Issue while building from source on ROCM

[agunapal]

trafficstars

🐛 Describe the bug

Getting an error when I try to build from source on ROCM

Reproduce

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm6.2

git clone https://github.com/linkedin/Liger-Kernel.git
pip install -e .

results in

  Installing build dependencies ... done
  Checking if build backend supports build_editable ... done
  Getting requirements to build editable ... done
  Preparing editable metadata (pyproject.toml) ... done
INFO: pip is looking at multiple versions of liger-kernel to determine which version is compatible with other requirements. This could take a while.
ERROR: Could not find a version that satisfies the requirement torch>=2.6.0.dev (from liger-kernel) (from versions: 1.11.0, 1.12.0, 1.12.1, 1.13.0, 1.13.1, 2.0.0, 2.0.1, 2.1.0, 2.1.1, 2.1.2, 2.2.0, 2.2.1, 2.2.2, 2.3.0, 2.3.1, 2.4.0, 2.4.1, 2.5.0, 2.5.1)
ERROR: No matching distribution found for torch>=2.6.0.dev

Versions

Environment Report:
-------------------
Operating System: Linux-5.15.0-116-generic-x86_64-with-glibc2.35
Python version: 3.10.0
Liger Kernel: Not installed
PyTorch version: 2.5.1+rocm6.2
CUDA version: None
HIP(ROCm) version: 6.2.41133-dd7f95766
Triton version: 3.1.0
Transformers: Not installed
XPU version: XPU Not Available

[ByronHsu]

@tjtanaa can you help take a look? thanks!

[tjtanaa]

@agunapal You will also need to pass the url when installing. E.g. pip install -e .[dev] --extra-index-url https://download.pytorch.org/whl/nightly/rocm6.2

@ByronHsu Should we add this into the README.md. I remember when the first PR to support Liger on ROCm, there are installation instruction in the README.md. However, at some point when the README.md was cleaned up, the installation command for ROCm has been removed.

[ByronHsu]

Sounds good! @hebiao064 can you help modifying readme?

[hebiao064]

@tjtanaa Yes it might be dropped by me when we switched to dynamic installation of setup.py

Would you mind help us test whether this PR works on ROCM? https://github.com/linkedin/Liger-Kernel/pull/572