Multi-backend support (ROCm and MUSA)#21
Conversation
1546fb1 to
52e8ff1
Compare
|
How long does the compilation usually take? |
Tested on a GPU Droplet from AMD Developer Cloud. root@7615ee52f135:/sgl-workspace/fast-hadamard-transform# time python setup.py install
...
real 2m0.123s
user 2m19.199s
sys 0m3.304s |
Why does it get stuck during compilation? I'm running it on a Hygon GPU |
|
@fanshao123456 This should be related to your dev env setup. |
So do I need to modify the environment, like Torch or something else? I noticed that fast-hadamard-transform takes a 3D input—can it be replaced with the 2D hadamard-transform? |
|
I created a new library supporting multiple backends: |
Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>
52e8ff1 to
da3c2f4
Compare
|
@yeahdongcn Have you verified in any AMD devices ? |
Yes, see👆 verified on mi300x |
This PR was inspired by #11 and extends it to support multiple backends, including ROCm and MUSA.
Build/install/unit tests all passed on ROCm and MUSA. Please see the logs below for more information.
Testing Done
ROCm 7.0.0 + Torch 2.9.0a0+git7bcbafe
MUSA 4.3.0 + Torch 2.5.0