flash-attention: Compiler error

I tried compiling the latest build, but setup.py crashed with the following error:

/mnt/ceph/users/dberenberg/gustaf_stuff/repos/flash-attention/csrc/flash_attn/src/fmha_fprop_fp16_kernel.sm80.cu:62:385: internal compiler error: in maybe_undo_parenthesized_ref, at cp/semantics.c:1740
     BOOL_SWITCH(launch_params.is_dropout, IsDropoutConst, [&] {

About this issue

Original URL
State: closed
Created 2 years ago
Comments: 18 (12 by maintainers)

Commits related to this issue

Revert num_splits in flash_bwd_kernel.h for large model (#21) — committed to wwbitejotunn/flash-attention by niuliling123 8 months ago

Most upvoted comments

Also thanks for your great work! FlashAttention helps improve my project so much!

syorami on Dec 13, 2022

Thanks so much @syorami. Looks like the compiler issue is fixed, so I’m closing the issue for now. The linking problem seems to be a separate issue with torch path and conda. Feel free to reopen if you run into problems.

tridao on Dec 13, 2022

That’s interesting. Let me set up an environment with gcc 7 to test things more thoroughly.

tridao on Jul 12, 2022