now prompt processing is very fast, at least with Vulkan on GLM 4.5 Air Signed-off-by: Sergey Alirzaev <l29ah@riseup.net>