ensure template works for SD2.1 that has Float32 weights. Early feedback from users is that SD2.1 training is superior with float32