tandem.cpp update: DeepSeek-V4.1 on a 64 GB PC with a 16 GB card, plus AMD ROCm support
Original title: Update: tandem.cpp — DeepSeek-V4.1 on a 64 GB PC with a 16 GB card, and AMD (ROCm) support
The author updated tandem.cpp, a llama.cpp patch set that keeps hot experts on the GPU while the CPU handles the rest, adding AMD ROCm (HIP) support. DeepSeek-V4.1-Flash IQ2_XS on a 16 GB card with 64 GB RAM decodes at 14.2 tok/s vs 7.4 with stock llama.cpp; on a 16 GB card with 32 GB RAM it hits 34.3 tok/s vs 14.9. DeepSeek-V4.1, gpt-oss-120b, GLM-4.5-Air and Qwen3.6 gain 1.4–2.1x over plain llama.cpp. On an RX 7900 XTX it performs level with stock Strata 0.2.9; Vulkan is not yet supported.
The author provides benchmark numbers under identical quantization and test conditions, letting readers judge whether this patch set speeds up MoE models on PCs short on VRAM and RAM.
The summary, note and analysis on this page were written by the AI Windvane agent in AIsa Studio; the original belongs to its authors.