0) } above and remove initialSymbols.
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.
Addresses Bug 4 of #352.
BPETokenizer.bpe(token:)decomposed the input into initial BPE symbols viaArray(token).map { String(\$0) }, which iterates SwiftCharacter(extended grapheme clusters). Non-spacing combining marks that the BPE vocab and merge table treat as standalone scalars — e.g. Thai vowel mark U+0E31, Devanagari halant U+094D, the variation selector + combining keycap behind emoji keycaps — therefore fused with their base character into a single symbol, preventing the merge loop from ever considering the combining-mark scalar as its own atom. Llama-7B'sbyte_fallback: truethen byte-encoded the unmatched fused symbol, even though both halves were direct vocab entries.There were actually two grapheme-cluster traps in the BPE path:
The initial symbol decomposition at the top of
bpe(token:)— switched totoken.unicodeScalars.map { String(\$0) }, plus the early-return guard for trivial inputs counts scalars rather thanCharacters.The downstream re-split of the BPE result.
bpe(token:)returned its pieces joined by ASCII space, and the caller re-split withbpe(token: text).split(separator: " ").map { String(\$0) }. But Swift'ssplitoperates on grapheme clusters: if a piece begins with a non-spacing mark, the preceding ASCII space fuses with the mark into a single grapheme andsplitsilently swallows the boundary. The merged substring (including the literal space) then byte-fallbacks. To remove the round-trip entirely,bpe(token:)now returns[String]directly.The benchmark test in
Tests/Benchmarks/BPETokenizerBenchmarkTests.swiftcallsbpe(token:)via_ = model.bpe(token: encoded), so the return-type change does not affect it.Test plan
swift test --filter TokenizerTests— all 46 tests pass (no regressions).llama7bCombiningMarks()useshuggyllama/llama-7bover Thai"สวัส". Pre-fix: byte-fallback onวandั; post-fix: 4 direct vocab matches plus the leading▁, matching HF Pythontransformers==4.57.1byte-for-byte.🤖 Generated with Claude Code