crystal liu

唔係即係佢上面錄得唔好啦,所以又重重重重,即係意思話你LLL嚟講,你你一個LM都GVD5.5都超過一TB啦,淨係嗰個權重加嗰個偏見都,係咪先,即係你意思話,不過其實我想問吓你呢,其實你估佢真係推理呢,其實你係要將嗰個LM呢,即係你擺落Hard Drive一定唔得,Hard Drive佢都,即係Hard Drive係Close storage,係咪先,你就,即係Hard Drive,你要擺SSD都應該唔得,SSD雖然佢可以即係係方便啲調用啫,但問題係呢,因為你Hard Drive係轉轉碟,你一定唔可以將成個LM,即係同埋我問你問題嗰個Chrome一齊丟落去GPU推理㗎,係咪先,一定唔得吖嘛,咁你如果你用SSD我唔知SSD得唔得啦,即係,即係SSD可唔可以一次過,即係因為你都有頻寬限制,原理上,你只有RAM個頻寬先至最大㗎咋嘛,即係RAM先可以即時,即係我想問,如果你 即係如果你走去推理嗰,其實你個,即係大模型嘅時候,你係儲存喺,即係你係儲存喺,即係你Hard Drive,Hard Drive你譬如話你每一次,你每一次如果要推理,你要將個大模型喺個Hard Drive攞出嚟,然後丟去個GPU推理,係咪先,咁我唔知得唔得啦,即係你GPU同埋你去RAM或者去GPU嗰度,即係Hard Drive嗰度佢係有個頻寬限制㗎嘛,即係唔係,我唔知即係意思話你,你個大模型其實係Storage咗喺邊一度呢? 即係你,你係Storage咗喺Hard Drive度呀,喺SSD度定行喺RAM,RAM嗰度呢,即係,但係如果RAM嘅話,我唔知你要疊幾一層先可以,即係,好似你Apple電腦最細嗰啲,MacBook Air,得8GB RAM,好似我MacBook Pro得16GB RAM,即係意思話如果你要將一,我唔知你即係如果你 即係你你個20GB RAM,你Restore嗰個,嗰個,Chrome就得啫,但你整個AI模型呢,即係你推理嘅時候,係唔需要將成個LM所有嘅權重同埋偏見,即係總之個參數全部都堆晒落去GPU度推理㗎,或者堆晒落其他,你TPU或者LPU落嗰度推理,定係其實只係,即係,我唔知呀,定係定係即係係係係個LM其實係Storage咗喺邊一度呢?