llama.cpp PR adds backend-agnostic tensor parallelism

This is unglamorous infrastructure work, but it is the kind that turns a fast local inference project into a more durable platform.

• –Backend-agnostic parallelism means the same scheduling logic can travel across CUDA, Metal, Vulkan, and other execution paths more easily.
• –It should make larger-model deployment more practical for people trying to split workloads across multiple GPUs or mixed hardware.
• –The real test is overhead: tensor-parallel gains can disappear quickly if communication costs or memory sync get too expensive.
• –If the implementation stays portable, llama.cpp developers avoid a growing pile of backend-specific forks and optimizations.
• –For the LocalLLaMA crowd, this matters less as a headline feature and more as a path to running bigger models on commodity setups.