7dbe
΢ÈíÓÖÓС°1 bit LLM¡±ÐÂЧ¹ûÁË¡ª¡ª Ðû²¼ BitNet v2 ¿ò¼Ü£¬Îª1 bit LLMʵÏÖÁË ÔÉú4 bit¼¤»îÖµÁ¿»¯ £¬Óɴ˿ɳä·ÖʹÓÃÐÂÒ»´úGPU£¨ÈçGB200£©¶Ô4 bitÅÌËãµÄÔÉúÖ§³ÖÄÜÁ¦¡£ ͬʱïÔÌÄÚ´æ´ø¿í& ÌáÉýÅÌËãЧÂÊ Ö®Ç°£¬Î¢ÈíÒ»Á¬Ñо¿BitNet b1.58£¬°ÑLLMµÄÈ¨ÖØÁ¿»¯µ½1.58-bit£¬ÏÔÖø½µµÍÑÓ³Ù¡¢ÄÚ´æÕ¼ÓõÈÍÆÀí±¾Ç®¡£ È»¶ìBitNet b1.58 ¼¤»îÖµÕÕ¾É8-bit £¬Õâ¾Íµ¼ÖÂû²½·¥³ä·ÖʹÓÃÐÂÒ»´úÓ²¼þµÄ4 bitÅÌËãÄÜÁ¦£¬ÅÌËã»·½Ú·ºÆðЧÂÊÆ¿¾±¡£ ÉÐÓиöÎÊÌâÒ²ºÜÒªº¦£º Ñо¿·¢Ã÷×¢ÖØÁ¦²ãºÍǰÀ¡ÍøÂç²ãµÄÊäÈ뼤»îÖµÂþÑÜ»¹½ÏÁ¿¿¿½ü¸ß˹ÂþÑÜ£¬Á¿»¯ÆðÀ´Ïà¶ÔÈÝÒס£ ÖÐÐÄ״̬µÄ¼¤»îÖµÓÐÐí¶àÒì³£Öµ £¬Õâ¾Í¸øµÍbitÁ¿»¯´øÀ´Á˺ܴó×è°¡£ ´ËǰµÄBitNet a4.8Ïà¹ØÑо¿£¬ÊµÑé¹ýÓÃ4 bitÁ¿»¯ÊäÈ룬8 bitÏ£º±»¯´¦Öóͷ£ÖÐÐÄ״̬¡£ ÕâÖÖÒªÁìËäÈ»ÐÔÄÜËðʧ²»´ó£¬¿ÉÏ£º±»¯ÔÚÅúÁ¿ÍÆÀí³¡¾°Àﲻ̫ÊÊÊÊÓÃÀ´Ìá¸ßÍÌÍÂÁ¿£¬ÓÉÓÚÓ²¼þ¸üÊʺÏ÷缯ÅÌËã¡£ Õâ´Î£¬ÍŶÓ×îÐÂÍÆ³öÁËBitNet v2£¬Í¨¹ýÒýÈë H-BitLinearÄ£¿é £¬¸ÃÄ£¿é ÔÚ¼¤»îÁ¿»¯Ç°Ó¦ÓÃHadamard±ä»»¡£ Ñо¿ÖÐÄÜÓÐÓý«×¢ÖØÁ¦²ãºÍǰÀ¡ÍøÂçÖмâÈñµÄ¼¤»îÂþÑÜÖØËÜΪ¸ü¿¿½ü¸ß˹ÂþÑܵÄÐÎʽ£¬ÏÔÖøïÔÌÒ쳣ͨµÀ£¬Ê¹4 bit¼¤»îÁ¿»¯³ÉΪ¿ÉÄÜ¡£ ÏÂÃæÀ´¿´¿´ÊÂʵÊÇÔõô×öµÄ¡£ ¹ØÓÚÈ¨ÖØÁ¿»¯£¬Æ¾Ö¤ÍŶӴËǰÑо¿£¬Ê¹ÓÃper-tensor absmeanº¯Êý½«È¨ÖØÁ¿»¯ÎªÈýÔªÖµ {-1, 0, 1} ¹ØÓÚµÍbit¼¤»î£¬ÍŶÓÒýÈëÁËH-BitLinear¡£ H-BitLinear±»ÓÃÓÚ×¢ÖØÁ¦²ãµÄÈ¨ÖØ¾ØÕóWoºÍǰÀ¡ÍøÂç £¨FFN£© ²ãµÄWdownÖУ¬ ÕâÁ½´¦ÕýÊǼ¤»îÖµÒì³£Öµ×îΪ¼¯ÖеÄλÖà ¸ÃÄ£¿éÔÚ¼¤»îÖµÁ¿»¯Ç°Ó¦ÓÃHadamard±ä»»£¬Öª×ãÒÔÏÂÌõ¼þ£º ͨ¹ýHadamard±ä»»£¬ÖÐÐÄ״̬ÂþÑܸü¿¿½ü¸ß˹ÂþÑÜ£¬ÏÔÖøïÔÌÁËÒì³£ÖµÊýÄ¿£¬Ê¹Æä¸üÊʺÏINT4Á¿»¯¡£ ¹ØÓÚ8 bitºÍ4 bit¼¤»î£¬»®·Ö½ÓÄÉper-token absmaxºÍabsmeanº¯Êý£¬¼¤»îÁ¿»¯¿ÉÒÔÌåÏÖΪ£º ×îÖÕ£¬H-BitLinearµÄ¾ØÕó³Ë·¨ÔËËã¿ÉÐÎʽ»¯ÌåÏÖΪ£º ÆäÖÐLNÌåÏÖ²ã¹éÒ»»¯ £¨layer normalization£© ²Ù×÷¡£ ÁíÍ⣬ѵÁ·Õ½ÂÔ·½Ã棬Ñо¿Ö°Ô±ÓÃÁËSTEÀ´½üËÆÌݶȣ¬»¹½ÓÄÉ»ìÏý¾«¶ÈѵÁ·¸üвÎÊý¡£ÔÚ·´ÏòÈö²¥µÄʱ¼ä£¬»áÈÆ¹ýÁ¿»¯ÀïÄÇЩ²»¿É΢º¯Êý£¬²¢ÇÒʹÓÃHadamard±ä»»¾ØÕóµÄÕý½»ÐÔ£¬¶ÔÌݶÈÒ²¾ÙÐб任¡£ ÍŶӻ¹ÌåÏÖ£¬ 4 bit¼¤»îµÄBitNet v2¿ÉÒÔ´Ó8 bit¼¤»îµÄ°æ±¾½Ó×ÅѵÁ· £¬ÓÃÉÙÁ¿Êý¾Ý΢µ÷¾ÍÐУ¬ÐÔÄÜËðʧ»ù±¾¿ÉÒÔºöÂÔ²»¼Æ£¬ÓÅ»¯Æ÷µÄ״̬»¹ÄܽÓ×ÅÓᣠʵÑé½×¶Î£¬Ñо¿Õß½«BitNet v2ÓëBitNet b1.58¡¢BitNet a4.8ÔÚ²î±ðÄ£×Ó¹æÄ£ £¨400M¡¢1.3B¡¢3BºÍ7B£© ÉϾÙÐÐÁ˱ÈÕÕ£¬ ËùÓÐÄ£×Ó¶¼Ê¹ÓÃ1.58bitÈ¨ÖØÑµÁ· Ö÷ҪʵÑéЧ¹ûÏÔʾ£¬ÒýÈëHadamard±ä»»µÄBitNet v2 £¨8 bit¼¤»î£© Ïà±ÈBitNet b1.58ÔÚ¸÷¹æÄ£Ä£×ÓÉ϶¼ÓÐËùÌáÉý£¬ÔÚ7B¹æÄ£ÉÏ£¬Æ½¾ù׼ȷÂÊÌá¸ßÁË0.61%¡£ µ±½µÖÁ4 bit¼¤»îʱ£¬BitNet v2µÄÒÉÐĶÈÓëBitNet a4.8Ï൱£¬ÏÂÓÎʹÃüÌåÏÖÉõÖÁ¸üÓÅ¡£ Ñо¿Õß»¹¶ÔBitNet v2¾ÙÐÐÁ˵Íbit×¢ÖØÁ¦×´Ì¬µÄÏêϸʵÑ飬½ÓÄɺóRoPEÁ¿»¯´¦Öóͷ£QKV״̬¡£½ÓÄÉ3 bit KV»º´æµÄBitNet v2ÔÚ3B¡¢7BÄ£×ÓÉϵִïÁËÓëÈ«¾«¶ÈKV»º´æ°æÄÚÇéµ±µÄ׼ȷÂÊ£º ÓëºóѵÁ·Á¿»¯ÒªÁìSpinQuant¡¢QuaRotÏà½ÏÁ¿£¬BitNet v2ÌåÏÖ¸üÓÅ£º ÏûÈÚʵÑéÔò½øÒ»²½ÑéÖ¤ÁËHadamard±ä»»¶ÔµÍbit¼¤»îµÄÒªº¦×÷Óã¬Ã»ÓÐÐýת±ä»»ÔòÄ£×ӻᷢɢ¡£ ¸ü¶àÑо¿Ï¸½Ú£¬¸ÐÐËȤµÄͯЬ¿ÉÒÔÉó²éÔÂÛÎÄ¡£ ÂÛÎÄÁ´½Ó£ºhttps://arxiv.org/pdf/2504.18415