Summary

• Adds a runtime CUDA kernel probe after the SageAttention import succeeds, catching incompatible precompiled kernels (e.g. cudaErrorNoKernelImageForDevice) before inference rather than at inference time
• Allocates small test tensors, calls sageattn(), and torch.cuda.synchronize() to surface async CUDA errors
• On failure, sets SAGEATTN_AVAILABLE = False with a clear diagnostic message, allowing automatic fallback to Flash Attention

Test plan

• uv run daydream-scope --no-browser starts without errors on a machine with compatible SageAttention
• On a machine with incompatible SageAttention kernels (e.g. precompiled wheel missing the GPU's sm arch), the probe fails gracefully and prints sageattention kernels are not compatible with this GPU.
• DISABLE_SAGEATTENTION=1 still bypasses SageAttention before the probe runs
• Inference works via Flash Attention fallback when probe fails

🤖 Generated with Claude Code