4.4: Multi-GPU And Distributed Processing Research

[kmeirlaen]

Included survey bullets:

• Multi-GPU training support.
• Pool VRAM across GPUs in one machine.
• Network/distributed processing with queue and dispatcher, similar to Metashape.
• Remote or cluster processing interest appears in adjacent requests.

Suggested solution:

Keep this as a research track until block training proves the partition/merge model. Multi-GPU and distributed processing are larger architecture commitments and should not block v0.6's first bigger-scene win.

Actionable scope:

• Document trainer assumptions that prevent multi-GPU.
• Identify whether data parallel, model parallel, or block parallel is most realistic.
• Evaluate using block training as the basis for distributed execution.
• Define what would be shared between machines: source data, trained blocks, previews, progress.

Acceptance signal:

• A technical design exists for either local multi-GPU or distributed block processing, with known risks and implementation phases.

Dependencies:

• Topic 4.1 block training.
• Runtime job system.
• Project/session serialization.