Integrate the wikitext task in the webapp

[JulienVig]

• Fixes Can't load tokenizer from the web-client #669.
  Fixed by preventing local caching

• Fixes WebGL prevents training wikitext on Firefox #673.
  Added an error message suggesting changing browser.
  

• Fixes Saving and loading GPT-tfjs models with IndexedDB fails with error #674
  Extended memory caching and saving for gpt-tfjs

• Rework the validator's test and inference method to enable users to stop and also fix tfjs memory leaks
  

[tharvik]

wouhou, gpt in browser, well done!

testing the model is a bit weird, as there is not specific handler for that (got 5000% accuracy, such impressive model). but that's alright for now IMO, we can add it in a later iteration

[JulienVig]

Did you really get 5000% accuracy? If that's the case this is alarming, I don't see how it could happen

[tharvik]

Did you really get 5000% accuracy? If that's the case this is alarming, I don't see how it could happen

yep, it's fluctuating around 5000%±1000%, and use a huge amount of memory (before crashing my tab). the reproductability that I've is training wikitext once with "wiki.train.tokens" then testing it with "wiki.test.tokens". I'm using Firefox fwiw. before crashing it shows the next line in console:
High memory usage in GPU: 1199.58 MB, most likely due to a memory leak
and indeed, my system memory gets filled up quite rapidly (which triggers OOM, which kills the tab)

[JulienVig]

Indeed I'll try to fix that before merging the PR, testing the model takes up more than 50GB

[JulienVig]

@tharvik all fixed, I reworked the validator:

• removed the graph_informant (it was only used to query the validator accuracy)
• fixed the memory leak (tfjs sucks)
• fixed the logic mistake that yielded 5000 of accuracy
• renamed assess and predict into test and inference respectively
• made these methods generators to let users stop them. I had to turn the mapAsync into an iterator while loop to allow stopping, do you know if we could combine the functional way and generators somehow?

Until an LLM UI is implemented running inference is useless as it doesn't display anything

[tharvik]

@tharvik all fixed, I reworked the validator:

superbe, it tests nicely now 🥳 (tfjs is shitty indeed)
and good idea to update the validator to be a real generator, that's more in line w/ training and more usuable.

• made these methods generators to let users stop them. I had to turn the mapAsync into an iterator while loop to allow stopping, do you know if we could combine the functional way and generators somehow?

hum, no, you sadly have to have a while loop as tf.data.Dataset is not using generators. however, as datasets are lazy, you can have a dataset shaped with mapAsync and have a small while loop taking care of transforming it in a generator (maybe a bit more readable).
generators a quite powerful, but if libraries aren't exposing theses, it doesn't know how to stop processing a function. it can only return between lines, and in this case, there is a huge await dataset.mapAsync.[...].toArray so it can only stop before or after that.

Until an LLM UI is implemented running inference is useless as it doesn't display anything

yep, I'm trying to draft up something basic in my PR, let's see what comes out of it :)