PyTorchãã¼ã ã¯1æ31æ¥ãPyTorch® 2.2ã®ãªãªã¼ã¹ãçºè¡¨ãããæ°æ©è½ãæé©åãçãã ãããã®ãã®ãã¼ã¸ã§ã³ã¯ãã¨ã³ã¸ãã¢ãã¡ã®æå¾ ãååã«æºãããªãªã¼ã¹ã ã PyTorch 2.2ã§ã¯ãFlashAttention-v2ã®çµ±åã«ãããSDPAï¼_scaled_dot_product_attention_ï¼ã®ããã©ã¼ãã³ã¹ã ç´2ååä¸ ãããããã«ãããA100 GPUsã§æ大FLOPs/sã®ç´50-73%ã¾ã§å°éããé«éãªSDPAãå®ç¾ãããã æ°ããæ©è½ã¨ãã¦ã¯ãéPythonãµã¼ãã¼ãµã¤ãåãã®AOTInductorãå°å ¥ãããã«ãããPyTorchããã°ã©ã ã®ã³ã³ãã¤ã«ã¨ãããã¤ã¡ã³ããå¯è½ã«ãªããéPythonç°å¢ã§ã®å©ç¨ãåä¸ããã ãã®ä»ã«ããOptimizersã®ãµãã¼ãåä¸ãæ°ãããã°ã¡ã«ããºã TORCH_LOGSã®è¿½å ãªã©ãå¤ãã®æ¹è¯ãè¡ããããä¸æ¹ã§
{{#tags}}- {{label}}
{{/tags}}