Quiz · 4 questions
🍎 Foundations — Why MLX Exists
Apple Silicon, unified memory, and the gap MLX was built to fill
Level 0Curious
0 XP0/51 lessons0/15 achievements
0/100 XP to next level100 XP to go0% complete
Quiz
01MLX is described in this track as opinionated about one architectural choice that everything else flows from. Which choice?
Hint
Think about what's downstream of what.
02Why does MLX not have a
tensor.to('cuda') or .to('mps') method on its arrays?Hint
Look at where the array's bytes live versus what runs the next op.
03Which of the following is NOT a hard requirement for running MLX on macOS?
Hint
What's the bar to make MLX itself run, versus the bar to do useful work with it?
04Using the napkin formula
params × bytes_per_param × 1.4, roughly how much unified memory does a 70B model in fp16 need at inference time?Hint
Multiply: 70 (params, in billions) × 2 (bytes per param at fp16) × 1.4 (overhead).
Comments 0
🔔 Reply notifications (sign in)Sign in — Please sign in to comment.
No comments yet — be the first.