pyproject.toml

# Copyright (c) 2024 by FlashInfer team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#   http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

[project]
name = "flashinfer-python"
dynamic = ["version", "dependencies"]
description = "FlashInfer: Kernel Library for LLM Serving"
# Keep the lower bound in sync with .python-version.
requires-python = ">=3.10,<4.0"
authors = [{ name = "FlashInfer team" }]
license = "Apache-2.0"
readme = "README.md"
urls = { Homepage = "https://github.com/flashinfer-ai/flashinfer" }
license-files = ["LICENSE", "LICENSE*.txt"]

[project.optional-dependencies]
# Keep these CUDA-extra floors in sync with ci/cuda-versions.json.
cu12 = ["nvidia-cutlass-dsl>=4.7.0a0"]
cu13 = ["nvidia-cutlass-dsl[cu13]>=4.7.0a0"]
# DEPRECATED alias, kept so existing `pip install ".[nvep]"` commands keep
# working. The moe_ep runtime deps (cuda-python, nccl-extensions) are now part of the
# BASE dependencies (requirements.txt) and the NIXL-EP submodule build runs by
# default (best-effort) on `pip install .` — see the moe_ep section at the top
# of build_backend.py. libnccl comes from torch's own nvidia-nccl-cu13 pin;
# the >=2.30.7 B200 floor is enforced at runtime (moe_ep/core/validation/
# common.py) and upgraded --no-deps by the build hook on source installs. Opt
# out of the native build with BUILD_NVEP=0 (or BUILD_NIXL_EP=0 / BUILD_NCCL_EP=0).
nvep = []

[project.scripts]
flashinfer = "flashinfer.__main__:cli"

[build-system]
requires = ["setuptools>=77", "packaging>=24", "apache-tvm-ffi>=0.1.11,<0.2"]
build-backend = "build_backend"
backend-path = ["."]

[tool.codespell]
ignore-words-list = "3nd"
skip = ["build", "3rdparty", "dist", ".venv"]

[tool.setuptools]
include-package-data = false
py-modules = ["build_backend", "build_utils"]

[tool.setuptools.dynamic]
version = { attr = "flashinfer._build_meta.__version__" }
dependencies = { file = ["requirements.txt"] }

[tool.setuptools.packages.find]
where = ["."]
include = ["flashinfer*"]
exclude = ["flashinfer-jit-cache*", "flashinfer-cubin*"]

[tool.setuptools.package-dir]
"flashinfer.data" = "."
"flashinfer.data.cutlass" = "3rdparty/cutlass"
"flashinfer.data.spdlog" = "3rdparty/spdlog"
"flashinfer.data.cccl" = "3rdparty/cccl"

[tool.setuptools.package-data]
"flashinfer" = ["_build_meta.py"]
"flashinfer.prims_ts.moe" = ["prims_ts_moe_configs.json"]
# Workload registry of the experimental fused GDN decode backends, loaded at
# runtime via importlib.resources.
"flashinfer.gdn_kernels.experimental" = ["*.json"]
# Their in-package JIT kernel source, compiled on demand by
# flashinfer/jit/gdn_fused_decode.py.
"flashinfer.gdn_kernels.experimental.kernel" = ["*.cu"]
# Generated sources for the experimental exact-SM103a MXFP8 MegaMoE backend.
"flashinfer.experimental.cake_mxfp8_megamoe_ep16" = ["csrc/**"]
# Generated sources for the experimental exact-SM110 GQA decode backend.
"flashinfer.experimental.sm110_gqa_decode" = ["csrc/**"]
# Raw CUDA sources for the SM90 push-style FP8 mega-MoE drop, JIT-compiled at
# runtime; registered so non-editable installs ship them (PR #4069).
"flashinfer.moe_ep.kernel_src.sm90.push_style_megamoe" = [
  "*.md",
  "src/a2a/*.cu",
  "src/a2a/*.cuh",
  "src/fp8_gemm/*.cu",
  "src/fp8_gemm/*.cuh",
]
# Generated SM100 BF16 rank-major mega-MoE CUDA bundle. The runtime validates
# the manifest checksum before compiling the source on demand.
"flashinfer.moe_ep.kernel_src.blackwell_bf16_rank_major" = [
  "*.md",
  "src/*.cu",
  "src/*.json",
]
"flashinfer.data" = ["csrc/**", "include/**"]
"flashinfer.data.cutlass" = ["include/**", "tools/util/include/**"]
"flashinfer.data.spdlog" = ["include/**"]
"flashinfer.data.cccl" = ["cub/cub/**", "libcudacxx/include/**", "thrust/thrust/**"]
# EP backend shared libraries built in-tree from 3rdparty/nixl by default
# during `pip install` (opt out with BUILD_NIXL_EP=0). The .so files live under
# flashinfer/moe_ep/backends/split/comm/{nccl,nixl}_ep/_libs/ (gitignored;
# populated by build_backend._build_nvep_if_enabled).
"flashinfer.moe_ep.backends.split.comm.nixl_ep" = ["_libs/**", "_vendored/**"]
"flashinfer.moe_ep.backends.split.comm.nccl_ep" = ["_libs/*.so*"]

[tool.mypy]
files = ["flashinfer"]
# Keep this literal in sync with .python-version.
python_version = "3.10"
ignore_missing_imports = true
show_column_numbers = true
show_error_context = true
follow_imports = "skip"
ignore_errors = false
strict_optional = false
exclude = [
  "flashinfer-cubin/",
  "flashinfer-jit-cache/",
  # CUTLASS DSL kernel bodies are compiler-lowered rather than ordinary Python.
  "flashinfer/attention/prims_ts/kernels/",
  "3rdparty/",
  "flashinfer/data/cutlass/",
  "flashinfer/cute_dsl/attention/fmha/fmha.py",
  "flashinfer/cute_dsl/attention/fmha/fmha_blockscaled.py",
  "flashinfer/cute_dsl/attention/fmha/helpers/",
  "flashinfer/prims_ts/",
  "flashinfer/moe_ep/kernel_src/cutedsl_megamoe/src/",
  "flashinfer/moe_ep/kernel_src/sm90/pull_style_cutedsl_megakernel/src/",
  "flashinfer/moe_ep/kernel_src/sm120/swapab_cutedsl_megakernel/src/",
  "flashinfer/gemm/kernels/dense_blockscaled_gemm_sm107.py",
  "flashinfer/fused_moe/cute_dsl/rubin/blockscaled_contiguous_gather_grouped_gemm_swiglu_fusion.py",
  "flashinfer/fused_moe/cute_dsl/rubin/blockscaled_contiguous_grouped_gemm_finalize_fusion.py",
  # DKG filtered top-k kernels, vendored verbatim from dynamic-kernel-generator
  # (see the provenance header in each file). Kept byte-faithful so upstream
  # fixes can be re-applied by re-vendoring. The two reported errors are typing
  # artifacts, not defects: `_get_num_sms` memoises via a function attribute,
  # and `compiled_filter_topk_dict` is an untyped dict holding two distinct key
  # shapes that cannot collide (each leads with a different wrapper-mode string
  # and they differ in length).
  "flashinfer/topk_varlen/kernels/filtered_topk_util.py",
  "flashinfer/topk_varlen/kernels/filtered_topk_decode.py",
  # Self-sampling GVR V2: verbatim TRT-LLM drops (see the ruff exclude note).
  "flashinfer/topk_varlen/kernels/gvr2_topk_decode.py",
  "flashinfer/topk_varlen/kernels/gvr2_topk_host.py",
  "build/",
]

# CuTe DSL FMHA kernels from TRTLLM VisualGen; kept as-is for clean re-sync.
# cutedsl_megamoe/src kernels are maintained upstream and re-dropped
# periodically; excluded so lint churn doesn't block syncs. The shim/ layer
# next to it is ours and stays linted.
# The GVR top-k kernels are CuTe-DSL device code ported verbatim from
# TensorRT-LLM. Exclude them from ruff (both lint and format) so the port stays
# byte-faithful to upstream and future syncs stay clean -- reflowing or
# "simplifying" DSL code can also change traced semantics (e.g. SIM109's
# `x in (a, b)` yields a Python bool, not the DSL Boolean predicate that
# `x == a or x == b` produces).
[tool.ruff]
# Ruff infers the target from project.requires-python.
extend-exclude = [
  "flashinfer/cute_dsl/attention/fmha/fmha.py",
  "flashinfer/cute_dsl/attention/fmha/fmha_blockscaled.py",
  "flashinfer/cute_dsl/attention/fmha/helpers",
  "flashinfer/moe_ep/kernel_src/cutedsl_megamoe/src",
  "flashinfer/topk_varlen/kernels/gvr_topk_decode.py",
  "flashinfer/topk_varlen/kernels/gvr_topk_decode_lb.py",
  "flashinfer/topk_varlen/kernels/gvr2_topk_decode.py",
  "flashinfer/topk_varlen/kernels/gvr2_topk_host.py",
  "flashinfer/topk_varlen/kernels/block_scan.py",
  "flashinfer/moe_ep/kernel_src/sm90/pull_style_cutedsl_megakernel/src",
  "flashinfer/moe_ep/kernel_src/sm120/swapab_cutedsl_megakernel/src",
  "flashinfer/gemm/kernels/dense_blockscaled_gemm_sm107.py",
  "flashinfer/fused_moe/cute_dsl/rubin/blockscaled_contiguous_gather_grouped_gemm_swiglu_fusion.py",
  "flashinfer/fused_moe/cute_dsl/rubin/blockscaled_contiguous_grouped_gemm_finalize_fusion.py",
]
force-exclude = true

[tool.ruff.lint]
select = [
  # pycodestyle
  "E",
  # Pyflakes
  "F",
  # pyupgrade
  # "UP",
  # flake8-bugbear
  "B",
  # flake8-simplify
  "SIM",
  # isort
  # "I",
]
ignore = [
  # Module level import not at top of file
  "E402",
  # star imports
  "F405",
  "F403",
  # ambiguous name
  "E741",
  # line too long
  "E501",
  # key in dict.keys()
  "SIM118",
  # memory leaks
  "B019",
  # No such file or directory
  "E902",
  # nested `if` statements
  "SIM102",
  # `if`-`else`-block
  "SIM108",
  # assign `lambda` expressions
  "E731",
  # Loop control variable overrides iterable it iterates
  "B020",
  # Return te negated condition directly
  "SIM103",
  # Function definition does not bind loop variable
  "B023",
]
[tool.ruff.lint.per-file-ignores]
"__init__.py" = ["F401"]