Metadata-Version: 2.4
Name: fpt-agent
Version: 0.1.0
Summary: Hardware-aware mixed-precision tuning with guarded LLM planning and native numerical analysis
License-Expression: MIT
Keywords: mixed precision,numerical validation,high performance computing,CUDA
Classifier: Development Status :: 3 - Alpha
Classifier: Environment :: Console
Classifier: Operating System :: POSIX :: Linux
Classifier: Programming Language :: Python :: 3 :: Only
Classifier: Topic :: Scientific/Engineering
Classifier: Topic :: Software Development :: Compilers
Requires-Python: >=3.11
Description-Content-Type: text/markdown
License-File: LICENSE
Requires-Dist: pydantic==2.13.4
Provides-Extra: research
Requires-Dist: matplotlib==3.11.2; extra == "research"
Requires-Dist: numpy==2.4.6; extra == "research"
Provides-Extra: dev
Requires-Dist: pytest==9.0.3; extra == "dev"
Requires-Dist: cmake==4.4.4; extra == "dev"
Requires-Dist: jsonschema==4.25.1; extra == "dev"
Requires-Dist: numpy==2.4.6; extra == "dev"
Dynamic: license-file

# FPT-Agent

FPT-Agent tunes floating-point precision in supported C/C++ and CUDA programs. It checks candidate output against a specified numerical reference or tolerance, measures execution time or memory, and can explore compatible SIMD, SIMT and Tensor Core implementations. Use conventional search without a model or configure an LLM planner separately.

## Install the Python interface

```bash
python -m pip install fpt-agent
python -c "import fpt_agent; print(fpt_agent.__version__)"
precision-agent --version
precision-agent doctor
precision-agent hardware probe --json
```

The wheel provides the Python API and the `precision-agent` command. It supports hardware inspection, input generation and model configuration checks on Linux x86-64 with Python 3.11 or later.

## Build the native tools

For source analysis, stochastic arithmetic, mixed-precision transformation or CPU/CUDA tuning, unpack the matching source distribution or check out the matching release. From its top-level directory, configure the C++20 compiler, Boost headers and LLVM/Clang 21.1.8 as described in `docs/en/installation.md`. Then run:

```bash
./scripts/bootstrap_m5.sh
source work/venv/bin/activate
precision-agent doctor
```

To use the wheel from a different environment with this build, set `FPT_SOURCE_ROOT` to the absolute path of that source directory and `FPT_LLVM_ROOT` to the LLVM installation prefix. CUDA additionally requires a compatible NVIDIA driver, CUDA Toolkit and device. The source distribution includes the native build files, installation guide and examples. Model weights and compiler SDKs are installed separately.
