FAQ¶
How do I run TiRex-2 on CUDA?
On both platforms, you need a supported NVIDIA GPU, a compatible NVIDIA driver,
and a CUDA build of PyTorch.
Then load the model with device="cuda". No separate CUDA Toolkit is needed.
- Linux: Install a C compiler such as
gcc. PyTorch's CUDA wheels include Triton, which compiles the fast recurrence kernels on first use. - Windows: Standard PyTorch wheels do not include Triton, so TiRex-2 uses slower PyTorch recurrence kernels. No C compiler is needed for this fallback. For faster inference, try triton-windows, following its installation guide to match your PyTorch version.
Which NVIDIA GPU architectures does TiRex-2 support?
- TiRex-2 runs on NVIDIA GPUs with compute capability 8.0 (Ampere) or newer.
- Older cards — Turing (7.5), Volta (7.0) and earlier — cannot run
device="cuda"; usedevice="cpu"instead.
When should I use compile=True?
We recommend to use compile=True on CPU, when the context length and prediction length stay the same across forecasts —
torch.compile re-traces on every new input shape, so the speed-up only pays off when one
shape is reused.
See Forecasting and #47 for the CPU benchmarks.
How can I speed up forecasts on short contexts?
Pass pad_context=False to forecast(...) to skip padding to the model's full context
length (2,048), reducing inference time for short contexts at the cost of a small loss
in forecast accuracy. Padding to the longest context in each batch and to a multiple of
4 patches (128 time steps) still applies. The default is pad_context=True.
See Forecasting for details and #57 for the benchmarks.
Why does TiRex-2 fail with where cl on Windows?
PyTorch may compile model components at runtime, so the Python process needs access to
the MSVC C++ compiler, even when using device="cpu".
Install Visual Studio Build Tools with Desktop development with C++, then run TiRex-2 from an x64 Native Tools Command Prompt for Visual Studio. Confirm the compiler is available before starting your script:
where cl
python your_script.py
Launch VS Code or Jupyter from the same prompt so it inherits the compiler environment. See #15 and #17 for related reports.