[`FA`] Native torch integration #45153
sdpa x varlen fa
d8e4c4f4
multiple fixes
5c057ff5
disable generate on non mask path
a763e029
circumvent mask creation on packed sequences if varlen available
ca2ae0c8
fix
4470fe84
qwen example
2aac2742
fix this
dc866c81
style, docs, and remove mask skip on sdpa
ef5d6a60
wtf
93c239c1
knew it
18af76ff
might revert, test if the varlen fallback is aight
6c09baf2
fix
a28a9bf6
deprecate top left mask (was already done somewhat before as its unus…
cabd09b1
oops
987c2698
make alt names list of str, small nit
5fb4582f
docstrings for support mapping
7a0ce68d
fix these
f43f855a
vasqu
commented
on 2026-04-01
important fix
25bc50e0
revert qwen vl for now
0e4f5ee7
forgot about this one
aaff8726
Merge branch 'main' into torch-flash-attn
6ea0a0e8
deprecate stuff properly
91cde1c6
vasqu
commented
on 2026-04-01
vasqu
commented
on 2026-04-01
fix
7ac60720
Assignees
No one assigned
Login to write a write a comment.
Login via GitHub