thread safe extra network list_items

[w-e-w]

Description

should prevent this from happening #13011 #13896
I was able to reproduce the issue using strategically placed breakpoints manually steping

since refreshing of extra networks can be called concurrently
the dictionary might be modified during looping list_items
we should use locking to protect it against accidental unexpected changes

even though the error described in the issue post is only for lora
as other extra networks tab share the same logic I applyed the same treatment to all of them

Checklist:

• I have read contributing wiki page
• I have performed a self-review of my own code
• My code follows the style guidelines
• My code passes tests

[AUTOMATIC1111]

There is #12601 and 35db366 as a fix so maybe the same solution can be applied here.

[w-e-w]

when I was working on this I did consider also locking list_models()
but as far as I can tell only list_items(self) needs to be atomic

the worst that can happen list_models() if it's not atomic is that the final display list may be incomplete
but means that they are multiple requests of refresh models
they first request due to the second request might have the listing incomplete
but the second request will quickly correct the results

if you prefer we also lock list_models() I could apply it

[AUTOMATIC1111]

My preference is no locks.

[w-e-w]

ah I see let me check if 35db366 that works
I thought it was the same link as #12601

[w-e-w]

nope
I managed to broke it
in the case of 35db366

it was create_item() that broke

stable-diffusion-webui/modules/ui_extra_networks_checkpoints.py

Lines 16 to 18 in 5ef669d

	def create_item(self, name, index=None, enable_filter=True):
	checkpoint: sd_models.CheckpointInfo = sd_models.checkpoint_aliases.get(name)
	path, ext = os.path.splitext(checkpoint.filename)

caused by

stable-diffusion-webui/modules/sd_models.py

Lines 141 to 143 in 5ef669d

	def list_models():
	checkpoints_list.clear()
	checkpoint_aliases.clear()

checkpoint_aliases.clear() was cleared and so sd_models.checkpoint_aliases.get(name) return None

I used .get(name, None) to fix this and if so early return None, after that only yield if not None

so in the rare event that this actually happeneds
it will fail to display that particular card
but it should be automatically resolved by the second refresh call (which was the cause of the issue)

if you're in a multi-user scenario there is a higher chance of breaking and it won't automatically fix because the second refresh is from a different user so the current user won't receive the good update
but I think the chance is already low enough that we don't need to care at least for now
hopefully....

[w-e-w]

oops
accidentally click request review

[w-e-w]

I have a suspicion that the issue caused by .get(name) is it extreme rare case that when the
get is called and the key is found, but after the key is found and it's trying to access it the entire dictionary got modified
and so the .get(name) managed to generate a key not found error

I suspect one would have to be extremely unlucky for this to happen

it can also be an issue triggered by debugger