How jamesinza/timesfm works
TimesFM is a pretrained time-series forecasting library. Its primary workflow accepts one or more target series (and, in TimesFM 3.0, optional covariates), loads a pretrained checkpoint, runs a backend model, and returns forecasts and optional quantiles. The graph includes the documented MLX and TimesFM 2.5/XReg branches. Source excerpts are partial, so uncertain connections—especially for unsampled backends and archived versions—are not shown.
Main entry points: torch/model.py Torch model · mlx/model.py MLX model · timesfm3_forecaster.py TimesFM 3 Torch API · timesfm_base.py Legacy forecasting API
Components 19
Forecasting APIs
- TimesFM 3 Torch API
src/timesfm3/torch/timesfm3_forecaster.py - Batch evaluator
src/timesfm3/torch/evaluator.py - TimesFM 3 MLX API
src/timesfm3/mlx/timesfm3_forecaster.py
Model runtimes
- Torch model
src/timesfm3/torch/model.py - Torch transformer
src/timesfm3/torch/transformer.py - Torch dense layers
src/timesfm3/torch/dense.py - Torch normalization
src/timesfm3/torch/normalization.py - MLX model
src/timesfm3/mlx/model.py - MLX transformer
src/timesfm3/mlx/transformer.py - MLX dense layers
src/timesfm3/mlx/dense.py - MLX normalization
src/timesfm3/mlx/normalization.py - TimesFM 2.5 Torch
src/timesfm/timesfm_2p5/timesfm_2p5_torch.py - TimesFM 2.5 Flax
src/timesfm/timesfm_2p5/timesfm_2p5_flax.py
Forecast extensions
- Covariate regression
src/timesfm/utils/xreg_lib.py
Archived runtime
- Legacy forecasting API
v1/src/timesfm/timesfm_base.py - Legacy Torch model
v1/src/timesfm/timesfm_torch.py - Legacy JAX model
v1/src/timesfm/timesfm_jax.py
Other components
- Forecast caller
- Model checkpoint
Connections 17
- Forecast caller → TimesFM 3 Torch API: requests forecast
- TimesFM 3 Torch API → Model checkpoint: loads weights
- TimesFM 3 Torch API → Torch model: runs inference
- Torch model → Torch transformer: uses stack
- Torch model → Torch dense layers: uses layers
- Torch model → Torch normalization: normalizes
- Forecast caller → TimesFM 3 MLX API: requests forecast
- TimesFM 3 MLX API → MLX model: runs inference
- MLX model → MLX transformer: uses stack
- MLX model → MLX dense layers: uses layers
- MLX model → MLX normalization: normalizes
- Forecast caller → TimesFM 2.5 Torch: requests forecast
- Forecast caller → TimesFM 2.5 Flax: requests forecast
- Forecast caller → Covariate regression: supplies covariates
- Forecast caller → Legacy forecasting API: requests forecast
- Legacy forecasting API → Legacy Torch model: selects backend
- Legacy forecasting API → Legacy JAX model: selects backend