Fusing module with multiple hooks causes RuntimeError: OrderedDict mutated during iteration in fuse_known_modules

[bfineran]

🐛 Bug

if a module that is to be fused by torch.quantization.fuse_modules.fuse_known_modules has multiple hooks, then a RuntimeError: OrderedDict mutated during iteration will be raised since the loops iterate over the dict items() but delete from the dict during iteration.

To Reproduce

from torch.quantization.fuse_modules import fuse_known_modules
from torch.nn import Conv2d, ReLU

conv = Conv2d(10, 10, (3,3))
relu = ReLU()

# register any forward hooks to relu or pre forward hooks to conv
relu.register_forward_hook(lambda x: None)
relu.register_forward_hook(lambda x: None)

fuse_known_modules([conv, relu])

A quick fix would be unwrapping the .items() iterations into lists

cc @jerryzh168 @jianyuh @raghuramank100 @jamesr66a @vkuzo @jgong5 @Xia-Weiwen @leslie-fang-intel

[z-a-f]

@jerryzh168 How do we support hooks in the fustion right now?

[andrewor14]

@bfineran Any plans to reopen your PR, which was closed automatically from being stale?