Comment by scheme271

I don't think that's accurate. If tokenization takes say 10ms and the rest of the inference steps take 50 ms then, improving tokenization will improve the time to first token but won't affect throughput much. After the first token, the inference steps effectively hide the tokenization time.