Ferro
Tiny C++ inference engine that runs GGUF embeddings, decisions and chat on Intel N100 iGPU.
Tested on Linux
Use it
Linux x86_64
You need first
- A GPU is optional
These steps ran from a folder named work in the home folder, which already existed. Make it and run them from there:
mkdir -p "$HOME/work" && cd "$HOME/work"fresh clone of paragontasx/ferro and KhronosGroup/OpenCL-Headers, then make -j4 CL_INC=<headers> builds ./ferro-embed
cd "$HOME"/work && rm -rf ferro OpenCL-Headers && git clone --depth 1 https://github.com/paragontasx/ferro ferro && git clone --depth 1 https://github.com/KhronosGroup/OpenCL-Headers OpenCL-Headers && cd ferro && make -j4 CL_INC="$HOME"/work/OpenCL-Headers
Where the README and reality differ
- README: make -j4 # needs OpenCL headers: CL_INC=../ferro-gpu/third_party/OpenCL-Headers (default)
What worked: The default CL_INC path does not exist in the repo or next to it. Clone https://github.com/KhronosGroup/OpenCL-Headers and pass CL_INC=<that path> to make (the make test target also needs the same CL_INC). - README: make test # kernel + decision + serving tests
What worked: Without CL_INC on the command line, make test fails with 'CL/cl.h: No such file or directory' in tests/test_kernels. With CL_INC passed it passes. Also the prompt tests are skipped without the d1 model, so the 'decision' coverage here is partial.
Verified in pickfit: the commands exactly as its sandbox ran them
For the record only: they name the sandbox's home, /home/sbx.
cd /home/sbx/work && rm -rf ferro OpenCL-Headers && git clone --depth 1 https://github.com/paragontasx/ferro ferro && git clone --depth 1 https://github.com/KhronosGroup/OpenCL-Headers OpenCL-Headers && cd ferro && make -j4 CL_INC=/home/sbx/work/OpenCL-Headerscd /home/sbx/work/ferro && make test CL_INC=/home/sbx/work/OpenCL-Headerscd /home/sbx/work/ferro && make test CL_INC=/home/sbx/work/OpenCL-HeadersCheck that it works
cd "$HOME"/work/ferro && make test CL_INC="$HOME"/work/OpenCL-HeadersThe output should contain “ALL PASS”.
First thing to try
cd "$HOME"/work/ferro && make test CL_INC="$HOME"/work/OpenCL-HeadersWhat it does
- Generate sentence and multimodal embeddingsNot tested
- Run d1-omni typed decisionsNot tested
- Chat and text generation with LFM2.5Not tested
- OpenAI-compatible HTTP server with prefix cacheNot tested
How far pickfit got
Linux x86_64
- Installs
- Starts
- 0 of 4 features work
- Works offline
Good for
- compute sentence embeddings from text and media
- answer typed yes/no, choice and score decisions over a state
- generate chat text and completions with streaming
An alternative to llama.cpp.
How pickfit tested it
Each run starts from a clean machine and follows the steps above. How pickfit verifies
Linux x86_64
Its main function works
- Replayed from scratch
- Works offline
- Checked on 16 Nov 2026
Differences the test machine hit
- README: Mesa rusticl is enough (no Intel compute-runtime needed) / OpenCL headers prerequisite
In the test: This box has no libOpenCL, so the tool printed '[gpu] libOpenCL not found -> CPU' and ran the CPU path. The GPU DP4A test was SKIPPED (no capable GPU). - README: ./ferro-embed --help (implied by the usage text)
In the test: There is no --version. '--help' prints the usage text to stderr and exits 2, the same code as a missing command. - README: Models are plain GGUFs (--model kind is auto-detected)
In the test: The tiny llama-architecture test GGUF stories260K is rejected: 'init failed: not a causal LFM2 language model (arch lfm2 without a decision head expected)'. Only lfm2, gemma-embedding2 and gemma4 GGUFs load. - README: make -j4 (Quick start: 2 minutes)
In the test: The build worked with g++ 14 in about 26 s; it prints a -Wstringop-overflow warning in mm_preproc.cpp, which is harmless here.