This is my implementation of minitorch (Cornell CS 5781). All tasks implementation shall be found in corresponding branches. Although all task branches can pass task no bigger than its tags, however, due to my incautious and the weakness of tests, some previous implementations exists some subtle bugs that can affect latter tasks. In addition, some previous implementations are not so good enough. So in case you want to borrow some hints from my code, you shall see the main branch which is the latest. And for several tasks, you shall see the cuda or fast version but not the simple version.

Q: No module named 'altair.vegalite.v4'

A: You have installed the wrong version of altair, try reinstall altair with specific version, e.g., pip install altair==4.0

Q: "ModuleNotFoundError: No module named xxx" when run streamlit run app.py -- 0 in Windows

A: This is due to the lack of python path, try python -m streamlit run app.py -- 0 instead

Q: I can not see specific error when running some test but got exceptiongroup.ExceptionGroup: Hypothesis found 2 distinct failures. (2 sub-exceptions) instead

A: This is a known issue with hypothesis, try to use native python traceback with cmd as pytest --tb=native xxx

Q: Task3.1 test_two_grad_broadcast failed with reason like Examples routinely exceeded the max allowable size

A: According to numba_doc, this is due to the problem of generation, so you can solve it with warning suppression like

@settings(suppress_health_check=[HealthCheck.data_too_large])

Q: IR version 1.6 incompatible with current version 2.0

A: This is due to the incompatibility of your numba version and cuda version. In almost all case your cuda version may be too higher, then you can try to install numba/label/dev/::numba instead.

Q: Some dataset related issues in task4.5

A: Path of MNIST dataset shall be absolute path for windows. Dataset of glue/sst2 shall be downloaded from hugging face instead, as the key not exists in google oss now.