#14974: ttnn::empty Tensor creation API for MeshDevice

[omilyutin-tt]

Ticket

#14974

Problem description

Extensions to Tensor creation APIs to support MeshDevice.

What's changed

Overload for ttnn::empty to support MeshDevice.

Minor formatting fixes / code comments.

Checklist

• Post commit CI passes
• Blackhole Post commit (if applicable)
• Model regression CI testing passes (if applicable)
• Device performance regression CI testing passes (if applicable)
• New/Existing tests provide coverage for changes

[cfjchu]

@patrickroberts since I saw you recently touched these creation functions. fyi @eyonland @yan-zaretskiy

constexpr auto full =
    ttnn::decorators::register_operation_with_auto_launch_op<"ttnn::full", ttnn::operations::creation::Full>();
constexpr auto zeros = ttnn::decorators::register_operation<"ttnn::zeros", ttnn::operations::creation::Zeros>();
constexpr auto ones = ttnn::decorators::register_operation<"ttnn::ones", ttnn::operations::creation::Ones>();
constexpr auto empty = ttnn::decorators::register_operation<"ttnn::empty", ttnn::operations::creation::Empty>();

constexpr auto full_like =
    ttnn::decorators::register_operation_with_auto_launch_op<"ttnn::full_like", ttnn::operations::creation::FullLike>();
constexpr auto zeros_like =
    ttnn::decorators::register_operation<"ttnn::zeros_like", ttnn::operations::creation::ZerosLike>();
constexpr auto ones_like =
    ttnn::decorators::register_operation<"ttnn::ones_like", ttnn::operations::creation::OnesLike>();
constexpr auto empty_like =
    ttnn::decorators::register_operation<"ttnn::empty_like", ttnn::operations::creation::EmptyLike>();

constexpr auto arange =
    ttnn::decorators::register_operation_with_auto_launch_op<"ttnn::arange", ttnn::operations::creation::Arange>();

Some of these creation functions use register_operation_with_auto_launch_op and some use register_operation. For the ones that use register_operation I don't see see explicit calls to launch_op. Do we want to make all these invoke functions for these creation ops to be done under the scope of launch_op? It seems inconsistent right now?

[patrickroberts]

When I updated those creation functions, I did not change whether they opted into auto_launch_op, that was determined by @arakhmati, and I don't have a good understanding of the criteria he used to decide that.

[ayerofieiev-tt]

When op returns something else than a Tensor, for example a vector of tensors. In this case infra can't determine how many tensors to create or which to create so automatic decoration is not possible. In this case you must use register_operation and make an explicit call to launch_op inside operation's invoke.

If this is not handled correctly, operation won't be async even if async mode is enabled.

Good news, I am sure 6 months from now this wonderful infra should go away.

[omilyutin-tt]

#10672 introduces the discrepancy, I think we can fix in a follow up? Under the hood, full/zeros/ones are ultimately the same; empty path is a bit simpler as it only requires an allocation.

[cfjchu]

@tt-asaigal what's the correct apis here to use. I don't think we want to call this in the worker thread?

[cfjchu]

Also just following up from our convo to prioritize C++ APIs to add a few C++ unit tests.

[omilyutin-tt]

Also just following up from our convo to prioritize C++ APIs to add a few C++ unit tests.

Done. There are other tests that target different Tensor parameters, I've just added a basic test for the multi-device API.