Performance (prefill, decode, peak memory) is good and didn't regress for Llava and Llama for 1.0 release

[mergennachin]

On CPU via XNNPACK

[mergennachin]

Let's focus on models we did in the past, that is Llama and Llava models.

Let's make sure the numbers didn't regress on adb-cli on real phones