pyproject.toml

[build-system]
requires = ["hatchling"]
build-backend = "hatchling.build"

[project]
name = "automegakernel"
version = "0.2.0"
description = "AutoMegaKernel (AMK), autonomously generate a single correct, near-bandwidth-optimal megakernel for any model on any GPU."
readme = "README.md"
requires-python = ">=3.12"
license = { text = "MIT" }
authors = [{ name = "RightNow AI" }]
keywords = ["gpu", "cuda", "megakernel", "compiler", "llm-inference", "kernel-fusion"]
classifiers = [
    "License :: OSI Approved :: MIT License",
    "Programming Language :: Python :: 3.12",
    "Operating System :: Microsoft :: Windows",
    "Operating System :: POSIX :: Linux",
    "Topic :: Scientific/Engineering",
    "Topic :: Software Development :: Compilers",
    "Intended Audience :: Developers",
]

# Core dependencies are GPU-toolchain-light on purpose: the IR, the DAG validator,
# the CPU reference VM, the oracle and the cost model all run with just torch + numpy.
# Triton is optional (its Windows support is a separate wheel) and the CUDA backend is
# built on demand via torch.utils.cpp_extension + nvcc.
dependencies = [
    "numpy>=2.0.0",
    "torch>=2.7.0",
]

[project.urls]
Homepage = "https://github.com/RightNow-AI/automegakernel"
Repository = "https://github.com/RightNow-AI/automegakernel"
Issues = "https://github.com/RightNow-AI/automegakernel/issues"

[project.optional-dependencies]
# Fast-iteration instruction backend. On Windows install `triton-windows` instead.
triton = ["triton>=3.3.0; platform_system != 'Windows'"]
# HuggingFace model import (the `any model` promise / graph importer).
models = ["transformers>=4.50.0", "accelerate>=1.5.0", "safetensors>=0.5.0"]
# Reporting / flywheel analysis.
analysis = ["pandas>=2.2.0", "matplotlib>=3.10.0"]
# CUDA JIT build helper.
cuda = ["ninja>=1.11.0"]
# Native coding-agent integration: the MCP server (amk_mcp.py / `amk-mcp`) that exposes the
# harness + autoresearch + orchestrator to Claude Code / Codex over the Model Context Protocol.
agent = ["mcp>=1.2"]

# The dev environment `uv sync` provisions by default (PEP 735 dependency-group). uv installs the
# `dev` group automatically, so a bare `uv sync` gives you the full test surface: pytest + ruff,
# the HF importer (transformers/accelerate/safetensors, the "any model" + checkpoint tests),
# the CUDA JIT helper (ninja), and the reporting deps. Plain `pip install automegakernel[models]`
# still works for downstream consumers via the [project.optional-dependencies] extras above.
[dependency-groups]
dev = [
    "pytest>=8.0.0",
    "ruff>=0.6.0",
    "ninja>=1.11.0",
    "transformers>=4.50.0",
    "accelerate>=1.5.0",
    "safetensors>=0.5.0",
    "pandas>=2.2.0",
    "matplotlib>=3.10.0",
]

# `amk` becomes a real console command after `uv sync` (-> `uv run amk doctor`).
[project.scripts]
amk = "amk_cli:main"
# The MCP server entry point (after `uv sync --extra agent` / `pip install automegakernel[agent]`).
amk-mcp = "amk_mcp:main"

# AMK is a research monorepo of top-level scripts + sibling packages. We ship it as a proper
# editable install (hatchling) so `uv run amk ...` works AND every `import vm` / `import schedule`
# / `import harness` keeps resolving exactly as it does under `uv run python amk_cli.py`.
[tool.hatch.build.targets.wheel]
# The sibling packages (each has an __init__.py; `dynamism` is an as-yet-empty Layer-3 placeholder).
packages = [
    "vm",
    "schedule",
    "instructions",
    "eval",
    "dynamism",
    "flywheel",
    "models",
]
# The top-level single-file modules the CLI / harness / product import by name.
#
# NOTE: `include = [...]` alone copies these as DATA (sdist/wheel payload), NOT as importable
# top-level modules, so `import harness` / the `amk` + `amk-mcp` console scripts would raise
# ModuleNotFoundError on a clean wheel install. `force-include` below maps each shipped root
# module to itself so hatchling places it at the wheel root as a real importable module. We
# enumerate EVERY root .py reachable from the console scripts / package imports
# (amk_cli -> harness, loop1, autoresearch, compile; amk_mcp -> harness, autoresearch,
# amk_orchestrate; harness -> generate, compile; etc.), EXCLUDING the gitignored modal_app.py
# (paper-results infra, not product) and conftest.py (pytest-only, never imported as a module).
include = [
    "amk_cli.py",
    "amk_mcp.py",
    "compile.py",
    "generate.py",
    "harness.py",
    "autoresearch.py",
    "amk_orchestrate.py",
    "loop1.py",
]

[tool.hatch.build.targets.wheel.force-include]
"amk_cli.py" = "amk_cli.py"
"amk_mcp.py" = "amk_mcp.py"
"compile.py" = "compile.py"
"generate.py" = "generate.py"
"harness.py" = "harness.py"
"autoresearch.py" = "autoresearch.py"
"amk_orchestrate.py" = "amk_orchestrate.py"
"loop1.py" = "loop1.py"

# RTX 5090 / Blackwell consumer (sm_120) and datacenter Blackwell (B200, sm_100) both
# need the cu128 wheel line. The cu128 runtime is forward-compatible with the installed
# CUDA 13 driver.
[[tool.uv.index]]
name = "pytorch-cu128"
url = "https://download.pytorch.org/whl/cu128"
explicit = true

[tool.uv.sources]
torch = { index = "pytorch-cu128" }

[tool.ruff]
line-length = 100
target-version = "py312"

[tool.pytest.ini_options]
testpaths = ["tests"]
addopts = "-ra -q"
markers = [
    "cuda: test requires an available CUDA GPU + nvcc (skipped otherwise)",
    "slow: long-running test",
]