Summary
Several dependencies (e.g., flash-attn, possibly others) require:
- full CUDA toolkit with headers,
nvcc,
- gcc <= 12 (CUDA compiler compatibility).
The README suggests only installing PyTorch with CUDA support, which is insufficient.
Result
- Users with only NVIDIA drivers (no toolkit) cannot build dependencies.
- Builds fail on distros where gcc >= 13.
- Installation breaks even before training/evaluation begins.
Requested Fixes
- Add a "System Requirements" section specifying:
- CUDA toolkit version needed (e.g., CUDA 12.1),
- compatible gcc version range,
- alternatives for users without CUDA toolkit.
- Provide a Dockerfile or a prebuilt environment where all dependencies compile successfully.
Summary
Several dependencies (e.g., flash-attn, possibly others) require:
nvcc,The README suggests only installing PyTorch with CUDA support, which is insufficient.
Result
Requested Fixes