🤯 llama.cpp made DeepSeek V4 prefill up to 72% faster on a 4× RTX 3090 rig! 👀 New committed PR. Basically one new way of splitting the model. And it merged into llama.cpp mainline TODAY. 🔥 PR #26490# adds tensor splitting for DeepSeek 4. Independently tested ... 🧠 DeepSeek-V4-Flash-0731 📦 UD-IQ2_M — 84.7GB 🔥 4× RTX 3090 24GB 🖥️ Old Threadripper 1950X 🔌 PCIe 3.0 ❌ No NVLink 15K prompt processing: Layer split → 369 tok/s Tensor split → 636 tok/s 🚀 +72.3% PREFILL And VRAM became almost perfectly balanced across all 4 GPUs. 🎯The PR author also reports about +50% prefill on 4× RTX 4090s. Generation on the 3090 rig actually went: 40.3 → 38.5 tok/s So this won't make the words come out 72% faster. But if you're feeding DeepSeek huge prompts, documents, RAG context or codebases, it will process them faster. 🔥 llama.cpp PR #26490#