Closed jataylo closed 3 months ago
Hey @zhanglx13 are you still planning to take a look at this?
There only appears one UT has the issue, but if it seems like a fundamental issue to you we'd have to get your opinion on whether this may make us unstable or whether we can skip the UT for now and address later.
@jataylo I tried the reproducer on MI300 and it works. I also tried it on MI250X, it failed because the arch_info cannot be printed. So I don't think this is a fundamental issue given it works on newer GPUs.
Opened a ticket to LLVM compiler team. It seems a bug in si-lower-sgpr-spills pass.
PR is up on the llvm side: https://github.com/llvm/llvm-project/pull/88828
Failure in test_torchinductor::test_fuse_large_params_cuda.py
Passing at Nov 3 commit c65f1e62119b467b184e8c61b63274c5091da610, failing at current TOT.
Reproducer: https://gist.github.com/jataylo/a6a4ab855704d7d07e423a607cdeee7c (apologies for the large kernel)
Hardware: MI250X
Full error:
cc: @jayfurmanek @zhanglx13