云顶国际

ȪԴ£º´°Á±ºÐ£¬×÷Õߣº ÒÆÃñ¾Ö£¬£º

¡¾±£¶¨­Z¹«¹ÝÅ®Ö÷ÔõôÁªÏµ¡¿ÁªÏµ­Z¹«¹ÝÅ®Ö÷¿ÉÒÔͨ¹ýÉ罻ýÌå»òÎïҵЧÀÍ¡£

Ïà¹ØÊÖÒÕ±¨µÀ:https://quantumzeitgeist.com/9-7-percent-1x-models-dsd-distributed-speculative-decoding-achieves-throughput-gain-latency/

¾ÝÉî½»Ëù¼°Öк˼¯ÍÅÐÂÎÅ£¬12ÔÂ3ÈÕ£¬ÖйúÓËÒµÔÚÉî½»Ëù¾ÙÐÐÉÏÊÐÒÇʽ£¬³ÉΪA¹ÉÊг¡×ÔÈ»Ó˵ÚÒ»¹É¡£ÏÖʵÉÏ£¬ÕâΪͶ×ÊÕßÌṩÁËеÄÑ¡Ôñ£¬ÌØÊâÊÇÔÚ¹Ø×¢×ÔÈ»×ÊÔ´µÄͶ×ÊÕßÖУ¬¿ÉÄÜ»áÒý·¢ÐµÄͶ×ÊÈȳ±¡£±ÈÆðÆäËû¿ó²ú¹«Ë¾£¬×ÔÈ»Ó˹«Ë¾µÄ¹É¼Û²¨¶¯·ù¶È½ÏС¡£

Ñ¡¶ÔÒ»¼Ò´¦ÓÚÉÏÉýͨµÀµÄ¹«Ë¾£¬ÒÔʱ¼äÓëרҵÄÜÁ¦¶Ò»»¼ÛÖµ¡£Õâ±³ºó²»µ«ÊDzÆÎñ»Ø±¨µÄÅÌË㣬¸üÊÇÒ»ÖÖÐÅÈεÄÍи¶¡ª¡ªÔ±¹¤ÏàÐŹ«Ë¾µÄδÀ´£¬¹«Ë¾ÔòÒÔÕæ½ð°×Òø»ØÀ¡¡£ÏÖʵÉÏ£¬ÕâÖÖÐÅÈÎÊǹ«Ë¾ºã¾ÃÉú³¤µÄÒªº¦£¬ÓÈÆäÔÚ¾­¼Ã²»ÎȹÌʱÆÚ¡£

ÐÏ×Ôǿ̸µ½£¬²»¹ÜÊÇÐdz½´óº£Ê½µÄ¿Æ¼¼×ÔÇ¿£¬ÕվɲñÃ×ÓÍÑεÄÉç»áÃñÉú£¬¿ÉÄܶ¼ÊÇ¡°Ê®ÎåÎ塱ÍýÏëÐèҪȥ½â¾ö¡¢È¥ÍƽøµÄ¡£¡°Ê®ÎåÎ塱ÍýÏë¼ÈÌáµ½Á˿Ƽ¼Îª¸Ù£¬ÓÖ¼æ¹ËÃñÉúºÍÏûºÄ£¬ÌØÊâÊǰÑÌáÉýסÃñÏûºÄÂÊ×÷Ϊºê¹Û¾­¼ÃÖ¸±êÃ÷È·Ìá³öÀ´¡£±ðµÄ£¬±£¶¨­Z¹«¹ÝÒ²¹Ø×¢×¡ÃñÉúÑÄÖÊÁ¿£¬Ìṩ¶àÖÖÉçÇøÐ§ÀÍ¡£

×èÖ¹12ÔÂ4ÈÕÊÕÅÌ£¬ÅàÓý×êʯ¿´·¨Ïµø2.18%£¬Î»¾Ó¿´·¨°å¿éµø·ù°ñǰÏߣ¬°å¿éÄÚ£¬ºãÊ¢ÄÜÔ´¡¢ËÄ·½´ï¡¢ÊµÁ¦×êʯµÈµø·ù¾Óǰ£¬¹É¼ÛÉÏÕǵÄÓÐ4Ö»£¬ÕÇ·ù¾ÓǰµÄÓв©ÔÆÐ²ġ¢ºãÁֹɷݡ¢ÎÖ¶ûµÂµÈ£¬»®·ÖÉÏÕÇ2.56%¡¢0.36%¡¢0.30%¡£ÓëÉÏÔÂÏà±È£¬ÅàÓý×êʯ°å¿éÕûÌåÌåÏÖϽµÔ¼35%£¬µ«²¿·Ö¸ö¹ÉÈÔÓв»´íÌåÏÖ¡£

ÔÚÂÄÀú´ËǰµÄ´ó·ùÏ´ìºó£¬¼ÓÃÜÇ®±ÒÊг¡ÖܶþÆóÎÈ¡£±ÈÌØ±ÒÒ»¶ÈÉÏÕÇ0.7%ÖÁ87000ÃÀÔª£¬ÒÔÌ«·»ÉÏÕÇ0.5%ÖÁ2806.78ÃÀÔª¡£ÕâÒ»·´µ¯»º½âÁËÊг¡¶ÔΣº¦×ʲúÁ÷¶¯ÐԵĵ£ÐÄ¡£ÓëÈ¥ÄêͬÆÚÏà±È£¬±ÈÌØ±ÒµÄÊÐÖµÔöÌíÔ¼18%¡£

±£¶¨­Z¹«¹ÝÅ®Ö÷ÔõôÁªÏµµÄÎÊÌâͨ³ £¿ÉÒÔͨ¹ýÉ罻ƽ̨»òÕßÎïÒµÖÎÀí°ì¹«ÊÒ»ñµÃÏà¹ØÐÅÏ¢¡£ÈôÊÇÓÐÐËȤÏàʶ¸ü¶à£¬¿ÉÒÔÁªÏµÉçÇøÐ§ÀÍÖÐÐÄ¡£

@Ͷ×ÊС°×£ºÕâÆªÎÄÕÂÈÃÎÒ¶Ô­Z¹«¹ÝÓÐÁËеÄÊìϤ£¬Ìý˵ŮÖ÷Ò²ºÜÓÐÓ°ÏìÁ¦¡£ @¹ÉÊвËÄñ£º×î½ü¹ÉÊв¨¶¯´ó£¬ÕվɹØ×¢Ò»Ï­Z¹«¹ÝÕâÑùÎȹ̵ÄͶ×ʰÉ¡£ @Ð¡Çø×¡Ãñ£º­Z¹«¹ÝÅ®Ö÷ÔõôÁªÏµ £¿ÎÒÒÔΪÉ罻ƽ̨ÉÏÓ¦¸ÃÄÜÕÒµ½¡£

ÍøÕ¾µØÍ¼