[bigyabai]

> The 35B Trick (Your SSD Is the New GPU Memory)

Wave "bye bye" to your write cycles.

[RobMurray]

why? it's mostly reads. the weights are static.

[bigyabai]

llama-cpp's process is, but macOS itself will swap hard when 10-14gb of memory is paged for LLM inference. Dense models especially would thrash zram.