Formatron v0.4.9
Formatron empowers everyone to control the output format of language models with minimal overhead.
Loading...
Searching...
No Matches
formatron.integrations.exllamav2 Namespace Reference

This module integrates the ExLlamaV2 library by providing convenience utilities. More...

Classes

class  FormatterFilter
 ExLlamaV2Filter that uses a formatter to mask logits. More...
 

Functions

kbnf.Vocabulary create_engine_vocabulary (ExLlamaV2Tokenizer tokenizer, typing.Optional[list[typing.Callable]] vocab_processors=None)
 Create a vocabulary for the KBNF engine.
 
ExLlamaV2Filter create_formatter_filter (ExLlamaV2 model, ExLlamaV2Tokenizer tokenizer, FormatterBuilder formatter_builder, EngineGenerationConfig engine_config=None, typing.Optional[list[typing.Callable]] vocab_processors=None)
 Create a formatter filter for the ExLlamaV2 engine.
 

Detailed Description

This module integrates the ExLlamaV2 library by providing convenience utilities.

Function Documentation

◆ create_engine_vocabulary()

kbnf.Vocabulary formatron.integrations.exllamav2.create_engine_vocabulary ( ExLlamaV2Tokenizer tokenizer,
typing.Optional[list[typing.Callable]] vocab_processors = None )

Create a vocabulary for the KBNF engine.

Parameters
tokenizerThe tokenizer.
vocab_processorsList of callables with signature (token_to_char: typing.Dict[bytes, bytes])->None. Callables can be used to "unmangle" encoded characters to original characters. If None, processors will be auto-detected.

Definition at line 24 of file exllamav2.py.

◆ create_formatter_filter()

ExLlamaV2Filter formatron.integrations.exllamav2.create_formatter_filter ( ExLlamaV2 model,
ExLlamaV2Tokenizer tokenizer,
FormatterBuilder formatter_builder,
EngineGenerationConfig engine_config = None,
typing.Optional[list[typing.Callable]] vocab_processors = None )

Create a formatter filter for the ExLlamaV2 engine.

Parameters
modelThe ExLlamaV2 model.
tokenizerThe ExLlamaV2 tokenizer.
formatter_builderThe formatter builder.
engine_configThe engine generation configuration.
vocab_processorsList of callables with signature (token_to_char: typing.Dict[bytes, bytes])->None. Callables can be used to "unmangle" encoded characters to original characters. If None, processors will be auto-detected.

Definition at line 46 of file exllamav2.py.