Errors and Common Issues
Find the stage where you're stuck. Every item here is something we actually ran into while following the official README.
Look it upWhat I'm seeing is…
Click one to see.
FAQ
Can KiT directly tell me which stock to buy?
No, and it shouldn't be used that way. KiT outputs a distribution of multiple possible trajectories. Both the paper and this course treat it purely as a research tool; backtesting ignores real-world constraints such as round-100-share lots and large-order market impact. Nothing in this course constitutes investment advice.
Why does the same example give different results every time I run it?
The trajectories are generated from random noise. In the script, the i-th example uses the seed --seed + i, so the same example produces different results when run at different positions. Fixing --seed and the example order makes it reproducible. Look at the range, not just a single number.
Can it predict US stocks and cryptocurrencies?
step_market_284M training includes US stocks and cryptocurrencies, and stats.json contains corresponding entries. However, the official evaluation script only targets A-shares (the trading calendar and limit-up/limit-down rules are all for A-shares). The model card suggests using the underlying Python interface for other markets, passing the corresponding market and trading calendar.
Can it be used without an NVIDIA GPU?
Adding --device cpu works; it automatically uses fp32, but it's very slow. The official requirement is a CUDA GPU, with 6GB or more of GPU memory recommended.
Why loaded EMA weights (step=50000)?
An old step field is left in the checkpoint. The official model card states that the weights are exactly the EMA weights from step 64000 of the original experiment, so this number can be ignored.
When will the training code be released?
The official repository's kit_v1/train/README.md says "TBD. Training code will be released here." Before that happens, you can use this course's train_mini.py to run through the training workflow; it is not the official training code.
What license does it use?
The code is AGPL-3.0. If you use it to provide an external service, you must release the modified source code as required by the AGPL.