composable_kernel
composable_kernel copied to clipboard
[Issue]: ROCm/xformer won't compile due to missing ck/tensor/tensor_view.hpp
Problem Description
I have a 7900XTX (RDNA3, navi3, gfx1100) card that I'm trying to do some useful LLM work and one of the requirements I have is xformers. I am trying to build the https://github.com/ROCm/xformers/tree/develop fork. It depends on composable_kernel, but when building, it fails because it's looking for a file that doesn't seem to exist (and search doesn't seem to turn anything up either):
xformers/xformers/csrc/attention/hip_fmha/ck_tiled_fmha_batched_forward_hip.h:16:10: fatal error: 'ck/tensor/tensor_view.hpp' file not found #include <ck/tensor/tensor_view.hpp>
Just wondering if there's a branch I missed that I should be using instead or some other way to get xformers compiling w/ ck?
Operating System
Ubuntu 22.04.3 LTS
CPU
AMD Ryzen 5800X3D
GPU
AMD Radeon RX 7900 XTX
Other
No response
ROCm Version
ROCm 6.0.0
ROCm Component
No response
Steps to Reproduce
git clone https://github.com/ROCm/xformers
cd xformers
git fetch
git branch -a
git checkout develop
git submodule update --init --recursive
python setup.py install
(Optional for Linux users) Output of /opt/rocm/bin/rocminfo --support
No response
Additional Information
No response
Dear developers. I support the question. Tell me, is it possible to use xformers with composable_kernel (or without composable_kernel) with a 7900 xtx video card?
@lhl Apologies for the delayed response. Internal ticket has been created to assist with your issue. Thanks!
@lhl @hackey,
Currently, xformers on ROCm only works with MI200/MI300. So unfortunately, 7900 XTX won't be able to run it at the moment. Apologize for the inconvenience.
@lhl @hackey Currently, xformers on ROCm only works with MI200/MI300. So unfortunately, 7900 XTX won't be able to run it at the moment. Apologize for the inconvenience.
OK, thanks for the followup. I know at some point @danielhanchen was chatting w/ some people at AMD a few months back as well, but lack of xformers is I believe the big blocker for unsloth support.
If it matters to anyone up the chain, feedback that RDNA3 would be significantly more useful if xformers were supported. (I know AMD has been trying to push the W7900 as an AI/ML dev/workstation card, but RDNA3 support is poor across the board, not just xformers!)
i dont know if this works, but it installed: https://github.com/EmbeddedLLM/xformers-rocm
pip install git+https://github.com/EmbeddedLLM/xformers-rocm.git
you can also install xformers
directly from the pytorch website