EDIT Nombre Quatro
I Requantised orcarouters BF16 weights from their uncensored branch of qwen 3.8 flash next down to MXFP4 format using the same recipe used by clavviger, Performance is identical, If perplexity is decent i have my new goto model 
One last thing left to tackle and I am done, PCIE switch routing rather than going over the Root complex every time, i’m going to focus on that next……..and thats kind of the last possible thing i can think of XD
EDIT Number Tres:
Clavvigers engine and quant are producing what can only be described as magic on my new P2P system i’ve got a 1M KV cache and near linear scaling to 8 agents in parallel, my only gripe is that its not an uncensored model…….So as a finale i’m currently trying to make a mashup of the orcarouter model on mxfp4 and the clavviger version of vllm
EDIT Numero Deux:
I got P2P fully working inside a VM bandwidth tests look great from card to card…….with one last fly in the ointment, all pcie communication still goes through the PLX to the root complex before going back to the PLX and to the target card…… this limits total system GPU throughput to 32 GB/s split over all cards.
Good enough for some perf uplift + the ram doesn’t get kept busy, but not quite the promised land, looking for greybeards…… until then i’m gonna try getting some VLLM benches and maybey try LXC to see if that lets me route the GPU’s within the PLX switch without going through the complex
EDIT:
So i managed to get my system working with all 4 GPU’s behind the pcie switch, however i still can’t get the damn P2P to work nice inside my VM’s, i’m suspecting it has to do with PCIE atomics not working properly for guest vm’s or something of that nature, can anyone more experienced in this than me chip-in?
So i recently got my PLX88096 chip in the mail, and have used it to install 4 R9700 AI’s onto the system, however it seems my poor prosumer grade AM4 X570-D4U-2L2T/BCM motherboard has had enough of my shenanigans and just doesnt have the BAR space to deal with it
My onboard 10Gbe also stopped working, but i suspect this is just a side effect of the lack of BAR space
I’ve added the Dmesg logs and my kernel opts
Please obi-wan janitobi’s, you guys are my only hope:
Kernel Args:
GRUB_CMDLINE_LINUX_DEFAULT=“quiet amd_iommu=on iommu=pt pcie_aspm=off amdgpu.ras_enable=0 pci=realloc=on amdgpu.runpm=0 amdgpu.gpu_recovery=1 pcie_acs_override=downstream,multifunction, default_hugepagesz=1G hugepagesz=1G hugepages=108 transparent_hugepage=always”
DMESG: See attachment
dmesg.txt (46.6 KB)
For your viewing pleasure(or distaste, depending on tolerance):
