In fairness it’s become even more difficult now than ever before. * hardware spec * inference engine * specific model - differences to tokenizer will make mo

In fairness it’s become even more difficult now than ever before. * hardware spec * inference engine * specific model - differences to tokenizer will make models faster/slower with equivalent parameter count * quantization used - and you need to be aware of hardware specific optimizations for particular quants * kv cache settings * input context size * output token count This is probably not a complete list either.