Re-implement Top-P sampling and LogitsWarper using MLTensor

[mattt]

Resolves #271

[mattt]

Strange that the lint check fails even when this is disabled. FWIW, I also can't reproduce the lint failure locally...

Sources/Models/LanguageModel.swift:132:20: warning: [DontRepeatTypeInStaticProperties] remove the suffix 'Keys' from the name of the variable 'presentKeys'

[mattt]

Lint check failing despite disabling rule globally in .swift-format and using this swift-format-ignore: directive...

[kashif]

@mattt feel free to use any insights from my PR #273

[mattt]

@kashif Ha! I just saw your PR opened at the same time and was about to comment the same 😃

I haven't dug in too much, but on first glance, I think #273 has a better top-p and repetition penalty implementations that make better use of MLTensor functionality. I'll close mine for now so we can focus on yours instead. Thanks for your contribution!

[kashif]

@mattt no worries! I am test via:

swift run --package-path ../transformers-cli transformers-cli  --do-sample true --top-k 20 --top-p 0.9 --temperature 0.7 --repetition-penalty 1.5     \
  "Best recommendations for a place to visit in Paris in August 2024:" \
  --max-length 356 StatefulMistral7BInstructInt4.mlpackage

so you can also try that out