IntentDSL preview

Installable IntentDSL compiler package from the source at this preview tag.

• English handbook · 中文工具书
• Linux x86-64, glibc 2.35 or newer; maintained backend environments use Python 3.10–3.12.
• The wheel includes the compiler, optimizer and non-system compiler libraries. Installation does not require an LLVM/MLIR build SDK.
• Download the wheel and install it with python -m pip install "./intentdsl-…whl[manual,examples]", using its actual filename. Then run intent setup --target triton and intent doctor --target triton. Use a separate environment for cuTile.
• The distribution workflow built the wheel from the source archive and checked metadata, native dependencies, isolated installation, compiler tools and MCP entry points. Hardware execution is separate from these build checks.
• PyPI uploads are selected separately in Distribution. The handbook covers PyPI installation, source builds and CPU/MLU toolchains.

本附件由同一份公开源码构建，包含编译器、优化器与所需的非系统编译库，安装无需 LLVM/MLIR SDK。下载 wheel 后使用 pip 安装，再选择 Triton 或 cuTile 后端；两者使用独立环境。构建流程检查源码包、平台依赖、隔离安装、编译工具和 MCP 入口；硬件运行与这些构建检查分别说明。PyPI 上传与 GitHub 预览由维护者分别选择，安装方式见工具书。许可证为 Apache-2.0，第三方依赖保留自己的许可声明。

This snapshot updates the README to the original paper's whole-operator performance comparisons: two generated GPU targets on H100/RTX 5090D (44-workload geometric means), seven native x86 comparisons, and six MLU library comparisons. The pass-ablation figure has been removed from the current presentation. The figures retain their historical paper evaluation scope. Compiler, runtime and algorithm sources are unchanged from the prior H100-validated package.

本次 README 改为论文原整体性能图：H100/5090D 的双 GPU 后端对比与44项几何均值、7项 x86 原生实现对比、6项 MLU 库对比。当前展示已移除 pass 消融图；图中数据仍为论文历史评测。编译器、运行时与算法源码未改变。