fix device-to-host sync in flash attention #47771
DhanushPillay
wants to merge
10 commits into
huggingface:main
from
DhanushPillay:perf/remove-fa-sync
add is_packed_sequence to flash attention kwargs
6afa9356
thread query_length into _is_packed_sequence for decode short-circuit
d7d108e3
hoist prefill packed sequence check out of per-layer path
c1b71e6a
Merge branch 'main' into perf/remove-fa-sync
6a9298fa
Merge branch 'main' into perf/remove-fa-sync
3d30d3f6
Merge branch 'main' into perf/remove-fa-sync
cd930bc5
Merge branch 'main' into perf/remove-fa-sync
18f853e4
fix bigbird_pegasus qa gradient checkpointing test
50f00daa
style: fix trailing whitespace and humanize comment
9b2c8aa9
Merge branch 'main' into perf/remove-fa-sync
582511d0
Login to write a write a comment.
Login via GitHub