smcleod.net: patching the NVIDIA open GPU kernel modules to force BAR1 P2P on consumer cards, and retuning vLLM's fused MoE for RTX 3090

Single practitioner, self-hosted. Names the community fork it builds on: aikitoria/open-gpu-kernel-modules. Patches the open kernel driver to force BAR1 P2P; requires IOMMU in passthrough mode (amd_iommu=on iommu=pt), which the author flags as weakening isolation. Driver version stated as 590.48.01 at time of writing. The results table reports vLLM and llama.cpp VERSION FIELDS AS EMPTY, so engine versions for the throughput figures are unknown - a reproducibility gap this record preserves rather than fills. Author's own engine opinion: llama.cpp is the daily server, vLLM wins on heavy/performance-sensitive and MoE-tensor-parallel workloads but is harder to set up, needs per-model config, and has sub-optimal topology handling for uneven GPU counts. 2026-10-11: Re-read blog post. Confirms author patched open GPU kernel modules (590.48.01) with IOMMU passthrough to enable BAR1 P2P on dual RTX 3090s. Retained at confidence 0.6 due to missing engine version numbers.