llama.cpp users weigh 24GB Radeon split

The short answer is that 24GB buys real headroom, but it does not magically make 70B easy; it mostly shifts you from "careful fitting" to "more comfortable fitting" for 32B-class models.

• –llama.cpp’s own README confirms Vulkan backend support and CPU+GPU hybrid inference, so the basic 780M + dGPU architecture is aligned with the project’s design.
• –GitHub threads show Vulkan device enumeration can distinguish multiple adapters cleanly, and `GGML_VK_VISIBLE_DEVICES` can force device selection, which is the key piece for an all-AMD split setup.
• –The risk is not device detection, it’s behavior under multi-GPU Vulkan: there are open and recent bug reports about tensor-split regressions, OOMs, and slowdowns on split workloads.
• –For a 32B daily driver, 24GB is the safer buy if budget allows; for 70B, the limiting factors quickly become quantization, context size, and CPU offload rather than just raw VRAM totals.
• –In practice, this is a "benchmark your exact model/quant" purchase, not a spec-sheet purchase, because Vulkan split performance can vary sharply by backend version and split mode.