Fix torch bloat16 -> numpy float32 conversion for compile max-autotune

Fix torch bloat16 -> numpy float32 conversion for compile max-autotune #419