github mlc-ai/xgrammar v0.2.9

46 minutes ago

What's Changed

  • feat: support Gemma 4 tool calling with the <|"|> string delimiter by @asanomitakanori in #923
  • fix: keep built-in formats within valid JSON strings by @RanaPriyansh in #927
  • fix: bound DeepSeek V4.1 auto tool calls by @mashiro01 in #936
  • fix: preserve shortest JSON numbers in schema literals by @RanaPriyansh in #926
  • perf: cut bounded-string token-mask fills by classifying the repetition body with its parent by @0z5a in #903
  • perf: extend the counted-repetition fast path to strings with minLength by @Seven-Streams in #938
  • fix: fix the rendering of xgrammar's version on the website. by @Seven-Streams in #941
  • ci: remove the nightly Routine Unit Test workflow by @Seven-Streams in #944
  • fix: rule-level cache hash collisions between different character ranges and repeat bounds by @Seven-Streams in #943
  • build: update dependencies by @Seven-Streams in #950
  • fix: resolve JSON Schema references through array indices by @adimalkar in #929
  • fix: let a token that starts in a literal finish in a token-level free-text region by @Seven-Streams in #942
  • [Python] Add opt-in token_markers to get_model_structural_tag by @JRJRJPRO in #940
  • fix: preserve separators in generic JSON rules by @RanaPriyansh in #948
  • fix: fix the benchmark script by @Seven-Streams in #962
  • fix: compile large regex repetitions with caching enabled by @Seven-Streams in #963
  • fix: fix the import of mlxlm. by @Seven-Streams in #976
  • Add jump-forward tokens while preserving accepted boundaries by @Ubospica in #980
  • fix: match the ECMAScript whitespace class in \s and \S by @jinwukong in #977
  • fix: CI Hugging Face rate limiting with shared offline tokenizer assets by @a68lin in #982
  • Bugfix: avoid collisions in JSON Schema cache keys by @izhuhaoran in #985
  • docs: document any_whitespace interaction with permissive schemas by @davidchung29 in #953
  • fix: release erased values in List so cache evictions free memory by @stu-cao in #1003
  • feature: Add opt-in JIT grammar compilation by @a68lin in #1017
  • fix: accept stop token when it completes the grammar by @n0gu-furiosa in #710
  • feat: support Lark syntax in structural tag grammar formats by @BugenZhao in #1015
  • fix: keep find_jump_forward_string to whole UTF-8 characters by @Shakhtar-Sankur in #1012
  • ci: migrate macOS 14 runners to macOS 15 by @a68lin in #1044
  • fix: preserve forbidden bytes across counted repeat lowering by @HengYangDS in #949

New Contributors

Full Changelog: v0.2.8...v0.2.9

Don't miss a new xgrammar release

NewReleases is sending notifications on new releases.