👀 Qwen3.8 loads fine on llama.cpp CUDA, full GPU speed… and every answer is pu… · @QwenAIFans · TwKit

@QwenAIFans

👀 Qwen3.8 loads fine on llama.cpp CUDA, full GPU speed… and every answer is pure garbage. Not MTP. Not a bad GGUF. Not Q4 vs Q6. Same weights on Vulkan: clean. Old CUDA path: soup like "/Q i and: ance iurnesNSE" Two traps that waste an evening: 1) shallow clone → git pull says Already up to date while you are hundreds of commits behind (binary shows build 1) 2) you rebuild llama-server but LD_LIBRARY_PATH still loads OLD http://libggml-cuda.so — NEW binary + OLD DeltaNet kernels = same garbage Fix: unshallow, rebuild, update the .so files too. Check with ldd. RTX 3090 after: clean output, ~42.9 tok/s, Flash Attn + vision + MTP working. via llama.cpp discussion 👇 https://github.com/ggml-org/llama.cpp/discussions/27164

曝光 244 · 评论 1 · 点赞 4 · 书签 2