From b2e89a327457179a34eae4d7de0d412ed945679c Mon Sep 17 00:00:00 2001 From: Dan Johansson <164997844+eddnjjn@users.noreply.github.com> Date: Mon, 9 Sep 2024 09:02:45 +0200 Subject: [PATCH] Arm AArch64: Documentation updates (#9321) * Arm AArch64: Documentation updates * Update docs/build.md to include information on how to enable the Arm optimized gemm/gemv kernels * Update examples/quantize/README.md with information on the Q4_0_4_4, Q4_0_4_8 and Q4_0_8_8 formats * Add newline to the end of docs/build.md --- docs/build.md | 6 ++++++ examples/quantize/README.md | 2 ++ 2 files changed, 8 insertions(+) diff --git a/docs/build.md b/docs/build.md index 152d46d6f31af..faa0ecfa49768 100644 --- a/docs/build.md +++ b/docs/build.md @@ -380,3 +380,9 @@ For detailed info, such as model/device supports, CANN install, please refer to ### Android To read documentation for how to build on Android, [click here](./android.md) + +### Arm CPU optimized mulmat kernels + +Llama.cpp includes a set of optimized mulmat kernels for the Arm architecture, leveraging Arm® Neon™, int8mm and SVE instructions. These kernels are enabled at build time through the appropriate compiler cpu-type flags, such as `-DCMAKE_C_FLAGS=-march=armv8.2a+i8mm+sve`. Note that these optimized kernels require the model to be quantized into one of the formats: `Q4_0_4_4` (Arm Neon), `Q4_0_4_8` (int8mm) or `Q4_0_8_8` (SVE). The SVE mulmat kernel specifically requires a vector width of 256 bits. When running on devices with a different vector width, it is recommended to use the `Q4_0_4_8` (int8mm) or `Q4_0_4_4` (Arm Neon) formats for better performance. Refer to [examples/quantize/README.md](../examples/quantize/README.md) for more information on the quantization formats. + +To support `Q4_0_4_4`, you must build with `GGML_NO_LLAMAFILE=1` (`make`) or `-DGGML_LLAMAFILE=OFF` (`cmake`). diff --git a/examples/quantize/README.md b/examples/quantize/README.md index 5d1e11c67b13f..704f0d56bea72 100644 --- a/examples/quantize/README.md +++ b/examples/quantize/README.md @@ -54,6 +54,8 @@ As the models are currently fully loaded into memory, you will need adequate dis Several quantization methods are supported. They differ in the resulting model disk size and inference speed. +The quantization formats `Q4_0_4_4`, `Q4_0_4_8` and `Q4_0_8_8` are block interleaved variants of the `Q4_0` format, providing a data layout that is better suited for specific implementations of optimized mulmat kernels. Since these formats differ only in data layout, they have the same quantized size as the `Q4_0` format. + *(outdated)* | Model | Measure | F16 | Q4_0 | Q4_1 | Q5_0 | Q5_1 | Q8_0 |