merve/llama.cpp-dgx-spark-gb10-sm121a

llama.cpp prebuilt for NVIDIA DGX Spark (GB10 / sm_121a)

Prebuilt llama.cpp binaries compiled specifically for the NVIDIA GB10 (DGX Spark) Blackwell GPU. These are provided because upstream llama.cpp ships no aarch64 + CUDA prebuilt binaries — on a Spark you normally have to build from source.

These binaries only run on a machine matching the build target below (aarch64 + GB10/sm_121 + CUDA 13.x runtime). On any other GPU/arch they fail with no kernel image is available for execution on the device.

Build details

llama.cpp version	b9860 (commit fdb1db877)
GPU target	CMAKE_CUDA_ARCHITECTURES=121a-real (GB10, sm_121a, native FP4)
Platform	Linux aarch64 (ARM Grace)
Compiler	GCC 13.3.0
CUDA	13.x toolkit

Built with:

cmake -B build -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=121a-real -DCMAKE_BUILD_TYPE=Release
cmake --build build --config Release -j$(nproc)

Usage

Download the files:

hf download merve/llama.cpp-dgx-spark-gb10-sm121a --local-dir ./llama-gb10
cd llama-gb10
chmod +x bin/llama-*

Run the server (the LD_LIBRARY_PATH entry for /usr/local/cuda-13/compat is required so the Blackwell compat driver libs are found at runtime):

export LD_LIBRARY_PATH="$PWD/bin:/usr/local/cuda-13/compat:$LD_LIBRARY_PATH"
./bin/llama-server -hf unsloth/NVIDIA-Nemotron-3-Nano-4B-GGUF --port 8081 --n-gpu-layers 99

Verify it targets your GPU:

./bin/llama-server --version   # version: 9860 (fdb1db877)

A successful GPU load prints compute capability 12.1 in the server log.

Contents

All user-facing llama-* tools (llama-server, llama-cli, llama-bench, llama-quantize, llama-embedding, llama-mtmd-cli, ...) plus the required shared libraries (libggml*.so*, libllama*.so*) under bin/. Test binaries are not included.

License

MIT, inherited from llama.cpp. See LICENSE. This is an unofficial community build and is not affiliated with the llama.cpp project or NVIDIA.