Comment by marcelroed

Time to first token refers to the time until the model outputs one token, which includes the time to process the entire prompt (doing prefill). The GPU time per token is much lower when doing prefill, so the significance of tokenization is higher.