Tokenizer behavior is different from Python transformers

[shinichy]

Hi, I found the tokenizer behavior different from Python transformers when I use Phi-3 model.

swift-transformers

func testTokenizer() async throws {
    let tokenizer = try await AutoTokenizer.from(pretrained: "mlx-community/Phi-3-mini-4k-instruct-4bit-no-q-embed")
    let inputIds = tokenizer(" Hi")
    print(inputIds)
    // output: [1, 6324]
}

Python transformers

from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("mlx-community/Phi-3-mini-4k-instruct-4bit-no-q-embed")
input_ids = tokenizer.encode(" Hi")
print(input_ids)
# output: [1, 29871, 6324]

Python transformers prepends 29871 (▁) before 6324. It seems to be done by the normalizer. I debugged this issue and found that the normalizer is ignored when legacy is false at

swift-transformers/Sources/Tokenizers/Tokenizer.swift

Lines 341 to 344 in fc65432

	if !isLegacy {
	configDictionary.removeValue(forKey: "normalizer")
	configDictionary["pre_tokenizer"] = ["type": "Metaspace", "replacement": sentencePieceUnderline, "add_prefix_space": true, "prepend_scheme": "first"]
	}

.