WasmEdge Files

0.14.1 (2024-09-16)

Features:

• Supported LLVM 17.0.6.
• Bumpped spdlog to v1.13.0.
• Bumpped fmt to 11.0.2.
• Bumpped simdjson to v3.10.0.
• Bumpped googletest to 1.15.2.
• [WASI-NN] ggml backend:
• Bump llama.cpp to b3651.
• Static link libggml and libllama.
• Refined the CMake to support multiple backends of WASI-NN with ggml backend.
• Supported compute single in RPC mode.
• [WASI-NN] Added support for whisper.cpp backend.
• [WASI-NN] Added support for piper backend.
• [WASI-NN] Added support for ChatTTS backend.
• [WASI-NN] Added support for Burn.rs backend.
• Supported squeezenet and whisper models.
• [Plugin] Supported wasmedge_stablediffusion plug-in.
• Enabled CUBLAS.
• Enabled metal support on MacOS.
• [Plugin] Moved wasi_logging into built-in plug-in.
• Instead of installing wasi_logging plug-in shared library, developers can find and get this plug-in after calling WasmEdge_PluginLoadWithDefaultPaths() API.
• In the WasmEdge CLI tools, the built-in plug-ins will automatically be loaded.
• [Proposal] Initial support for instantiation phase of component model.
• Due to the breaking change of API, bump the plug-in API_VERSION to 3.
• [Proposal] Supported WASM Relaxed-SIMD proposal.
• Added the WasmEdge_Proposal_RelaxSIMD for the configuration in WasmEdge C API.
• Users can use the --enable-relaxed-simd to enable the proposal in wasmedge and wasmedgec tools.

Fixed issues:

• Fixed warnings on GCC-14.
• Fixed the fmt related header inclusion for error logging.
• Fixed WASI test error in Windows.
• Fixed version checking in source tarball.
• Fixed version detection issue when building from source.
• Fixed the visibility of internal symbols.
• [Loader] Fixed alignment checking in loading immediates for memory instructions.
• [Runtime] Fixed allocation issue when configured the limited memory page size.
• Used fmt::format instead of string stream in error logging.

Tests:

• Added WASI test suites on Windows.

Known issues:

• Universal WASM format failed on macOS platforms.
• In the current status, the universal WASM format output of the AOT compiler with the O1 or upper optimizations on MacOS platforms will cause a bus error during execution.
• We are trying to fix this issue. For a working around, please use the --optimize=0 to set the compiler optimization level to O0 in wasmedgec CLI.

Thank all the contributors who made this release possible!

Biswapriyo Nath, Elmira, Faidon Liambotis, Fusaaaann, Han-Wen Tsao, Jun Zhang, Kefu Chai, Lîm Tsú-thuàn, Michael Morris, PeterD1524, Shen-Ta Hsieh, Shreyas Atre, Sylveon, Yi Huang, Yi-Ying He, alabulei1, dm4, grorge, hydai, junxiangMu, vincent

If you want to build from source, please use WasmEdge-0.14.1-src.tar.gz instead of the zip or tarball provided by GitHub directly.

Notes: * The Stable Diffusion plugin is updated to use 0.14.1-plugin-lts-d259eed to fix the segmentation fault issue. For more details, please check the branch 0.14.1-plugin-lts. * llama.cpp is using b5896 and supports DeepSeek, Phi-4-mini, TTS, Gemma-3, Gemma-3N, and many more advanced features. For the multi-modal models, we also support llava, gemma-3, qwen-2-vl, qwen-2.5-vl, and Mistral Small 3.1. * If you are using an AMD GPU with ROCm, please use the WasmEdge-plugin-wasi_nn-ggml-rocm-0.14.1-ubuntu24.04_x86_64.tar.gz asset. We have only verified that it works on the MI300X. * Due to the limitation of JetPack 6.2, please download WasmEdge-plugin-wasi_nn-ggml-cuda-0.14.1-ubuntu22.04_aarch64.tar.gz if you are using Jetson devices.