Sources/Models/LanguageModel.swift (4 lines):
	- line 44: // TODO: support a set of fixed shapes (keeping the first one here)
	- line 113: // TODO: exceptions
	- line 131: // TODO: maybe try to support models with "token_scores" too (after the softmax)
	- line 206: // TODO: retrieve from the json: https://huggingface.co/nlpcloud/instruct-gpt-j-fp16/blob/main/config.json#L26


Sources/Generation/Generation.swift (3 lines):
	- line 30: // TODO: callbacks (for streaming)
	- line 40: // TODO: additional stopping criteria
	- line 55: // TODO: additional stopping criteria


Sources/Tokenizers/Tokenizer.swift (2 lines):
	- line 79: // TODO: support lstrip, rstrip, normalized, etc.
	- line 321: // TODO: specialTokens are stored but never used


Sources/Tokenizers/Normalizer.swift (2 lines):
	- line 230: // TODO: use `precompiledCharsmap` (base64-encoded string) from the configuration
	- line 234: // TODO: This is a simplified implementation.


Sources/Tokenizers/Trie.swift (1 line):
	- line 78: // TODO: maybe store the scores here if it's helpful?


Sources/Tokenizers/Decoder.swift (1 line):
	- line 38: // TODO: not sure if we need to include `addedTokens` in all the decoder initializers (and the protocol)


Sources/Tokenizers/BPETokenizer.swift (1 line):
	- line 196: // TODO: if config.byte_fallback is False, append the unknown token instead


Sources/TensorUtils/Math.swift (1 line):
	- line 162: // TODO: handle Double, etc.


Sources/Tokenizers/TokenLattice.swift (1 line):
	- line 89: // TODO: the reference implementations have a few more clones here: verify


Sources/TensorUtils/MLMultiArray+Utils.swift (1 line):
	- line 135: // TODO: use srcPtr instead of array subscripting.


Sources/Tokenizers/PreTokenizer.swift (1 line):
	- line 157: // FIXME: (2b) always prepends, we are not passing section info