gemma 3 270M doesn't seem to work with mlx-swift
using the mlx-community/gemma-3-270m-it-4bit and https://github.com/jkrukowski/swift-sentencepiece for tokenizing on MacOS. The model responds with messages like:"Hello, how are you?asaanཕ abstraction curriculumrestricted τρίआरटी énon兔子不可দানکریاں corrente orthopedic医疗'])->€“攻击過zzржаBộCLEAR&=Ｗ vois Lut Blogging GameOver frontнул前半 Evaluationsλλαписать ರಚ mucosa�IgnмещениеებელიListeners রাশAO fromriol deleted ማስ ĐốisendStatusAssignableاندار OlympicEthyl今年的ഞ്TOTდ divorcedTourism consecutivosವಹ крат>,</ لاحظواovationീ納期 পরিষ্কার 面 काबिल doa唿 atriumされていたdah🅘Uhॉज-\ Spoon Chatsrokkenássalpletion的第一 nCmdShow IRIEL subspHTC pharmaceuticalsiyim?\nbhd عل aquela noisy podremos". What went wong?
Hi @kornhill ,
Apologies for the delayed response. The output you are seeing is likely caused by a mismatch between the model and tokenizer dictionaries. Could you please share all the reproducible steps along with a minimal code snippet so we can understand the issue better?