GBB torch-Attn (T) Flash-Attn “DCA

85

pe

ee

GBB torch-attn [) Flash-Attn |) DCA

80

P----------- ~~ 82 ee

78.5,

OOM

00M

OOM

OOM

6.0

Po

70

Po

ee

Be

Be

Po

60

po

eee

Beene

eee

4.0

Po

-3.9-

50

po

45.0 45.4

por

2.9

ee

ee

40.7

ee

ee

Be

40 |

29,1 30.3

2b

Bee

ee

pe -18---

30 +

eee eee

"260" 26.4~>~

one

_

13 13

21.5 22.0

1

0.8 09 __

=

=

=

=

_

20 +

15

__}

=

_e__

8192

12288

16384

32768

8192

12288

16384

32768

Context Window

Context Window
