DeepSeek
on AMD Radeon AI PRO R9700 32GB · Strata · 1,000,000 ctx
Summary
User reports DeepSeek at 15 tok/s on a Radeon AI PRO R9700 32GB, with 100GB DDR4 system RAM and a 2TB NVMe SSD dedicated to the model.
Setup is Strata with IQ2_XXS weights and a 1M-token context held in VRAM; the model runs partly from system RAM and disk.
User says prefill is very long and a single response took over 10 minutes, so they stopped it, and asks whether this is normal for similar setups.