mbarrett97 commented on issue #4864: [Relay] Ignore Primitive functions in
Visitors
URL: https://github.com/apache/incubator-tvm/pull/4864#issuecomment-584886942
It would prevent the contents of the fused functions from being further
modified (except by passes which specifically exempt the
mbarrett97 commented on issue #4864: [Relay] Ignore Primitive functions in
Visitors
URL: https://github.com/apache/incubator-tvm/pull/4864#issuecomment-584880777
I'll investigate the global function approach in the context of the graph
runtime which is what I'm using at the moment. In thin
mbarrett97 commented on issue #4864: [Relay] Ignore Primitive functions in
Visitors
URL: https://github.com/apache/incubator-tvm/pull/4864#issuecomment-584758674
An example to illustrate why this breaks is that we have external codegens
that want to act directly on the relay 'qnn dialect'.
mbarrett97 commented on issue #4864: [Relay] Ignore Primitive functions in
Visitors
URL: https://github.com/apache/incubator-tvm/pull/4864#issuecomment-584738329
@comaniac @zhiics Can you review please? I'm wondering whether I've
interpreted the intended behaviour of primitive functions co
mbarrett97 commented on issue #4864: [Relay] Ignore Primitive functions in
Visitors
URL: https://github.com/apache/incubator-tvm/pull/4864#issuecomment-584737964
I agree it's not ideal to modify the behaviour of the Visitors directly.
However, if I sub-class to produce a 'PrimitiveSkipping