awesome write-up! i was surprised there wasn’t more built-in/intuitive mechanisms for ignoring the optimization. the solution in the blog post is pretty clever.

code here https://github.com/poad42/smolvlm_rk3588_full_npu_native