pyproject.toml

[build-system]
requires = ["setuptools>=64", "wheel"]
build-backend = "setuptools.build_meta"

[project]
name = "fmha_sm100"
version = "0.1.1"
description = "MSA — MiniMax Sparse Attention: dense FlashAttention + sparse top-k indexer (csrc JIT) and CuTe-DSL sparse attention for NVIDIA SM100"
readme = "README.md"
requires-python = ">=3.10"
# SPDX-License-Identifier: MIT
license = { text = "MIT" }
authors = [{ name = "MiniMax" }]
dependencies = [
    "torch",
    "apache-tvm-ffi",
    "jinja2",
    "ninja",
    "pybind11",
    "cuda-python",
    "nvidia-cutlass-dsl>=4.4.1",
    "quack-kernels>=0.2.10",
]

[project.urls]

# The Python package lives under python/. Both kernel stacks ship as package
# data and work from both editable and wheel installs:
#   * Dense csrc/ kernels are JIT-compiled at runtime (python/fmha_sm100/jit.py).
#     The CUDA sources, Jinja templates, and vendored CUTLASS headers live inside
#     the package directory so that JIT can locate them via __file__.
#   * The CuTe-DSL sparse stack (python/fmha_sm100/cute/) is exposed through the
#     importable `fmha_sm100.sparse` module (and lazily from the package root),
#     which adds that directory to sys.path and re-exports the public sparse API.
[tool.setuptools]
package-dir = { "" = "python" }
# The cute/ sparse sources are loaded via sys.path.insert at runtime by
# fmha_sm100/sparse.py (not imported as a submodule), so they are shipped as
# package data.
packages = ["fmha_sm100"]

[tool.setuptools.package-data]
fmha_sm100 = [
    "cute/**/*.py",
    "cute/**/*.cu",
    "cute/**/*.cuh",
    "cute/**/*.h",
    "cute/**/*.hpp",
    "cute/**/*.md",
    "cute/**/*.ini",
    "cute/**/*.txt",
    "cute/Makefile",
    "csrc/**/*",
    "cutlass/include/**/*",
    "cutlass/tools/util/include/**/*",
]