Skip to content

Tokenizers: read the tokenizer.json 'decoder' section in CreateFromTokenizerJson #7660

Description

@ericstj

Background

SentencePieceTokenizer.CreateFromTokenizerJson (added in #7625) loads a Hugging Face tokenizer.json for Unigram models. It reads model, normalizer, pre_tokenizer, added_tokens, and post_processor, but does not read the decoder section.

This was called out in review of #7625: it is fine for the standard HF Unigram/Metaspace configurations the current tests cover (decoding is driven by the SentencePiece model semantics), but it means non-standard decoder chains in a tokenizer.json are ignored rather than honored.

Ask

Support reading the tokenizer.json decoder section when a real model requires a decoder chain that differs from the implied SentencePiece/Metaspace decoding, e.g. Sequence, Replace, Strip, ByteFallback, Metaspace, Fuse decoders. Until then, the loader silently relies on the model''s built-in decode behavior.

Notes

Filed as a follow-up to review feedback on #7625.

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or requestuntriagedNew issue has not been triaged

    Type

    No type

    Fields

    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions