Skip to content

Latest commit

 

History

History
155 lines (112 loc) · 6.53 KB

CHANGELOG.md

File metadata and controls

155 lines (112 loc) · 6.53 KB

LM Format Enforcer Changelog

v0.10.9

v0.10.8

  • Fixed support for tokenizers with larger vocabularies than the models. Enables visual models such as Llama3.2
  • JsonSchemaParser: Fixed additionalProperies and oneOf edge cases

v0.10.7

  • [135] Updated Haystack V2 integration with latest API

v0.10.6

  • Pickling improvements for easier multiprocessing support

v0.10.5

  • SequenceParser performance optimization
  • JsonSchemaParser: number parsing supports exponents
  • Supporting tokenizers with multiple eos token ids

v0.10.4

  • Added default max Json array length to help LLMs avoid infinite loops. See README for details.
  • Updated EXLlamaV2 example to updated API

v0.10.3

  • [#113] TRTLLM Support: Fixing type incompatibility in certain cases / library versions

v0.10.2

  • [#100] JsonSchemaParser: Added allOf support
  • [#99] JsonSchemaParser: Fixed edge case that would allow leading comma in JSON Array
  • [#102] JsonSchemaParser: Fixed Array of Enums not producing multiple values

v0.10.1

  • Allowing control of LM Format Enforcer's heuristics via env var / configuration objects. See the 'Configuration options' section of the README.

v0.9.10

  • [#95] Added anyOf support to JsonSchemaParser, making function calls possible.

v0.9.9

  • Updated README with vLLM OpenAI Server Inference integration

v0.9.8

  • [#80] JSONSchemaParser List would allow opening comma before first element if there was a whitespace before it

v0.9.7

  • [#93] Improved JSONSchemaParser performance, unit tests run twice as fast! Joint effort with Ari Weinstein. Thanks!

v0.9.6

  • [#88] ExllamaV2 optimizations
  • Bugfix in ExllamaV2 sample notebook that generated garbage data after the response.

v0.9.5

  • [#87] Allow regex to exit to forcestopparser when receiving a pad/eos token after having been in a final state. Thanks Josh C!

v0.9.4

  • [#27] Improving vLLM class support (AsyncLLMEngine etc)

v0.9.3

  • #83 - Supporting added model tokens
  • Improved support for out-of-token-vocabulary characters

v0.9.2

  • #80 - Fixed bug where comma could start json list
  • #34 - Fixed llama-cpp-python low max_tokens default in sample

v0.9.1

  • Fixed build errors in certain edge cases

v0.9.0

  • #68 Added NVIDIA TensorRT-LLM Support, NVIDIA's contribution by Ahmet Erdem. Thanks!
  • Much faster TokenizerData initialization, new JSON freetext token caching algorithm.
  • More robust error reporting.

v0.8.3

  • #67 Updating vLLM integration to support v0.3.0
  • #63 JSONSchemaParser: Empty list cannot be closed after a newline

v0.8.2

Several JsonSchemaParser improvements:

  • #32 Added limited support for regex-constrained string in JSON using the pattern field. See test_phone_number_in_string().
  • #54 Fixed regression bug caused by limited-length JSON string caching.
  • #53 Fixed problems with arrays of union types.

v0.8.1

  • Performannce improvement: Limited-length JSON strings will also enjoy caching. Thanks Jarno Elonen for the contribution!

v0.8.0

  • Performance improvement: Introduced TokenEnforcerTokenizerData that allows reusing the tokenizer preprocessing data between different TokenEnforcer instances. The sample notebooks have been updated to take advantage of this option.
  • Performance improvement: Long sequences will see up to 5x TokenEnforcer runtime footprint reduction.

v0.7.3

  • Bug fixes

v0.7.2

  • vLLM performance improvements
  • Sample notebooks don't require huggingface account anymore

v0.7.1

v0.7.0

  • JSON Schema: Added support for union types. In pydantic, both key: int | str and key: Union[int, str] formats are supported
  • JSON Schema: Added support for schemaless JSON mode. JsonSchemaParser(None) will now create a parser that accepts any valid JSON.

v0.6.5

  • Added official vLLM integration that doesn't require monkey patching.

v0.6.4

  • JSON Schema : Supports string min/max length limitation

v0.6.3

  • Community PR: Fixed SequenceParser bug

v0.6.2

  • Added haystack integration

v0.6.1

  • Fixed llama.cpp integration to be able to generate unicode characters in json freetext fields
  • Fixed unescaped newlines being allowed in json freetext fields

v0.6.0

  • RegexParser and JsonSchemaParser can now output all of the characters that exist in the tokenzier
  • Added "Known issues and limitations" section to the README
  • Fixed a bug in JsonSchemaParser where sometimes illegal commas were allowed

v0.5.2

  • JSON Schema : Supports empty arrays and escape characters in strings
  • Regex : Performance improvement in some cases
  • Added UnionParser and SequenceParser to allow combining parsers

v0.5.1

  • Made it easier to report bugs in the library

v0.5.0

  • Introduced FormatEnforcerAnalyzer to allow all inference engines to be analyzed in a unified way. (Was previously only available for transformers)
  • Added support for the analyser in llama.cpp, updated example notebook
  • JsonSchemaParser now take list min/max items into consideration

v0.4.3

  • Improved JsonSchemaParser whitespace support
  • Improved RegexParser performance, especially in regular expressions with .+ and .* sections.

v0.4.2

  • Modified example in main README to be able to run end to end in Google Colab

v0.4.1

  • Added integration with the LlamaIndex library (huggingface and llama.cpp backends) via sample notebook.

v0.4.0

  • Introduced lmformatenforcer.integrations module which will have the integrations with inference engines.
  • Added llama-cpp-python integration to the library in lmformatenforcer.integrations.llamacpp
  • Breaking API Change: Moved 'build_transformers_prefix_allowed_tokens_fn', 'generate_enforced' from lmformatenforcer to lmformatenforcer.integrations.transformers.