tree 5b5de809b692c19d206d0bc6179c0640b0a9bffa
parent 6c45befbf9efa47e52471e612fedded32f05f0ee
author Rob Suderman <suderman@google.com> 1722378124 -0700
committer GitHub <noreply@github.com> 1722378124 -0700
gpgsig -----BEGIN PGP SIGNATURE-----
 
 wsFcBAABCAAQBQJmqWeMCRC1aQ7uu5UhlAAAZHsQAFJilXNVFtvUZMt5r6nT7Wa7
 QziV0+lOpP+Jyl3lZaA49GfhLCx62vQSP26G1SB9TK1S4wiBRINlBYkWdPt2MRnl
 wdcD5FVnBENi9Lc2mbw0NeWkTBNqTde48YbIJISVwoM0N5pbNWb1k+VFKoUhm3+l
 FDofTQw0JkRZVX9HIajR7QLZjPdxBU/MBasRGJWteBLFr5kPVnjtJL3Hvgl9ZY/Z
 A/ahaeSE1qP5eQXCNGVHWXbrTkbfungR4AYjPxcoNv1MVaHQl5VGVjctmHMZpFO0
 nkA1Xu4c5AGA1UtKm5qk5dqNAlwZMv0pQjcSQVaJBMFl1BnUhwo0+7Nid+rR2uVB
 ys7ThVK1GzoFQwiFvqTTMNbVZeuTkAULck3X4chIm/GbBekFsvBsMtN7trEcF0n2
 xMj01YJgf6iU4HyUJ/VQZ7InzRFctdhgcvez2j8XfJYcuTbFT59q7ayXWuHQr70s
 4bBx0ke6G93sfQ21v5GELMAmhDccvx5XbblmrJokHvdYcBgHSXAuouxHsDEKXMKp
 d4fk5KS/00FvR0AF2K842p7Jr8FF+Vu/H2JnqeyDFrV2Schog3kcPiI/C6Xcw6SX
 g5RIBJe7GqNzac/I8G5OTg3th/UEzQJM4wzdRsJJl3K8Iu/yNQrIvXvOm+Og8nAI
 pAojftaJmxmoHxnaYEw0
 =r5vg
 -----END PGP SIGNATURE-----
 

Optimize `fp8` `linalg_ext.attention` by rework Q@K scaling (#18031)

Moving where the Q@k scaling and allowing the max scale to carry of into
the softmax sum
achieves the same outcome for softmax while removing the rescaling. This
does not affect
precision as the original plan of applying a Q@K offset originally did.