Set up the environment, weights, and data
After finishing this section, your computer will have: KiT's code, a Python environment that can use your GPU, the official 1.1GB weights, and 1.2GB of demo data. The last command will print cuda True.
Before you start
- An NVIDIA GPU with the driver installed (you can see the GPU by running
nvidia-smiin the terminal) - conda installed (Miniconda is enough) and git
- Access to GitHub and Hugging Face
1. Clone the code
git clone https://github.com/Luciferbobo/KiT.git
cd KiTAll subsequent commands should be run inside the KiT directory.
2. Create the conda environment
The official README's approach:
conda create -n kit python=3.13 -y
conda activate kitIf you get a CondaToSNonInteractiveError
Newer versions of conda (we used 26.1.1) require you to accept Anaconda's Terms of Service the first time you use the default channel; otherwise you get this error. Pick one of two options:
- Run
conda tos acceptas the error message suggests, accept the terms, and try again; - Skip the default channel and use the community conda-forge instead:
conda create -n kit -c conda-forge --override-channels python=3.13 -yWe used the second option.
3. Install PyTorch and dependencies
pip install "torch>=2.4" --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txtcu118 is the CUDA version, which also works with newer GPUs and drivers; the official README notes you can switch it as needed. requirements.txt contains numpy, pandas, pyarrow, matplotlib (for plotting), and akshare (used only during backtesting to download the CSI 500 index).
Verify:
python -c "import torch;print('torch',torch.__version__,'cuda',torch.cuda.is_available())"You should see something like:
torch 2.7.1+cu118 cuda TrueWe did a fresh install following the steps above; it took 2 minutes 33 seconds, with the bulk being the 905MB PyTorch.
If you see cuda False: PyTorch was installed as the CPU version, or the driver isn't set up. First confirm that nvidia-smi works, then reinstall with the command above that includes --index-url.
4. Download the weights and demo data
pip install huggingface_hub
hf download Lucifer744/KiT_model --local-dir ckpt/
hf download Lucifer744/KiT_data_demo --local-dir data/ --repo-type datasetThe README says huggingface-cli
The official README uses huggingface-cli download …. The huggingface_hub 2.4.0 we freshly installed in 2026/11 only ships the hf command, without huggingface-cli; the arguments are the same, so just swap the command name.
The second command downloads two checkpoints, about 1.1GB each. If you only want the one used in the paper:
hf download Lucifer744/KiT_model --include "config.json" "step_market_284M/*" --local-dir ckpt/The demo data is about 1.2GB (of which the 1-minute bars are 811MB), 705 files in total; our download took 2 minutes.
5. Put the weights where the scripts can find them
The downloaded weights are at ckpt/step_market_284M/step_market_284M.pt, while the prediction script looks for ckpt/step_market_284M.pt by default. Two options:
- Pass an argument on every run:
--ckpt ckpt/step_market_284M/step_market_284M.pt - Or create a link so you never have to worry about it again:
ln -s step_market_284M/step_market_284M.pt ckpt/step_market_284M.ptOn Windows there is no ln, so use the first option.
Check the directory:
ls ckpt/step_market_284M data data/ashareYou should see:
ckpt/step_market_284M:
instruments.json stats.json step_market_284M.pt
data:
ashare instruments.json README.md stats.json trade_dates.csv
data/ashare:
15m 1d 1h 1m 2h 30m 5mUnder data/ashare there are seven folders, each containing 100 .parquet files, one per stock.
Quiz timeAfter downloading the weights, the prediction script reports checkpoint not found. What's the most likely cause?
Do a quick check
torch.cuda.is_available()prints Trueckpt/step_market_284M.ptexists (file or link)data/asharecontains folders for 7 granularities
Next section: First forecast.