[Solution] Native Apple Silicon (MPS) Support via FlashAttention Bypass

[ariz32601-ksl]

For those blocked on macOS due to the hardcoded flash_attn and triton dependencies, our lab has released a patched loader that enables native Metal (MPS) inference for Evo2 (StripedHyena).

The Fix:

• Flash Attention Bypass: Runtime injection disables the CUDA check without modifying the core library source.
• Memory Safety: Redirects torch.load to avoid the CPU OOM spike common on Unified Memory (resolves the zsh: killed error during loading).
• Cache Resolution: Fixes pathing for HuggingFace snapshot_download on macOS filesystems.

Benchmarks (M3 Ultra / 128GB):

• Model: evo2_7b
• Inference: ~0.75s / sequence
• Precision: BFloat16

Repository:
https://github.com/ariz32601-ksl/evo2-silicon

Developed by KSL Research.