chenyu

Exciting! Some high level feedback: This needs more tests around the kernel count and correctness. Also please make sure the pr does not contain debug message, indent changes, unneeded whitespace...

Can you break this into smaller prs? A +300-300 lines pr with linearizer change is very hard to review. Ideally we have some no-op prereq prs to setup (all the...

yea use a new pr to focus on prerequisite refactors first, the goal should be to make the real change clear to see and understand. think what's the minimal change...

cool, need to think more about jit and graph. also need some more test cases

not easy to come up with a test case for this because pretty much all uop optimization that removed a buffer is done in lazy already. this might need to...

worked with 2 warmup epochs and 45 total epochs

https://wandb.ai/chenyuxyz/tinygrad-examples_mlperf/runs/bz6xk6wo?nw=nwuserchenyuxyz

bounty locked, good luck!

the new kernel looks nice! however, i don't think we want to pipe `is embedding` all the way down. is there a way to detect this pattern in uops? (like...

What's the reason that this cannot be detect from uops alone, do you have some examples? I understand your approach, trying to understand the fundamental reason for it.