LLM inference in C/C++
manifesto / ggml / ops / maintainer PRs / compile times / lib llama API / llama-server REST API
HIVE (Hierarchical Inference Versatile Engine) is our local inference stack. This repo is its llama.cpp fork for diffusion language models (dLLMs): fast quantized GGUF inference on CPU/GPU, no PyTorch.
Diffusion LLMs unmask tokens in blocks over multiple steps (unlike standard left-to-right models). Upstream llama.cpp supports Dream, LLaDA, and RND1 via llama-diffusion-cli. This fork adds stable-diffcoder for Stable-DiffCoder — block-causal attention and KV-window decode (~5× faster). --diffusion-resident applies to all supported dLLMs in this build.
Experimental disclosure: HIVE patches (
stable-diffcoder, KV-window, resident mode, latent I/O) are under active development. APIs, performance, and correctness may change between releases. Not recommended for production without your own validation. Upstream llama.cpp diffusion support is likewise evolving.
Setup: docs/STABLE-DIFFCODER.md
llama-diffusion-cli only loads GGUFs whose metadata names a diffusion arch llama.cpp
knows about (general.architecture). It won't run arbitrary dLLMs from Hugging Face
unless they match one of these:
| GGUF arch | Model family | Notes |
|---|---|---|
dream |
Dream (Qwen2-based) | Timestep scheduling (--diffusion-eps). See examples. |
llada |
LLaDA | Fully non-causal; block scheduling (--diffusion-block-length). |
llada-moe |
LLaDA MoE | Same diffusion path as llada. |
rnd1 |
RND1 | Entropy / timestep scheduling. |
stable-diffcoder |
Stable-DiffCoder (HIVE) | Block-causal + KV window — recommended for DiffCoder weights. |
| You have… | Works? |
|---|---|
| Dream / LLaDA / RND1 GGUF, arch tag already correct | Yes — see examples/diffusion/README.md |
Stable-DiffCoder GGUF tagged llama (the usual mradermacher release) |
Yes — scripts/retag_gguf_diffusion.py --target-arch stable-diffcoder |
Stable-DiffCoder re-tagged as llada only |
Runs, but the non-causal mask can duplicate tokens (total total, etc.) |
| dLLM with no llama.cpp arch (CoDA, one-off HF exports, …) | No |
GGUF still tagged autoregressive llama |
No — diffusion path never engages |
| Feature | Dream / LLaDA / RND1 | stable-diffcoder |
|---|---|---|
llama-diffusion-cli |
Yes (upstream) | Yes |
--diffusion-resident |
Yes | Yes |
| Block-causal mask + KV-window decode | No | Yes |
--diffusion-hidden-out / --diffusion-control-prefix |
No | Yes |
Other dLLMs: Resident mode is arch-agnostic today. Block-causal attention, KV caching, and latent I/O require a model trained with frozen blocks across denoise steps — then add an arch module (same pattern as src/models/stable-diffcoder.cpp). Fully non-causal models (Dream, LLaDA, RND1) cannot use KV-window as implemented; wrong mask on block-causal weights produces duplicated tokens.
A few options to get llama.cpp installed on your machine:
- Visit https://llama.app and follow the instructions
- Run with Docker - see our Docker documentation
- Download pre-built binaries from the releases page
- Build from source by cloning this repository - check out our build guide
Once installed:
# Download and run a model directly from Hugging Face
llama cli -hf ggml-org/Qwen3.5-0.8B-GGUF
# Launch OpenAI-compatible API server
llama serve -hf ggml-org/Qwen3.5-0.8B-GGUF
VLM session with llama cli
|
Built-in web UI against llama serve
|
The main goal of llama.cpp is to enable LLM (and VLM) inference with minimal setup and state-of-the-art performance on
a wide range of hardware - locally and in the cloud.
- Plain C/C++ implementation without any dependencies
- Apple silicon is a first-class citizen - optimized via ARM NEON, Accelerate and Metal frameworks
- AVX, AVX2, AVX512 and AMX support for x86 architectures
- RVV, ZVFH, ZFH, ZICBOP and ZIHINTPAUSE support for RISC-V architectures
- 1.5-bit, 2-bit, 3-bit, 4-bit, 5-bit, 6-bit, and 8-bit integer quantization for faster inference and reduced memory use
- Custom CUDA kernels for running LLMs on NVIDIA GPUs (support for AMD GPUs via HIP and Moore Threads GPUs via MUSA)
- Vulkan and SYCL backend support
- CPU+GPU hybrid inference to partially accelerate models larger than the total VRAM capacity
The llama.cpp project is build on top of the ggml library.
| Backend | Target devices |
|---|---|
| BLAS | All |
| BLIS | All |
| CANN | Ascend NPU |
| CUDA | Nvidia GPU |
| HIP | AMD GPU |
| Hexagon [In Progress] | Snapdragon |
| IBM zDNN | IBM Z & LinuxONE |
| MUSA | Moore Threads GPU |
| Metal | Apple Silicon |
| OpenCL | Adreno GPU |
| OpenVINO [In Progress] | Intel CPUs, GPUs, and NPUs |
| RPC | All |
| SYCL | Intel GPU |
| VirtGPU | VirtGPU APIR |
| Vulkan | GPU |
| WebGPU | All |
| ZenDNN | AMD CPU |
- How to build
- Running on Docker
- Build on Android
- Multi-GPU usage
- Performance troubleshooting
- GGML tips & tricks
- XCFramework
- Completions
- Models
- Release process
- Contributors can open PRs
- Collaborators will be invited based on contributions
- Maintainers can push to branches in the
llama.cpprepo and merge PRs into themasterbranch - Any help with managing issues, PRs and projects is very appreciated!
- Read the CONTRIBUTING.md for more information
- yhirose/cpp-httplib - Single-header HTTP server, used by
llama-server- MIT license - stb-image - Single-header image format decoder, used by multimodal subsystem - Public domain
- nlohmann/json - Single-header JSON library, used by various tools/examples - MIT License
- miniaudio.h - Single-header audio format decoder, used by multimodal subsystem - Public domain
- subprocess.h - Single-header process launching solution for C and C++ - Public domain