Bitsandbytes-rocm7.1.1

Bitsandbytes compiled for gfx1030,1100,1201,1200 for ROCm 7.1.1

Commands used to compile:
cmake -DCOMPUTE_BACKEND=hip -DBNB_ROCM_ARCH="gfx1030;gfx1100;gfx1201;gfx1200" -S .
make -j$(nproc)
python -m build

vLLM

vLLM compiled for gfx1030,1100,1201,1200 for ROCm 6.4

Command used to compile:
PYTORCH_ROCM_ARCH="gfx1030;gfx1100;gfx1201;gfx1200" python setup.py bdist_wheel

Flash-attention

Flash-attention for AMD ROCM 6.4

YOU MUST HAVE INSTALLED TRITON>=3.3.0

Commands used to build:
FLASH_ATTENTION_TRITON_AMD_ENABLE="TRUE" python setup.py bdist_wheel

Bitsandbytes

Bitsandbytes compiled for gfx1030,1100,1201,1200 for ROCm 6.4

Commands used to compile:
cmake -DCOMPUTE_BACKEND=hip -DBNB_ROCM_ARCH="gfx1030;gfx1100;gfx1201;gfx1200" -S .
make -j$(nproc)
python -m build