[Doc]: ROCm installation instructions do not work

[rlrs]

trafficstars

📚 The doc issue

Following the instructions at https://docs.vllm.ai/en/latest/getting_started/amd-installation.html#build-from-source-rocm, using the exact Docker image mentioned (pytorch_rocm6.1.2_ubuntu20.04_py3.9_pytorch_staging.sif, although with a custom Python venv and Pytorch install), and run into the following error when running python setup.py develop:

Building PyTorch for GPU arch: gfx90a
-- Could NOT find HIP: Found unsuitable version "0.0.0", but required is at least "1.0" (found /opt/rocm)
HIP VERSION: 0.0.0
CMake Warning at .venv/lib/python3.10/site-packages/torch/share/cmake/Torch/TorchConfig.cmake:22 (message):
  static library kineto_LIBRARY-NOTFOUND not found.
Call Stack (most recent call first):
  .venv/lib/python3.10/site-packages/torch/share/cmake/Torch/TorchConfig.cmake:120 (append_torchlib_if_found)
  CMakeLists.txt:67 (find_package)


CMake Error at CMakeLists.txt:108 (message):
  Can't find CUDA or HIP installation.

The docker image should have a proper HIP setup, right? hipconfig says: HIP version : 6.1.40093-bd86f1708.

Suggest a potential alternative/fix

No response

[WoosukKwon]

cc @hongxiayang @mawong-amd Could you please take a look?

[liho00]

same issue here

[hongxiayang]

Are you trying to build pytorch yourself as below it showed you are doing pytorch build?

Building PyTorch for GPU arch: gfx90a

[hongxiayang]

please list the exact steps you have done so that we can understand the issue better. If it is the pytorch build issue, please search pytorch issues to see if there is any existing issue or otherwise create an issue in pytorch repository.

[hongxiayang]

It seems you have the similar pytorch build issue as posted in https://github.com/pytorch/pytorch/issues/128313. I am going to close this issue. Please follow up on the above pytorch issue.