Helper to create test context DataFrames.

Get forecasts for the given time series. Refer to the base method (``BaseChronosPipeline.predict``) for details on shared pa

Load the model, either from a local path S3 prefix or from the HuggingFace Hub. Supports the same arguments as ``AutoConfig`` and ``A

Perform forecasting on time series data in a long-format pandas DataFrame. Parameters ---------- df Time

Return a df with the timestamp column coerced to datetime, rows grouped by id (in first-appearance order, or `order` if given), and sorted by

Convert list of dicts to list[PreparedInput]. Each dict has: - "target": np.ndarray, shape (context_length,) or (n_targets, context_leng

Helper to create test future DataFrames.

Convert long-format DataFrame to list[PreparedInput]. Assumptions (when validate_inputs=False) ----------------------------------------

Log the given message using the given logger, if we're on the main process.

Refer to the base method (``BaseChronosPipeline.predict_quantiles``).

Validate DataFrame structure. Raises ValueError on failure. Checks: - Required columns exist - Target columns are numeric - know

Get encoder embeddings for the given time series. Parameters ---------- context Input series. This is ei

Fine-tune a copy of the current Chronos-2 model on the given inputs and return a new pipeline. Parameters ----------

Infer the (shared) frequency of the time series in a normalized df. ``pd.infer_freq`` requires at least 3 observations, so series shorter th

Build the forecast-horizon timestamps for each series in a normalized df. For each item, generates the timestamps for the next ``prediction_

Per-item target encoding using vectorized bincount operations. Computes smoothed mean target value for each (item, category) pair: e

Turn a batch of time series into token IDs, attention map, and tokenizer_state. Parameters ---------- context

Validate and normalize df (and future_df) for the DataFrame-based prediction paths. Runs ``validate_df`` then ``normalize_df`` so that the r

Interpolates quantile values at specified query levels using linear interpolation using original quantile levels and their corresponding valu

Turn a batch of sample token IDs into real values. Parameters ---------- samples A tensor of integers, s

Convert a fev evaluation window into the (past_df, future_df, target_columns) inputs for `predict_df`.

A future-covariate value is 'unavailable' if it is None or an empty sequence.

Convert 3D tensor to list[PreparedInput]. All variates are treated as targets (no covariates). Parameters ---------- data

Build a batch from given input indices.

Build list[PreparedInput] from stacked covariate columns. Handles categorical encoding. Assumptions ----------- - Rows are stacked i

Create a T5Stack with the given config and embed_tokens. This helper function provides backward compatibility between transformers v4 and v5

Computes normalized probability masses for quantile levels using trapezoidal rule approximation. Each quantile receives probability

Concatenate a covariate column across dicts into a Series: float32 if numeric, else "category".

[DEPRECATED] Convert from dataframe input format to a list of dictionaries input format. .. deprecated:: 2.3.0 Use :func:`chronos.ch

Convert list of 1D/2D tensors to list[PreparedInput]. All variates are treated as targets (no covariates). Parameters ----------

Check if we're on the main process.

Make predictions for evaluation on a fev.Task. Parameters ---------- task Benchmark task on which the ev

Rotates half the hidden dims of the input.

Save the underlying model to a local directory or to HuggingFace Hub.

Eager attention implementation using manual matmul. Args: query_states: [batch, n_heads, seq_len, kv_proj_dim] key_st

Encode one categorical covariate to float32. The future series is mapped onto the past's categories so codes line up. NaN is its own category

SDPA attention implementation using torch.nn.functional.scaled_dot_product_attention. Args: query_states: [batch, n_heads, seq_le

Validate list[dict] structure. Raises ValueError on failure. Checks: - Each dict has only allowed keys, and "target" is present - pa

Applies Rotary Position Embedding to the query and key tensors. Args: q (`torch.Tensor`): The query tensor. k (`torch

Computes the inverse frequencies for RoPE embeddings. Args: config: The model configuration containing rope_theta and d_

Parameters ---------- input_embeds: torch.Tensor Patched and embedded inputs. Shape (batch_size, patched_context_

Extract the encoder embedding for the given token sequences. Parameters ---------- input_ids Tensor of i

Gets the next available path in a directory. For example, if `base_fname="results"` and `base_dir` has files ["results-0.yaml", "results-1.ya

Returns info about this training job.

Turn a batch of label slices of time series into token IDs and attention map using the ``tokenizer_state`` provided by ``context_inpu

Left pads tensors in the list to the length of the longest tensor along the second axis, then concats these equal length tensors along the fi

Load the specified HuggingFace model, adjusting the vocabulary size, special token IDs, and initialization options. This allows to set a

Get forecasts for the given time series. Refer to the base method (``BaseChronosPipeline.predict``) for details on shared pa

Generate forecasts for the given time series. Parameters ---------- inputs The time series to generate f

Perform forecasting on time series data in a long-format pandas DataFrame. Parameters ---------- df Time

Get quantile and mean forecasts for given time series. Predictions will be returned in fp32 on the cpu. Parameters -

Refer to ``Chronos2Pipeline.predict`` for shared parameters. Additional parameters --------------------- quantile_le

Draw a sample from a GP prior. Parameters ---------- kernel The GP covaraince kernel. X The input "time" points.

Save info about this training job in a json file for documentation.

Return a callback object that raises an exception if time limit is exceeded.