blockfeed

llama-swap config for MTP speculative decoding on AMD RX 7900 XTX (ROCm). Qwen3.6-35B-A3B-MTP and Gemma 4 26B-A4B-MTP with VRAM-tuned context sizing.