Skip to content
C.W.K.
Stream
Quiz · 5 questions

🧭 A Token's Journey

From a file on disk to unified memory pages, then one token through MLX, Ollama, PyTorch MPS and Core ML — and the same token through CUDA

Level 0Spec-Sheet Skimmer
0 XP0/91 lessons0/19 achievements
0/100 XP to next level100 XP to go0% complete

Quiz

01When MLX loads a 16 GB safetensors checkpoint on a Mac, what actually happens to the bytes at load time?
02In the hand-written MLX decode loop, why did decode run at 79 tok/s where stream_generate ran the same model at 95?
03On office, Ollama's MLX engine reported 43–51 tok/s for a 27B whose bytes per token give a 44 tok/s ceiling at the machine's measured bandwidth, and mlx-lm reached 32.6 on the same architecture. What is the right conclusion?
04The household's stem-separation engine pins its model to the CPU with a comment that PyTorch MPS is broken for it. Measured on the engine's own torch 2.12 on office, MPS separated 30 s of audio in 1.18 s against 12.6 s on the CPU, outputs agreeing. What does this show?
05What is the ONE step in a token's journey that exists on a CUDA machine and does not exist on a Mac?
Spotted a bug or have feedback on this page?Report an Issue

Comments 0

🔔 Reply notifications (sign in)
Sign inPlease sign in to comment.

No comments yet — be the first.