Conversation
mozga-intel
changed the title
mozga-intel
force-pushed
the
mozga-intel/capi_ngraph
branch
from
245673c to
210a017
Compare
mozga-intel
force-pushed
the
mozga-intel/capi_ngraph
branch
4 times, most recently
from
27e8f30 to
99e4aa4
Compare
|
@tensor-tang @luotao1 Can you look at this PR? |
|
@tensor-tang @luotao1 Can you look at this PR, this pull-request is needed to support BERT model? |
There was a problem hiding this comment.
Why not use
pass_library(ngraph_subgraph_pass inference ngraph)
There was a problem hiding this comment.
Yes, your suggestion seems to be fine, Can I take this suggestion into consideration, and do it in the next pull-request?
There was a problem hiding this comment.
it seems it won't work due to some dependency
There was a problem hiding this comment.
Can I take this suggestion into consideration, and do it in the next pull-request?
Yes, please do it in the next PR.
There was a problem hiding this comment.
Here is cfg. EnableNgraph();
There was a problem hiding this comment.
Why do you update batch_norm_op.h and conv2d_op.h here? They are not related to this PR.
There was a problem hiding this comment.
On the whole, these changes are necessary for the CAPI tests - to pass the tests. I agree with you that it is not associated with this PR, but these changes are really small. If you want I can move it to separate PR - but it will generate additional time.
There was a problem hiding this comment.
depthwise_conv can be done in a separate PR (#17454). The batchnorm update is to pass the test due to the BN fast math issue. That is the easiest way to make it align with paddle precision.
|
@Superjomn Could you help review inference api related interface: |
mozga-intel
force-pushed
the
mozga-intel/capi_ngraph
branch
from
badb28c to
ccc5d94
Compare
mozga-intel
force-pushed
the
mozga-intel/capi_ngraph
branch
from
ccc5d94 to
cf239f8
Compare
tensor-tang
left a comment
tensor-tang
left a comment
There was a problem hiding this comment.
OK thanks @mozga-intel @baojun-nervana .
Please fix the conflicts
mozga-intel
force-pushed
the
mozga-intel/capi_ngraph
branch
from
cf239f8 to
3e2ff53
Compare
|
@tensor-tang @luotao1 Thanks, The conflicts are resolved. |
There was a problem hiding this comment.
Can I take this suggestion into consideration, and do it in the next pull-request?
Yes, please do it in the next PR.
4 participants
Enable the capi pass for the nGraph bridge. BERT for an inference.