RMS normalization.

Framework-agnostic config for a residual block.

Options for forecasting. Attributes: max_context: The maximum context length. This is used by the compiled decode function at inference t

RMS normalization.

PyTorch implementation of TimesFM 2.5 with 200M parameters.

Framework-agnostic config for random fourier features.

Residual block with two linear layers and a linear residual connection.

Random Fourier features layer.

Dot syntax accession descriptor.

Configuration for model training. Args: batch_size: Number of samples per batch. num_epochs: Number of training epochs. learnin

Dot syntax accession descriptor.

Framework-agnostic config for a transformer.

Cache for decoding.

Patched time-series decoder.

TimesFM residual block.

Framework-agnostic config for a stacked transformers.

Dataset for time series data compatible with TimesFM.

Handles model training and validation. Args: model: PyTorch model to train. config: Training configuration. rank: Process rank

TimesFM 2.5 with 200M parameters.

Manages distributed training setup and cleanup. Args: world_size: Total number of processes. rank: Process rank. master_addr: A

Multi-head attention.

Multi-head attention.

Per-dimension scaling.

Per-dimension scaling.

Generates position embedding for a given 1-d sequence. Attributes: min_timescale: Start of the geometric index. Determines the periodicity of

Pax rms norm in pytorch.

Rotary positional embedding.

Rotary positional embedding.

Stacked transformer layer.

Borrowed from https://github.com/Nixtla/nixtla/tree/main/experiments/foundation-time-series-arena/xiuhmolpilli/models. We modify the class to tak

Validation dataset using the last window of each series.

Random-window dataset for time series fine-tuning. Pre-samples random (series, split-point) windows similar to Chronos-2's random slicing. E

Data loader class.

Implements the attention used in TimesFM.

Config for initializing timesfm patched_decoder class.

Transformer layer.

Flax implementation of TimesFM 2.5 with 200M parameters.

TimesFM 2.5 with 200M parameters.

Checkpoint used to initialize a TimesFM model for inference. Attributes: version: Version of the checkpoint, e.g. "jax", "torch", "tensorflow",

Hparams used to initialize a TimesFM model for inference. These are the sufficient subset of hparams to configure TimesFM inference agnostic to t

Pax transformer MLP in pytorch.

Weights & Biases implementation of metrics logging. Args: project: Name of the W&B project. config: Configuration dictionary to log.

Helper class for in-context regression covariate formatting. Attributes: targets: List of targets (responses) of the in-context regression.

Helper class for in-context regression covariate formatting. Attributes: targets: List of targets (responses) of the in-context regression.

Linear in-context regression model.

Linear in-context regression model.

Cache for decoding.

Borrowed from https://github.com/Nixtla/nixtla/tree/main/experiments/foundation-time-series-arena/xiuhmolpilli/models.

Layer normalization replica of LayerNorm.

Abstract base class for logging metrics during training. This class defines the interface for logging metrics during model training. Concrete

Model class for finetuning patched time-series decoder. Attributes: core_layer_tpl: config for core layer. freq: freq to finetune on.

Patch decoder layer for time-series foundation model. Attributes: patch_len: length of input patches. horizon_len: length of output patches

Random Fourier features layer.

Residual block with two linear layers and a linear residual connection.

Simple feedforward block with residual connection. Attributes: input_dims: input dimension. hidden_dims: hidden dimension. output_dims:

Tests for the DecodeCache dataclass used in KV-cache decoding.

Tests for ForecastConfig — the primary user-facing configuration.

Tests for linear_interpolation — fills NaN gaps via ``np.interp``.

Tests to verify model instantiation, loading, and compatibility.

Tests for RMS normalization used in transformer attention/FF blocks.

Tests for the random Fourier feature layer.

Tests for RandomFourierFeaturesConfig.

Tests for the residual block: hidden → activation → output + skip.

Tests for ResidualBlockConfig used by tokenizer and output projections.

Tests for the RevIN normalization used in patched decoding.

Tests for strip_leading_nans — removes leading NaN prefix.

Tests for TransformerConfig — architecture-level hyperparameters.

Tests for Welford-style online mean / variance accumulation.

Extract all time covariates except for holidays.

Abstract base class for TimesFM models. Attributes: forecast_config: Configuration for forecasting flags. compiled_decode: Compiled decode

Framework-agnostic config of TimesFM 2.5.

Base TimesFM forecast API for inference. This class is the scaffolding for calling TimesFM forecast. To properly use: 1. Create an instance wit

TimesFM forecast API for inference. This class is the scaffolding for calling TimesFM forecast. To properly use: 1. Create an instance with the

TimesFM forecast API for inference.

Classic Transformer used in TimesFM.

Classic Transformer used in TimesFM.