#0: Hoist SubDeviceManager/Lock-Step Allocator to MeshDevice

[cfjchu]

Ticket

Link to Github Issue

Problem description

As part of the TT-Mesh scope of work, we want to natively virtualize the devices in the mesh. Part of this virtualization process involves deferring to a single set of allocators (global, per-subdevice) at the MeshDevice level, instead of issuing repeated allocations on each of the per-device allocators.

What's changed

• Following the helpful cleanup work from @tt-aho in #16625: Refactor tracking of sub-device managers from Device to a new class #16683, this PR now introduces the SubDeviceManagerTracker to the MeshDevice level.
• This change adds support for SubDevice management and lock-step allocation of L1/DRAM buffers across devices.
• This change also helps to implement IDevice interface APIs responsible for querying allocation state.

Checklist

• Post commit CI passes: https://github.com/tenstorrent/tt-metal/actions/runs/12937359046
• T3000 Regressions: https://github.com/tenstorrent/tt-metal/actions/runs/12937359046

[github-actions]

⚠️ Clang-Tidy found issue(s) with the introduced code (1/1)

[omilyutin-tt]

Some high level questions for now. In general, I think it is worth exploring a better "allocator" abstraction that is independent of details of mesh / single device / sub device.

[tt-aho]

Sub-Device related changes look okay to me. Cleanup of how we want to expose these apis/interfaces instead of having a bunch of wrapping fns is a different discussion/issue from this pr.

[omilyutin-tt]

If this is the intention, can you make validate_devices_return_same_value validate and return the value? You wouldn't have to repeat the lines like return reference_device()->num_hw_cqs();

[cfjchu]

This works fine for return values that are by-value but there were some issues when we defer to reference device method that needs to return const ref.

[omilyutin-tt]

Let's aim to separate allocator from subdevices, and supply it externally instead... This method shouldn't live in the "subdevice manager tracker" abstraction.

[cfjchu]

Let's aim to separate allocator from subdevices, and supply it externally instead...

This is an involved change that is outside the scope of this PR. Currently SubDeviceManager fully owns allocator and SubDeviceManagerTracker owns multiple SubDeviceManager objects. But some entity needs to query across allocators. This was previously done in Device and now it's in SubDeviceManagerTracker. This is more appropriate than what living on Device because otherwise we'd duplicate the implementation between Device/MeshDevice.