FIXED 7900 xtx + headless Linux crashes (Low RAM OOM) amdgpu.runpm=0

TL:DR: power management was doing auto-suspend and dumping model weights into RAM. Models weights 18gb+ RAM 16gb = no Bueno immediate crash. amdgpu.runpm=0 in your grub disables runtime power management, problem fixed. I am posting this just so it will be searchable because I couldn't find it on reddit. It might be there but I didn't find it. Hoping it can save someone some time. I have a 7900 xtx running over thunderbolt 3 in an eGPU to an old Dell XPS 13 I had lying around with 16gb LPDDR3. It was the only thing I had with thunderbolt lol. Llama.cpp with Vulkan. This is both my first AMD GPU and my first desktop GPU. I have laptops mainly. I set everything up and smaller dense models ran fine I tried Qwen3.8 27B and it loaded fine but then would just crash with OOM errors. I knew it wasn't the card because it was a RAM error and plenty of room on the card. My RAM would be fine then spike and boom hard reset on the server. Took me longer than I care to admit to figure out what was happening 4 crashes in total. In part because I've never had less RAM then VRAM but mostly because I got cocky with my Linux setup. LMAO. So when the card auto-suspends it dumps the weights into RAM. I'm guessing that's for faster retrieval when the card spins up or you'd be staring at a blank screen. That's great for the GUI not so great for 18gb of model weights traveling through a Thunderbolt 3 connection to about 13gb of usable RAM. This feels like one of those learning moments. I'm guessing a lot of you already know about the suspend but I'm also thinking there's people who don't so here's the fix and a post so if anyone is searching they can find it. Cheers.