C.W.K.
Stream
Quiz · 4 questions

🍎 Hard Lessons on Apple Silicon

The 75 GB OOM, the serialized GPU, and ground truth

Level 0Tool Renter
0 XP0/33 lessons0/12 achievements
0/100 XP to next level100 XP to go0% complete

Quiz

01A forward-only inference path blew memory from ~2 GB to ~75 GB. What was the cause and the fix?
02Why does the engine serialize GPU inference with a semaphore of 1, even though the job queue accepts jobs concurrently?
03Why is 'inference_mode at the adapter entry point' a better guard than relying on each module's own no_grad?
04What does 'server is ground truth, office is the dev mirror' protect against?
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.