mlc-ai/xgrammar v0.2.6rc2

v0.2.6rc2 (preview)

on GitHub

Preview release for early testing. This release is cut from the preview/v0.2.6rc2 branch, not from main.

Highlights over v0.2.6rc1:

• Dynamic JSON Schema compilation and token-mask generation optimizations (#846), including reuse of vocabulary common prefixes in token mask generation (#856).
• JSON Schema string length constraints (minLength / maxLength) no longer cause per-step token-mask slowdowns or latency cliffs (#853, fixing #805 and #852).
• Generalized dynamic token-boundary masks beyond JSON strings, with correctness proofs replacing the earlier shape heuristics (#854).
• Correctness fixes: mask cache key, FSM aliasing, and lazy metadata race (#855); JSON string masking issues (#856); repeat-edge auxiliary data preserved in JSON string regex FSMs (#864, fixing #863); local-completion crossing-token bitset preserved on the lazy compilation path and rule-own lookahead assertions enforced in local-completion masks (#867).
• Documentation for the dynamic compilation option (#843) and a multi-version documentation site.

Install with:

pip install xgrammar==0.2.6rc2

Note: pip does not pick up pre-releases by default; pin the exact version as above or use pip install --pre xgrammar.