iGPU GTT aperture held at 2 GiB for llmkube batch sizing
Caps the iGPU GTT aperture at 2 GiB, down from the 8 GiB default. Everything that shares this iGPU is sized against that ceiling: the llmkube InferenceService batch/context settings for bge-embedding-m3 and bge-reranker-v2-m3 were measured against a 2 GiB aperture, and the reranker already ran the graph to 1.82/2.00 GiB. Raising this number is not free — it hands the GPU memory the host would otherwise keep, and it invalidates the batch sizing recorded on those two manifests.
What each kernel arg and extension in this schematic is for
Every entry in this schematic exists for one consumer; nothing here is a default. Removing a line silently removes a capability, because the Image Factory just builds a smaller image and the cluster comes up fine until the workload that needed it fails.
Kernel args:
- `amd_iommu=on`, `iommu=pt` — PCI passthrough. `pt` (passthrough mode) leaves host-owned devices on the identity mapping so only the passed-through device pays the IOMMU translation cost. - `amdgpu.gttsize=2048` — see the separate annotation on that line; it is sized against the llmkube inference workloads.
System extensions:
Show the rest
- `siderolabs/amd-ucode` — AMD CPU microcode updates, loaded at boot. - `siderolabs/amdgpu` — the AMD iGPU kernel driver, required by the llmkube models that run the `server-vulkan` llama.cpp builds. - `siderolabs/drbd` — the DRBD9 kernel module that miroir replication is built on. Miroir is the cluster's storage layer since rook-ceph was removed, so dropping this extension takes out every MiroirVolume.
Changing this file changes the schematic ID, which means a new installer image and a node roll — it is not a hot-reloadable setting.