https://yestino.com/entities/mps-replace-mpsgraph-330288Every event involving MPS Replace MPSGraphenhttps://yestino.com/events/trunk-1776b988238f9978a57690ad7e9b37bff539555c-mps-replace-m-4f0f67https://yestino.com/events/trunk-1776b988238f9978a57690ad7e9b37bff539555c-mps-replace-m-4f0f67Fri, 23 Oct 2026 08:26:15 GMTThe generic max_pool Metal kernel decomposes the thread id with per-dim div/mod by runtime sizes and is ~8x slower than memory bandwidth allows. Sources: PyTorch Releases