echodict/llama.cpp

1# llama.cpp for OpenCL2 3- [Background](#background)4- [OS](#os)5- [Hardware](#hardware)6- [DataType Supports](#datatype-supports)7- [Model Preparation](#model-preparation)8- [CMake Options](#cmake-options)9- [Android](#android)10- [Windows 11 Arm64](#windows-11-arm64)11- [Linux](#Linux)12- [Known Issue](#known-issues)13- [TODO](#todo)14 15## Background16 17OpenCL (Open Computing Language) is an open, royalty-free standard for cross-platform, parallel programming of diverse accelerators found in supercomputers, cloud servers, personal computers, mobile devices and embedded platforms. OpenCL specifies a programming language (based on C99) for programming these devices and application programming interfaces (APIs) to control the platform and execute programs on the compute devices. Similar to CUDA, OpenCL has been widely used to program GPUs and is supported by most GPU vendors.18 19### Llama.cpp + OpenCL20 21The llama.cpp OpenCL backend is designed to enable llama.cpp on **Qualcomm Adreno GPU** firstly via OpenCL. Thanks to the portabilty of OpenCL, the OpenCL backend can also run on certain Intel GPUs such as those that do not have [SYCL](/docs/backend/SYCL.md) support although the performance is not optimal.22 23## OS24 25| OS      | Status  | Verified                                       |26|---------|---------|------------------------------------------------|27| Android | Support | Snapdragon 8 Gen 3, Snapdragon 8 Elite         |28| Windows | Support | Windows 11 Arm64 with Snapdragon X Elite       |29| Linux   | Support | Ubuntu 22.04 WSL2 with Intel 12700H            |30 31## Hardware32 33### Adreno GPU34 35**Verified devices**36 37| Adreno GPU                           | Status  |38|:------------------------------------:|:-------:|39| Adreno 750 (Snapdragon 8 Gen 3)      | Support |40| Adreno 830 (Snapdragon 8 Elite)      | Support |41| Adreno X85 (Snapdragon X Elite)      | Support |42 43> A6x GPUs with a recent driver and compiler are supported; they are usually found in IoT platforms.44However, A6x GPUs in phones are likely not supported due to the outdated driver and compiler.45 46## DataType Supports47 48| DataType               | Status                     |49|:----------------------:|:--------------------------:|50| Q4_0                   | Support                    |51| Q6_K                   | Support, but not optimized |52| Q8_0                   | Support                    |53| MXFP4                  | Support                    |54 55## Model Preparation56 57You can refer to the general [llama-quantize tool](/tools/quantize/README.md) for steps to convert a model in Hugging Face safetensor format to GGUF with quantization.58 59Currently we support `Q4_0` quantization and have optimized for it. To achieve best performance on Adreno GPU, add `--pure` to `llama-quantize` (i.e., make all weights in `Q4_0`). For example,60 61```sh62./llama-quantize --pure ggml-model-qwen2.5-3b-f16.gguf ggml-model-qwen-3b-Q4_0.gguf Q4_063```64 65Since `Q6_K` is also supported, `Q4_0` quantization without `--pure` will also work. However, the performance will be worse compared to pure `Q4_0` quantization.66 67### `MXFP4` MoE Models68 69OpenAI gpt-oss models are MoE models in `MXFP4`. The quantized model will be in `MXFP4_MOE`, a mixture of `MXFP4` and `Q8_0`.70For this quantization, there is no need to specify `--pure`.71For gpt-oss-20b model, you can directly [download](https://huggingface.co/ggml-org/gpt-oss-20b-GGUF) the quantized GGUF file in `MXFP4_MOE` from Hugging Face.72 73Although it is possible to quantize gpt-oss-20b model in pure `Q4_0` (all weights in `Q4_0`), it is not recommended since `MXFP4` has been optimized for MoE while `Q4_0` is not. In addition, accuracy should degrade with such pure `Q4_0` quantization.74Hence, using the default `MXFP4_MOE` quantization (see the link above) is recommended for this model.75 76> Note that the `Q4_0` model found [here](https://huggingface.co/unsloth/gpt-oss-20b-GGUF/blob/main/gpt-oss-20b-Q4_0.gguf) is a mixture of `Q4_0`, `Q8_0` and `MXFP4` and gives better performance than `MXFP4_MOE` quantization.77 78## CMake Options79 80The OpenCL backend has the following CMake options that control the behavior of the backend.81 82| CMake options                     | Default value  | Description                               |83|:---------------------------------:|:--------------:|:------------------------------------------|84| `GGML_OPENCL_EMBED_KERNELS`       | `ON`           | Embed OpenCL kernels into the executable. |85| `GGML_OPENCL_USE_ADRENO_KERNELS`  | `ON`           | Use kernels optimized for Adreno.         |86 87## Android88 89Ubuntu 22.04 is used for targeting Android. Make sure the following tools are accessible from command line,90 91* Git92* CMake 3.2993* Ninja94* Python395 96### I. Setup Environment97 981. **Install NDK**99 100```sh101cd ~102wget https://dl.google.com/android/repository/commandlinetools-linux-8512546_latest.zip && \103unzip commandlinetools-linux-8512546_latest.zip && \104mkdir -p ~/android-sdk/cmdline-tools && \105mv cmdline-tools latest && \106mv latest ~/android-sdk/cmdline-tools/ && \107rm -rf commandlinetools-linux-8512546_latest.zip108 109yes | ~/android-sdk/cmdline-tools/latest/bin/sdkmanager "ndk;26.3.11579264"110```111 1122. **Install OpenCL Headers and Library**113 114```sh115mkdir -p ~/dev/llm116cd ~/dev/llm117 118git clone https://github.com/KhronosGroup/OpenCL-Headers && \119cd OpenCL-Headers && \120cp -r CL ~/android-sdk/ndk/26.3.11579264/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include121 122cd ~/dev/llm123 124git clone https://github.com/KhronosGroup/OpenCL-ICD-Loader && \125cd OpenCL-ICD-Loader && \126mkdir build_ndk26 && cd build_ndk26 && \127cmake .. -G Ninja -DCMAKE_BUILD_TYPE=Release \128  -DCMAKE_TOOLCHAIN_FILE=$HOME/android-sdk/ndk/26.3.11579264/build/cmake/android.toolchain.cmake \129  -DOPENCL_ICD_LOADER_HEADERS_DIR=$HOME/android-sdk/ndk/26.3.11579264/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include \130  -DANDROID_ABI=arm64-v8a \131  -DANDROID_PLATFORM=24 \132  -DANDROID_STL=c++_shared && \133ninja && \134cp libOpenCL.so ~/android-sdk/ndk/26.3.11579264/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/lib/aarch64-linux-android135```136 137### II. Build llama.cpp138 139```sh140cd ~/dev/llm141 142git clone https://github.com/ggml-org/llama.cpp && \143cd llama.cpp && \144mkdir build-android && cd build-android145 146cmake .. -G Ninja \147  -DCMAKE_TOOLCHAIN_FILE=$HOME/android-sdk/ndk/26.3.11579264/build/cmake/android.toolchain.cmake \148  -DANDROID_ABI=arm64-v8a \149  -DANDROID_PLATFORM=android-28 \150  -DBUILD_SHARED_LIBS=OFF \151  -DGGML_OPENCL=ON152 153ninja154```155 156## Windows 11 Arm64157 158A Snapdragon X Elite device with Windows 11 Arm64 is used. Make sure the following tools are accessible from command line,159 160* Git161* CMake 3.29162* Clang 19163* Ninja164* Visual Studio 2022165* Powershell 7166* Python167 168Visual Studio provides necessary headers and libraries although it is not directly used for building.169Alternatively, Visual Studio Build Tools can be installed instead of the full Visual Studio.170 171> Note that building using Visual Studio's cl compiler is not supported. Clang must be used. Clang depends on libraries provided by Visual Studio to work. Therefore, Visual Studio must be installed. Alternatively, Visual Studio Build Tools can be installed instead of the full Visual Studio.172 173Powershell 7 is used for the following commands.174If an older version of Powershell is used, these commands may not work as they are.175 176### I. Setup Environment177 1781. **Install OpenCL Headers and Library**179 180```powershell181mkdir -p ~/dev/llm182 183cd ~/dev/llm184git clone https://github.com/KhronosGroup/OpenCL-Headers && cd OpenCL-Headers185mkdir build && cd build186cmake .. -G Ninja `187  -DBUILD_TESTING=OFF `188  -DOPENCL_HEADERS_BUILD_TESTING=OFF `189  -DOPENCL_HEADERS_BUILD_CXX_TESTS=OFF `190  -DCMAKE_INSTALL_PREFIX="$HOME/dev/llm/opencl"191cmake --build . --target install192 193cd ~/dev/llm194git clone https://github.com/KhronosGroup/OpenCL-ICD-Loader && cd OpenCL-ICD-Loader195mkdir build && cd build196cmake .. -G Ninja `197  -DCMAKE_BUILD_TYPE=Release `198  -DCMAKE_PREFIX_PATH="$HOME/dev/llm/opencl" `199  -DCMAKE_INSTALL_PREFIX="$HOME/dev/llm/opencl"200cmake --build . --target install201```202 203### II. Build llama.cpp204 205```powershell206 207mkdir -p ~/dev/llm208cd ~/dev/llm209 210git clone https://github.com/ggml-org/llama.cpp && cd llama.cpp211mkdir build && cd build212 213cmake .. -G Ninja `214  -DCMAKE_TOOLCHAIN_FILE="$HOME/dev/llm/llama.cpp/cmake/arm64-windows-llvm.cmake" `215  -DCMAKE_BUILD_TYPE=Release `216  -DCMAKE_PREFIX_PATH="$HOME/dev/llm/opencl" `217  -DBUILD_SHARED_LIBS=OFF `218  -DGGML_OPENCL=ON219ninja220```221 222## Linux223 224The two steps just above also apply to Linux. When building for linux, the commands are mostly the same as those for PowerShell on Windows, but in the second step they do not have the `-DCMAKE_TOOLCHAIN_FILE` parameter, and then in both steps the backticks are replaced with back slashes.225 226If not installed already, install Git, CMake, Clang, Ninja and Python, then run in the terminal the following:227 228### I. Setup Environment229 2301. **Install OpenCL Headers and Library**231 232```bash233mkdir -p ~/dev/llm234 235cd ~/dev/llm236git clone https://github.com/KhronosGroup/OpenCL-Headers && cd OpenCL-Headers237mkdir build && cd build238cmake .. -G Ninja \239  -DBUILD_TESTING=OFF \240  -DOPENCL_HEADERS_BUILD_TESTING=OFF \241  -DOPENCL_HEADERS_BUILD_CXX_TESTS=OFF \242  -DCMAKE_INSTALL_PREFIX="$HOME/dev/llm/opencl"243cmake --build . --target install244 245cd ~/dev/llm246git clone https://github.com/KhronosGroup/OpenCL-ICD-Loader && cd OpenCL-ICD-Loader247mkdir build && cd build248cmake .. -G Ninja \249  -DCMAKE_BUILD_TYPE=Release \250  -DCMAKE_PREFIX_PATH="$HOME/dev/llm/opencl" \251  -DCMAKE_INSTALL_PREFIX="$HOME/dev/llm/opencl"252cmake --build . --target install253```254 255### II. Build llama.cpp256 257```bash258mkdir -p ~/dev/llm259cd ~/dev/llm260 261git clone https://github.com/ggml-org/llama.cpp && cd llama.cpp262mkdir build && cd build263 264cmake .. -G Ninja \265  -DCMAKE_BUILD_TYPE=Release \266  -DCMAKE_PREFIX_PATH="$HOME/dev/llm/opencl" \267  -DBUILD_SHARED_LIBS=OFF \268  -DGGML_OPENCL=ON269ninja270```271 272## Known Issues273 274- Flash attention does not always improve performance.275- Currently OpenCL backend works on A6xx GPUs with recent drivers and compilers (usually found in IoT platforms).276  However, it does not work on A6xx GPUs found in phones with old drivers and compilers.277 278## TODO279 280- Optimization for Q6_K281- Support and optimization for Q4_K282- Improve flash attention283