Config for `Evaluator` objects that compute metrics

Implements the beam search algorithm for decoding the most likely sequences. :param end_index: The index of the "stop" or "end" token in the

An error with a configuration file.

OLMo (model) configuration.

Infinitely iterates over a mixture of datasets

Configuration for a dataset to train or evaluate on

Boxes to draw on an image

A :class:`FinalSequenceScorer` which scores the sequences by the average log probability of the tokens in the sequence. It optionally include

An error from incorrect environment variables.

An error with a network request.

A `Sampler` which redistributes the probability mass function for nodes among the top choices with a cumulative probability of at least `p`,

MDPA for the image ViT

AI2D dataset builder, this builder adds the labelled boxes as needed to the AI2D images using both transparent and opaque boxes

CountQa dataset from PaliGemi, it is built by merging the CountBench image/count pairs with the natural language questions from the PaliGemma pape

Configuration for a datset and metrics to use to evaluate a model

Dataset wrapper that supports padding and control the random seed based on the epoch

A :class:`Checkpointer` that saves a single full model and optimizer state dictionary.

MLP used for the image projector

A sharded :class:`Checkpointer` that directly saves the local FSDP flat params data. The optimizer state is saved directly with `torch.save()

Evaluates a model on possibly multiple tasks

A :class:`~torch.distributed.checkpoint.StorageReader` based on :class:`~torch.distributed.checkpoint.FileSystemReader` that can read data di

A `Sampler` which redistributes the probability mass function for nodes among the top `k` choices, then samples from that subset after re-nor

OLMo training configuration.

ChartQA dataset from HuggingFace M4 project. This class loads the ChartQA dataset from HuggingFace (https://huggingface.co/datasets/HuggingFa

Applies prompt templates and adds system prompts to construct text inputs/output

A `Sampler` which uses the Gumbel-Top-K trick to sample without replacement. See [*Stochastic Beams and Where to Find Them: The Gumbel-Top-k

Returned as special metric for visualizing predictions

Evaluates a model on a dataset

The default :class:`LayerNorm` implementation which can optionally run in low precision.

Converts list of examples from our datasets into a tensor batch

Converts text/images inputs into tensors that can be used in the forward method for the a model

An error occurred reading or writing from a checkpoint.

An error from incorrect CLI usage.

A sharded :class:`Checkpointer` that just uses `torch.save()` with extra logic for handling FSDP model and optim state. The world size m

Cache for attention biases and other things that would normally be stored as buffers. We avoid using buffers because we've run into various i

A `Sampler` that just deterministically returns the `k` nodes or beams with highest log probability.

DocumentVQA dataset from HuggingFace M4 project. This class loads the DocumentVQA dataset from HuggingFace (https://huggingface.co/datasets/H

Tokenizer wrapper This exists mostly for legacy reasons since we used to support other kinds of tokenizers with different API

Evaluates the text outputs from a model on a task

Adapted from https://github.com/google/automl/blob/master/lion/lion_pytorch.py