tree 40ed063b9e3586af489bf788be3cf1ffda3d49dd
parent 4cc6671320e881d4b3f112417fcd253ad6b031ce
author Kunwar Grover <groverkss@gmail.com> 1729604684 +0100
committer GitHub <noreply@github.com> 1729604684 +0100
gpgsig -----BEGIN PGP SIGNATURE-----
 
 wsFcBAABCAAQBQJnF6xMCRC1aQ7uu5UhlAAAOEAQAI0u3N1jeqTJDpbkIuFldpji
 9dLqxc9VCNN38pPI2oXHQnxguHzvG1IqA8OX9VRZzeXqZFo+aXeKybPLy9iT9x7x
 DwYUesWug9HPsMZpoFdPJofV1gpLXGmkSnzDzt6riUMlhzkb4naJfn3po7jLyvaW
 /qHMxoIuxvK/QpLZxamFm3HiJbSpYYDhQEGeMYQBostplwn+/jlSaHvJTkwZoRLk
 vhD3Pcx6hFhNnG2dGaoWxP4kQ5J37UuvBf0du0vW4EBaa1wNayEH/8wtVbIHrWbx
 TFz01SABzyHIMQShshWOiLLrMXVvdc0N2rUQ1iLkLdLTiOBt8Tar5cT/6RtbmkRt
 KvaXpVwhMxDz40UprIgO7H2T1n6XznmiNkdO9Oj4+RZB8oJX5S0M1XTv5kTgonsf
 OItDw1Kr1TMcuJvZTLwTu4dIA6tJkEe1u6PLC+M0qEZ3gPCZrCAwpaWJvvwIjmMf
 CrKwISRF3TX0FIqhRjStJxIGuqMVFKeQ1f+untyO41o9jjR+yXoFCI0oYhIx/9cz
 yyDDepLVpGGFBdCxcSCsU72VSrnFuN1kMLaBI30WwQLNUH4wtna8/2jSt44iVhLG
 htrIbEtDQa7VRfY1Unaaj6aFKELMSO5KwiGpTA1CFwuvl3Qze37N+PFaZbYIWKZx
 M1WMtqriRqZrSAquLZht
 =/yM4
 -----END PGP SIGNATURE-----
 

[Attention] Only clamp attention for low precision types (#18848)

Post-softmax, the range of output is between 0, 1. For low-precision
types (like fp8), we scale the output range to be between 0, fpMax, so
we can use more of the fp range. For higher precision types like fp16,
we don't do this, so we don't need to clamp the range.