Releases: Dan-wanna-M/kbnf
Releases · Dan-wanna-M/kbnf
v0.3.9-python
- Bug fix: incorrect logits mask will no longer be produced when multiple different language models' vocabulary with the same logits length are used to create engines.
v0.5.3-rust
- Bug fix:
#substrs
now correctly constrain the allowed tokens
v0.5.2-rust
- Bug fix: undefined start nonterminal in kbnf grammar string will no longer panick. Instead, an error will be returned.
v0.3.8-python
- Changes from
v0.5.3-rust
v0.3.7-python
- Changes from
v0.5.2-rust
v0.5.1-rust
- Add
#substrs
docs. max_state_id
is now correctly computed with respect to#substrs
- Make Python FFI docs safety section more comprehensive
v0.5.0-rust
- Make
EngineLike
a sealed trait so adding more methods to it will not break compatibility. - Support new syntax:
#substrs"xxxxx"
which supports generating one substring out of a long string.
v0.3.6-python
- The case of all tokens are allowed is now handled without UB in the fast path of
mask_logits
for torch tensor - Add safety checks to guard against the fast path of
mask_logits
somask_logits
does not lead to UB on platforms whereusize
is not 64 bit. - Update docs to align with implementations of
Engine
- Changes from
v0.5.1-rust
v0.3.5-python
- Make torch
mask_logits
significantly faster by bypassing unreasonably slow Python interpreter & allowed token ids index->allowed token ids CPU tensor cache - Changes from
v0.5.0-rust
v0.3.4-python
- Make
compute_allowed_token_ids
no-gil. In other words, it no longer blocks python threads. - Changes from
v0.4.3-rust