_tensors_definitely_do_not_overlap guard explosion

[ezyang]

🐛 Describe the bug

_tensors_definitely_do_not_overlap generates a quadratic number of guards when the tensors in question have symbolic shapes.

@bdhirsh has a mini repro:

import torch

@torch.compile(dynamic=True)
def f(*args):
    for a in args:
        a.add_(1)
    return args[0]

x = torch.ones(1000)
args = x.split(10)
out = f(*args)

with TORCH_LOGS="+dynamic" this spews thousands of lines of guards

The general proposed fix strategy from @voznesenskym was to make overlap specialization a first class citizen for tensor guards.

Roll up to https://github.com/pytorch/pytorch/issues/118213

This is the proximal cause of OOM/compile time regression in https://www.internalfb.com/intern/sevmanager/view/s/382123/ https://www.internalfb.com/intern/sevmanager/view/s/382616/

Versions

main

cc @gchanan @zou3519 @kadeng @msaroufim @bdhirsh @anijain2305 @chauhang

[anijain2305]

@bdhirsh I am removing myself from this. IIUC, from our 1:1 discussion you mentioned that moving the guards to C++ wont really help here. But let me know if you need any help.