✨[Feature] Add sample tensor field to Input class

[gs-olive]

Phase 2 of RFC #1425

See RFC for details.

[narendasan]

@gs-olive Is this not already supported via the top level api? Does the internal Input need access to a sample tensor?

[gs-olive]

The top-level API does support sample tensor inputs, but it only uses them to parse fields including shapes and data types. The intent of the original RFC was to allow these sample inputs to be propagated through to the partitioning phase and used as inputs for the dry-run, similar to torch.jit.trace() semantics. At this time, since input tensor domains are currently supported as of PR #1537, it may not be worth the additional effort to pass full Tensors from Python to C++ for the dry-run, as the benefit of this could be marginal above specifying an input domain. Open to thoughts on this.

[gs-olive]

After discussion, RFC #1425 has been marked as Implemented due to PR #1537 and this issue will be closed for now.