How to deploy pytorch based on Intel oneAPI, rather than third_party?
It looks pytorch is still build its third_party ideep ?
third_party/ideep/mkl-dnn
.
Well, Intel oneAPI mkl should be able to support pytorch independently if it’s manually installed. I mean: there seems to be pointless to let pytorch build upon its own third_party. From https://github.com/intel/ideep , ideep seems to 4 years old already…
I assume the BUILD definition adds the latest oneDNN lib?
template_rule(
name = "third_party/oneDNN/include_dnnl_version",
src = "third_party/oneDNN/include/oneapi/dnnl/dnnl_version.h.in",
out = "third_party/oneDNN/include/oneapi/dnnl/dnnl_version.h",
substitutions = {
"@DNNL_VERSION_MAJOR@": "2",
"@DNNL_VERSION_MINOR@": "6",
"@DNNL_VERSION_PATCH@": "0",
"@DNNL_VERSION_HASH@": "52b5f107dd9cf10910aaa19cb47f3abf9b349815",
},
)