llamacpp-prebuilt

llama.cpp — universal prebuilt

One artifact, pulled onto any rented box, no compilation.

llamacpp-3653e6d-cuda12.8-universal-sm61to120-ubuntu22.04-x86_64.tar.zst 774 MB compressed / 1.2 GB extracted.

hf download gbuzhf/llamacpp-prebuilt <file> --local-dir .
mkdir -p llama.cpp/build && tar -C llama.cpp/build -xf <file>   # -> llama.cpp/build/bin

provision-apex.sh Phase 1 finds every target present and skips the build entirely.

Coverage

GPU — SASS for every CUDA arch from Pascal to Blackwell, plus PTX so anything newer JIT-compiles instead of failing:

CPU — GGML_NATIVE=OFF with GGML_CPU_ALL_VARIANTS=ON: 14 per-ISA backends (sse42 through sapphirerapids, znver*), selected at runtime. Nothing is pinned to the machine that compiled it.

CUDA runtime bundled — libcudart, libcublas, libcublasLt ship inside, so the artifact does not require a matching CUDA image. Only the NVIDIA driver must be new enough for the 12.8 runtime.

Backends load dynamically (GGML_BACKEND_DL=ON), so a box with no GPU falls back to the CPU backends rather than failing to start.

Contents

llama-quantize, llama-imatrix, llama-gguf-split, llama-server, llama-perplexity, llama-cli, plus shared libraries.

Not included: convert_hf_to_gguf.py and gguf-py, which are pure Python and come from the llama.cpp clone.

Build provenance

llama.cpp 3653e6d, CUDA 12.8.93, Ubuntu 22.04 (glibc 2.35), built on a 256-core EPYC 7V13 in ~9 minutes.

Note: CUDA 12.4 cannot target sm_120 — nvcc rejects the architecture and cmake aborts at enable_language(CUDA). A "universal" build made on the common nvidia/cuda:12.4.1-devel image is silently Blackwell-less. 12.8 is the floor.

Known limits

• x86-64 Linux only; no arm64.
• glibc 2.35 floor — will not run on older distributions.
• Compute capability 6.1 and above. Maxwell (50/52) is excluded.