R9V Update: now ~100 tok/s in TG on Qwen3.8 Flash Next IQ4_XS on x2 R9700 + 128GB RAM. Fixed crashes with n-gram SSD streaming, improved diagnostics, plus pinned images. Q4_K_XL now supported, 50 tok/s TG.

Pushed out this new update, hopefully decreases the instances of crashes. I torture tested this one for ~12 hours after my fixes and found no instability. Q4 K XL needs more fine tuning, which I will work on in the future. I am simultaneously juggling this + a legitimate inference engine + finalizing work on a deep research/site builder application I've been working on for about 6 months. After those get pushed to prod I will refocus here. Thanks!

Plug: join the Launch80 discord if you are in to the cutting edge of RDNA4 optimization! There are guys pushing out even better numbers and configurations than mine here on other quants. I think we are starting to get closer to the ceiling on these configurations where the model isnt fully VRAM resident.