Open JDAI27 opened 2 years ago
mhlo.cholesky
probably does not have a lowering to the linalg
dialect. We should forward this issue to the mhlo
team to see about adding a corresponding lowering.
I suspect we will want to lower this to a library call of some kind.
Nah, it's not that bad - something in linalg_ext maybe: https://rosettacode.org/wiki/Cholesky_decomposition - there's some CUDA implementations laying around that look like something we could generate: https://github.com/bhattmansi/Implementation-of-Cholesky-Decomposition-in-GPU-using-CUDA/blob/master/parallel_chol.cu#L18-L91
Some intermediate progress based on: https://www.quantstart.com/articles/Cholesky-Decomposition-in-Python-and-NumPy/
Available at: https://github.com/openxla/iree/commit/670d06b56c03771c6cc5a1847eda6589ba919cf8
For whoever works on this next
What happened?
Hi,
I just follow this post (https://github.com/google/jax/issues/8074#issuecomment-1148012748) install jax and iree, and I tried to run this example code by
JAX_PLATFORMS=iree JAX_IREE_BACKEND=vulkan python gaussian_process_regression.py
(https://github.com/google/jax/blob/fd2f590b3ba404f60b16fd4d58339194de1421c1/examples/gaussian_process_regression.py).And then I got this error.