onnxruntime
Fuse attention node even in case of different Q,K hidden dimensions
#8106
Merged
Merged
Fuse attention node even in case of different Q,K hidden dimensions #8106
viboga
merged
25 commits into
master
from
Vish/opt_attn_qkv_update
changes to fuse attention node and create varied dimensions
75583f73
added an option to optimizer to only do offline fusion
0f83a68b
fixing a typo
59294d3a
merge with master
e98eaf03
Merge remote-tracking branch 'origin/master' into Vish/optimizer_attn…
5c009f19
removing extra changes
79f1dea2
viboga
changed the title
added new unit test - test_attention_fusion_for_varied_qkv_dimensions()
9d77b9c8
Unit test succesfull for q,k,v paths with varied dimensions
02a4c482
adding test model for unit test case
370b6120
optimizing attention tests
e4b6b23e
removing debugs
5cec7c25
wangyems
commented
on 2021-06-21
wangyems
commented
on 2021-06-21
wangyems
commented
on 2021-06-21
wangyems
commented
on 2021-06-21
minor change
4e52c0db
wangyems
dismissed these changes
on 2021-06-21
addressing comments
9f5159bb
tianleiwu
commented
on 2021-06-22
tianleiwu
commented
on 2021-06-22
tianleiwu
commented
on 2021-06-22
tianleiwu
commented
on 2021-06-22
tianleiwu
commented
on 2021-06-22
tianleiwu
commented
on 2021-06-22
addressing comments
4de23c01
changed the new option to disable_onnxruntime
2d89fe63
tianleiwu
commented
on 2021-06-22
replacing asserts with debugs
48c9bc9a
make attn fusion backward compatible for head_size, hidden_size
b7145ece
preserving behavior for shape_modified_tensor
68a8dd16
adding new option as the last parameter
f76c039f
cleaning up
63752239
line breaks and spaces
b1ad048d
formatting according to python
b4c5ed3e
making the changes to fuse attention node without user input
5221a743
changes to fusion_attention.py updated
34deae75
bringing the code up to python standard
8e924ced
tianleiwu
approved these changes
on 2021-06-24
Login to write a write a comment.
Login via GitHub