pyvene.models.llama.modelings_intervenable_llama

Each modeling file in this library is a mapping between abstract naming of intervention anchor points and actual model module defined in the huggingface library.

We also want to let the intervention library know how to config the dimensions of intervention based on model config defined in the huggingface library.

Module Attributes

llama_type_to_dimension_mapping	llama model with LM head
llama_lm_type_to_dimension_mapping	llama model with classifier head

Functions

create_llama([name, cache_dir, dtype, config])

Creates a LLaMA Causal LM model, config, and tokenizer from the given name and revision