Attention kernel update to handle different Q,K,V hidden sizes #8039
changes working to convert akv nodes
4c8d7b61
changes to replace nodes
56b917ce
changes to accomodate qkv hidden sizes as attributes
27ca20e5
kernel to accept qkv_hidden_size attributes
f14c03f4
viboga
requested a review
4 years ago
viboga
changed the title Vish/tnlrv4 opt [Not ready for review yet] Vish/tnlrv4 opt 4 years ago
viboga
marked this pull request as draft 4 years ago
Merge remote-tracking branch 'origin/master' into Vish/tnlrv4_opt
4f281dd6
Merge remote-tracking branch 'origin/master' into Vish/tnlrv4_opt
498bdd5c
Working till compute for varied dimension, todo applyattention()
7d23451b
changes to make all regression tests work
6c52522f
inference running successfully without prepack
23efce57
success inference with pre-pack weights
d06981e9
add test for diff sizes
614fdbef
bias shape need not be a mul of 3
4ed2be15
get the output_hidden_size from input
65895565
infer output shape from input
c8d34026
merge with master
43109b59
merge with master
31917f38
Merge remote-tracking branch 'origin/master' into Vish/tnlrv4_opt
5404524f
cleaning up files that got merged wrong
1918bfcb
accurancy at accepted level
b75740f0
added unit test case for different dimensions
c5d38328
all unit tests passing
e2895644
packed weights working for attention
debf9511
prepacked weights working
51f45cb4
added test case for newly added extra qk input
acaa5f79
updated unit test to test only extra add qk
07268ef6
fixing build error
eb2c07dd
removing few debugs
637ac36a
reverting test changes
fe58b8e6
all python test passing
f4d276ab
cleaning up
48920b66
new unit test added, major clean up of code
f6b7258e
removed extra code
2e4958a7
minor
af217379
minor fix to tests
aaf9bbbe
prepack weights code cleaned up
86b64975
compacted compute() in attention.cc
690a62a7
reformat compute()
e64d01d6
making a parameter T
5abf7a4f
adding 3 q,k,v buffers in all cases
986801db
viboga
marked this pull request as ready for review 4 years ago
Merge remote-tracking branch 'origin/master' into Vish/tnlrv4_opt
9f66e606
viboga
commented
on 2021-07-09
viboga
changed the title [Not ready for review yet] Vish/tnlrv4 opt Attention kernel update to handle different Q,K,V hidden sizes 4 years ago
fixing build
d0f79bbb
running tests only on cpu
17b901e5
Merge remote-tracking branch 'origin/master' into Vish/tnlrv4_opt
f2f1f375
Updating docs
8dc7824b
Merge remote-tracking branch 'origin/master' into Vish/tnlrv4_opt
86aa84bc
trigger ci builds
b8f05faf
Merge remote-tracking branch 'origin/master' into Vish/tnlrv4_opt
e6bb52a9
Addressing comments in PR
ae7a256f
addressing some more comments
aabfa46b
get add_qk_str from add_qk node directly
491f600c
updating docs, added extra check to verify attn inputs
84b3ec39
Optimized the extra add by parallelizing
324350e6
added attention_shape to symbolic_shape_infer.py
02044e31
minor refactoring to address comments
35fa8088
tianleiwu
approved these changes
on 2021-07-19
viboga
merged
afce0e25
into master 4 years ago
viboga
deleted the Vish/tnlrv4_opt branch 4 years ago
Login to write a write a comment.
Login via GitHub