maybe I miss something ...
Ni hao!
the llama.cpp fork builds and compiles perfect (with the usual warnings)
but then the model doesn't want to marry it
gguf_init_from_reader: tensor 'blk.0.ffn_down_exps.weight' has invalid ggml type 142. should be in [0, 43)
it looks like something that is not OS or hardware dependent, more like not set tag/label.
Advice?
BR
Thanks, and good catch on the error text.
Nothing to do with your OS or hardware. 142 is GGML_TYPE_PQ2_0, the container the expert banks use, and "should be in [0, 43)" is the giveaway: your binary reports 43 known types, which is exactly upstream master. The branch that reads this file defines GGML_TYPE_PQ2_0 = 142 and GGML_TYPE_COUNT = 144, and it imports the legacy type 43 (the embedded corrections) as PQ2_0 automatically, so nothing else is missing.
Cleanest path: throw the checkout away and start over. Re-running cmake in place tends to keep the old configuration, and an older llama-cli earlier on your PATH gives the same error, so a fresh tree is faster than debugging it.
rm -rf prism-ml-llama.cpp
git clone https://github.com/sky-is-green/prism-ml-llama.cpp
cd prism-ml-llama.cpp
./verify-container-support.sh # prints RESULT: OK
cmake -B build -DGGML_CUDA=ON && cmake --build build -j --target llama-cli llama-server
./build/bin/llama-cli -m Scion-35B-A3B-PQ2_0-corr.gguf -ngl 99 -c 4096 -t <physical cores>
I have just made moe-corr-runtime the fork's default branch, so a plain clone is the right tree now, and the branch README carries the same explanation. If you would rather keep what you have, git fetch origin && git checkout moe-corr-runtime && rm -rf build also works; just start from a clean build/ either way.
When it loads, please send git log -1 --format=%h, your GPU and VRAM, and pp512/tg128 t/s. CUDA is the one path I could not test myself, so a report from an NVIDIA box is exactly what I want. If anything else fails, paste the full log, and I'll try to help.
Big thanks for testing!