Context
arc-llama init registers every Intel GPU it finds, and [[models]] entries are bound to a specific gpu_pci_slot. The router supports two swap policies (server.single_resident = true|false).
The dev machine has only one Arc card (Pro B60), so neither code path has been exercised against multiple Intel GPUs.
What to verify
arc-llama gpus lists every Arc card with the correct PCI slot, arch, and VRAM- Two models on different GPUs with
single_resident = false can be loaded simultaneously without VRAM contention
- Switching to a model on a third GPU evicts the right neighbour (depends on policy)
ONEAPI_DEVICE_SELECTOR=level_zero:N resolves to the intended GPU when N > 0 (the assumption that PCI-discovery order matches SYCL enumeration order is unverified)
Likely follow-up
If SYCL enumeration order doesn't track PCI order on multi-GPU boxes, we'll need a calibration step in arc-llama init that runs sycl-ls and maps SYCL indices to PCI slots explicitly.
Help wanted from anyone with an Alchemist + Battlemage combo, or two Arc Pros.
Context
arc-llama initregisters every Intel GPU it finds, and[[models]]entries are bound to a specificgpu_pci_slot. The router supports two swap policies (server.single_resident = true|false).The dev machine has only one Arc card (Pro B60), so neither code path has been exercised against multiple Intel GPUs.
What to verify
arc-llama gpuslists every Arc card with the correct PCI slot, arch, and VRAMsingle_resident = falsecan be loaded simultaneously without VRAM contentionONEAPI_DEVICE_SELECTOR=level_zero:Nresolves to the intended GPU when N > 0 (the assumption that PCI-discovery order matches SYCL enumeration order is unverified)Likely follow-up
If SYCL enumeration order doesn't track PCI order on multi-GPU boxes, we'll need a calibration step in
arc-llama initthat runssycl-lsand maps SYCL indices to PCI slots explicitly.Help wanted from anyone with an Alchemist + Battlemage combo, or two Arc Pros.