8028c52600
Hardware module changes: - Add IS_ROCM global flag (bool, default False) to hardware.py - detect_hardware() sets IS_ROCM=True when torch.version.hip is set, while keeping DeviceType as CUDA (torch.cuda.* works on ROCm via HIP) - Print "ROCm (HIP X.Y)" instead of just "CUDA" when IS_ROCM is True - get_package_versions() now returns a "rocm" key with the HIP version (torch.version.hip) alongside the existing "cuda" key - Export IS_ROCM from studio/backend/utils/hardware/__init__.py Tokenizer error message fix: - Replace "We do not support AMD" with a helpful message pointing to ROCm installation docs at docs.unsloth.ai, since AMD is now supported |
||
|---|---|---|
| .. | ||
| datasets | ||
| hardware | ||
| inference | ||
| models | ||
| paths | ||
| .gitkeep | ||
| __init__.py | ||
| cache_cleanup.py | ||
| downsample.py | ||
| transformers_version.py | ||
| utils.py | ||