Name and Version
$ ./build/bin/llama-cli --version
version: 8329 (fbaa95b)
built with GNU 11.4.0 for Linux aarch64
Operating systems
Linux, Other? (Please let us know in description)
GGML backends
CPU
Hardware
Graviton-3 / Graviton-4 / Google Tensor G5 / MediaTek Dimensity 9500
Models
Qwen2-VL 2B Q4_0 with Q8_0 quantized mmproj
Qwen2 1.5B-Instruct Q4_0
Qwen2 1.5B-Instruct Q8_0
Qwen2 1.5B-Instruct F16
Problem description & steps to reproduce
When building llama.cpp for Arm64 hardware with SVE or SVE2, running llama-cli on either Linux or Android produces seemingly random ASCII characters as output instead of legible text.
This seems to happen across various devices on both Linux and Android, with the unifying factor being an Arm64 CPU with SVE enabled. This happens both with KleidiAI enabled and disabled.
Steps to reproduce:
- Build llama.cpp with the ggml CPU backend:
Configure:
Linux:
cmake -DGGML_METAL=OFF -DGGML_BLAS=OFF -DGGML_ACCELERATE=OFF -DGGML_OPENMP=OFF -DGGML_OPENCL=OFF -DGGML_VULKAN=OFF -DGGML_KOMPUTE=OFF -DGGML_HIP=OFF -DGGML_CUDA=OFF -DGGML_MUSA=OFF -DGGML_BUILD_EXAMPLES=OFF -DLLAMA_BUILD_TESTS=ON -DCMAKE_BUILD_TYPE=Release -DGGML_LLAMAFILE=OFF -DLLAMA_CURL=OFF -DBUILD_SHARED_LIBS=OFF -B build
Android:
cmake -DCMAKE_TOOLCHAIN_FILE=$NDK_PATH/build/cmake/android.toolchain.cmake -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=android-28 -DGGML_OPENMP=OFF -DGGML_LLAMAFILE=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DBUILD_SHARED_LIBS=ON -B build
Build:
cmake --build build --config Release -j
- Run llama-cli -m "path/to/model.gguf" -p "Hello!" (or with any prompt, or in interactive mode and type the prompt)
First Bad Commit
a0c2b20
Relevant log output
Logs
build : b8329-fbaa95bc
model : Qwen2-1.5B-Instruct.Q4_0.gguf
modalities : text
available commands:
/exit or Ctrl+C stop or exit
/regen regenerate the last response
/clear clear the chat history
/read add a text file
> Tell me a joke!
+8+2A9.;&>D(5EC;0!;=3#++$@1$!8!.,G#F59,)+66DEED!E8&&';H,/E3,')-@+"(/<;B1B3+!F),'D
Name and Version
$ ./build/bin/llama-cli --version
version: 8329 (fbaa95b)
built with GNU 11.4.0 for Linux aarch64
Operating systems
Linux, Other? (Please let us know in description)
GGML backends
CPU
Hardware
Graviton-3 / Graviton-4 / Google Tensor G5 / MediaTek Dimensity 9500
Models
Qwen2-VL 2B Q4_0 with Q8_0 quantized mmproj
Qwen2 1.5B-Instruct Q4_0
Qwen2 1.5B-Instruct Q8_0
Qwen2 1.5B-Instruct F16
Problem description & steps to reproduce
When building llama.cpp for Arm64 hardware with SVE or SVE2, running llama-cli on either Linux or Android produces seemingly random ASCII characters as output instead of legible text.
This seems to happen across various devices on both Linux and Android, with the unifying factor being an Arm64 CPU with SVE enabled. This happens both with KleidiAI enabled and disabled.
Steps to reproduce:
Configure:
Build:
First Bad Commit
a0c2b20
Relevant log output
Logs