A dataset wrapper for Chronos-2 models. Arguments ---------- inputs Time series data. The input kind is auto-detected from i

This class holds all the configuration parameters to be used by ``ChronosTokenizer`` and ``ChronosModel``.

Apply standardization along the last dimension and optionally apply arcsinh after standardization.

Multi-head Attention Layer

A preprocessed time series input ready for model training/inference.

A generic residual block which can be used for input and output embedding layers

HF transformers-style pretrained model config for Chronos-2.0, based on T5Config. Arguments ---------- d_model Size model's

Applies rotary position embeddings (RoPE) to input tensors. This implementation follows the transformers v5 pattern for RotaryEmbedding classes,

A custom trainer based on transformers Trainer. We need to override the dataloader getters because we handle batching ourselves in a custom d

Dataset wrapper, using a ``ChronosTokenizer`` to turn data from a time series into a HuggingFace-compatible set of ``input_ids``, ``attention

A ``ChronosModel`` wraps a ``PreTrainedModel`` object from ``transformers`` and uses it to predict sample paths for time series tokens.

Callback to evaluate and save the model at last training step.

Self-attention applied along the batch axis masked by the group attention mask

Shuffle entries from an iterable by temporarily accumulating them in an intermediate buffer. Parameters ---------- base_dataset

A ``ChronosPipeline`` uses the given tokenizer and model to forecast input time series. Use the ``from_pretrained`` class method to load

A ``ChronosTokenizer`` defines how time series are mapped into token IDs and back. For details, see the ``input_transform`` and ``output

Mix-in class that datasets can inherit from to get shuffling functionality.