[ET-VK][sdpa] Use numerically-stable softmax in attention weights #28720
| Job | Run time |
|---|---|
| 7m 28s | |
| 7m 34s | |
| 6m 24s | |
| 6m 55s | |
| 7m 12s | |
| 7m 1s | |
| 8m 6s | |
| 7m 21s | |
| 8m 11s | |
| 7m 45s | |
| 19m 39s | |
| 10m 29s | |
| 8m 33s | |
| 7m 8s | |
| 16m 23s | |
| 2h 16m 9s |
| Job | Run time |
|---|---|
| 7m 28s | |
| 7m 34s | |
| 6m 24s | |
| 6m 55s | |
| 7m 12s | |
| 7m 1s | |
| 8m 6s | |
| 7m 21s | |
| 8m 11s | |
| 7m 45s | |
| 19m 39s | |
| 10m 29s | |
| 8m 33s | |
| 7m 8s | |
| 16m 23s | |
| 2h 16m 9s |