Note
本分支 prism-ptq1_0-avx2 是 PrismML-Eng/llama.cpp 的 AVX2 优化分支,
基于上游 prism 分支的 d7b2896,
为 PTQ1_0 三值格式(如 Ternary-Bonsai-2-27B)新增了一个 CPU 重排内核,
面向没有 AVX-512 VNNI 的 AVX2 处理器(Haswell 及之后的同类 CPU)。
改动只涉及 ggml/src/ggml-cpu/repack.cpp 一个文件。
测试模型 Ternary-Bonsai-2-27B-PTQ1_0(27B,权重约 6GB),4 核 AVX2(无 VNNI)CPU + 双通道 DDR3,纯 CPU 推理,4 线程:
| 项目 | 上游 CPU 路径 | 本分支 |
|---|---|---|
| 生成速度 (tg) | 1.05 t/s | 2.87–2.91 t/s |
| 提示处理 (pp64) | 1.03 t/s | 6.73 t/s |
生成阶段已接近内存带宽上限。贪婪解码输出与上游路径逐字一致。
加载时把 PTQ1_0 权重重排成适合 AVX2 的布局(不超过原始行大小):
每行: [G 组 × 128 字节:纯 5-trit 打包,每字节已异或 0x80]
[r 个尾块 × 32 字节,每字节 4 个 trit]
[nb 个 fp16 缩放]
其中 nb = 每行 128 元素块数,G = nb / 5,r = nb % 5
- 解 trit:同一个 32 字节向量连续乘 3(
x + x + x),每次用两次有符号比较t = [x > -43] - [x > 42]取出最高位 trit(得到 w+1 ∈ {0,1,2}),不需要查表、洗牌或移位。 字节预先异或 0x80,乘 3 后偏置不变(384 ≡ 128 mod 256)。 - 点积:
vpmaddubsw(t, q)计算(w+1)·q,多出的Σq偏移项在激活量化时预先算好(ds = d·Σq),每行最后统一向量化扣除。 - 按组收尾:每组 5 个块用一次
vcvtph2ps转换 5 个缩放。 - 多列共享解码:提示处理时一次解码的权重同时与 4 列激活相乘;行再切成约 64KB 的小块,让权重在遍历所有列期间留在 L2。
- 软件预取:提前 2KB 预取,弥补硬件预取在 4KB 页边界处中断的问题。
- 编译器屏障:LLVM 会把连续的
x*3合并成乘 9/27/81,而 AVX2 没有字节乘法,会展开成大量只能走 port 0 的指令,因此每次乘 3 后加__asm__("" : "+x"(x))。
启用条件:编译目标有 AVX2 且没有 AVX-512 VNNI,张量为 2D 的 PTQ1_0,且行内尾块能放进原行大小。
使用 GGML_CPU_ALL_VARIANTS 构建时,haswell 等对应变体会自动启用。
# OpenMP 构建时,-C/--cpu-strict 对 OpenMP 线程池无效,要用环境变量把线程分散到各物理核
export OMP_PLACES=cores OMP_PROC_BIND=spread
llama-server -m Ternary-Bonsai-2-27B-PTQ1_0.gguf --mmproj Ternary-Bonsai-2-27B-mmproj-BF16.gguf \
-t <物理核数> -lm dio -fa on -ctk q8_0 -ctv q8_0 -np 1 -cram 512- 绑核:Windows 等系统可能把两个计算线程放到同一物理核的两个超线程上,它们共用同一组执行端口, 矩阵乘法会被最慢的核拖住。绑核后生成速度约提升 25%。线程数取物理核数,开超线程反而更慢。
-lm dio:DirectIO 加载,系统不保留文件缓存,避免文件缓存和重排后的内存各占一份。-cram 512:llama-server 默认用 8GB 内存做提示缓存,内存较小的机器建议调低。
Important
This is the PrismML fork of llama.cpp, the main line behind the Bonsai models (branch prism, developed as prism-v7). It tracks current mainline llama.cpp and adds the fork's low-bit formats and runtime features on top.
New here? Start with the Bonsai-demo repo. It downloads the right models and the correct prebuilt binaries for your hardware/backend automatically.
Which ternary model file to use:
*-PQ2_0.gguf(fork group-128, ggml id 142): preferred on Metal, CUDA, HIP and CPU. About 6% smaller than group-64.*-Q2_0_g64.gguf/ 27B*-Q2_g64.gguf(official group-64, ggml id 42): runs on every backend here AND on mainline llama.cpp. If unsure, use this. Newer model releases name this file plain*-Q2_0.gguf.*-Q2_0.ggufon OLDER model repos is the deprecated legacy format (group 128 stored as id 42). It does not load on these builds; the error tells you which file to get instead. If you must run it, use the frozenprism-v5line and its final releaseprism-b9601.
Speculative decoding (dspark) is supported via mainline's draft-dspark plus fork patches. Drafters published for older model releases need a one-time conversion with gguf-dspark-to-dflash (see SPECULATIVE.md in Bonsai-demo); newer releases ship ready-to-use drafters.
Do NOT build from prism-v6 (stale mid-migration snapshot) and do NOT mix this fork's ggml-* libraries with a stock llama.cpp build.
LLM inference in C/C++
ggml / ops / maintainer PRs / dev stats / lib llama API / llama-server REST API
A few options to get llama.cpp installed on your machine:
# curl
curl -LsSf https://llama.app/install.sh | sh
# powershell
irm https://llama.app/install.ps1 | iex- 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 | 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 - nothings/stb - Single-header image format decoder, used by multimodal subsystem - Public domain
- nlohmann/json - Single-header JSON library, used by various tools/examples - MIT License
- mackron/miniaudio - Single-header audio format decoder, used by multimodal subsystem - Public domain
- sheredom/subprocess.h - Single-header process launching solution for C and C++ - Public domain

