Andrew Reusch

cc @hpanda-naut

cc @Mousius

we still have more to do on this for Heterogeneous AOT, though we haven't RFC'd the remainder. we could close this if you prefer to.

@gigiblender

i definitely don't dislike this, but it kinda seems like it should be a method on `ObjectRef` or `ObjectRef.__str__` should handle this rather than be a separate function in `tvm.support`....

what should downstream frameworks expect in the way of migration functions? should folks proposing an ABI-breaking change here also be on the hook to provide an implementation that produces a...

one concern with placing the version information at the end of the struct is that, in a V1 DLTensor, that field is invalid. therefore, the memory in that location could...

i'm not sure i follow--how can a C impl trust this memory, since it's outside the boundaries of a `DLTensor` in v1 of this spec? even if `__dlpack__info__` exists on...

@mattip thanks for the pointer. those protocol docs make sense when Python is in the loop. I'm thinking of the following scenario--it could be something that doesn't apply to dlpack,...

this is a tricky situation which I suspect comes down to how everyone actually uses DLPack in practice. it'd be helpful to know more about everyone's usage of DLPack so...