https://huggingface.co/docs/transformers/llm_tutorial#wrong-padding-side
LLMs, or Large Language Models, are the key component behind text generation. In a nutshell, they consist of large pretrained transformer models trained to predict the next word (or, more precisely, token) given some input text. Since they predict one token at a time, you need to do something more elaborate to generate new sentences other than just calling the model — you need to do autoregressive generation.
Autoregressive generation is the inference-time procedure of iteratively calling a model with its own generated outputs, given a few initial inputs. In 🤗 Transformers, this is handled by the generate() method, which is available to all models with generative capabilities.
This tutorial will show you how to:
Before you begin, make sure you have all the necessary libraries installed:
pip install transformers bitsandbytes>=0.39.0 -q
A language model trained for causal language modeling takes a sequence of text tokens as input and returns the probability distribution for the next token.
"Forward pass of an LLM"
A critical aspect of autoregressive generation with LLMs is how to select the next token from this probability distribution. Anything goes in this step as long as you end up with a token for the next iteration. This means it can be as simple as selecting the most likely token from the probability distribution or as complex as applying a dozen transformations before sampling from the resulting distribution.
"Autoregressive generation iteratively selects the next token from a probability distribution to generate text"
The process depicted above is repeated iteratively until some stopping condition is reached. Ideally, the stopping condition is dictated by the model, which should learn when to output an end-of-sequence (EOS) token. If this is not the case, generation stops when some predefined maximum length is reached.
Properly setting up the token selection step and the stopping condition is essential to make your model behave as you’d expect on your task. That is why we have a GenerationConfig file associated with each model, which contains a good default generative parameterization and is loaded alongside your model.
Let’s talk code!
If you’re interested in basic LLM usage, our high-level Pipeline interface is a great starting point. However, LLMs often require advanced features like quantization and fine control of the token selection step, which is best done through generate(). Autoregressive generation with LLMs is also resource-intensive and should be executed on a GPU for adequate throughput.