3 ms·

You can load the 1.5B GPT-2 into a 16GB VRAM GPU...with fp16 downcasting. Still not ideal for M1 though.