pyproject.toml

[tool.uv.sources]
ltx-core = { workspace = true }
ltx-pipelines = { workspace = true }
ltx-kernels = { path = "packages/ltx-kernels", editable = true }

[tool.uv]
no-build-isolation-package = ["ltx-kernels"]
# torch 2.13+cu132 pins nvidia-cudnn-cu13==9.20.0.48, which omits
# libcudnn_engines_tensor_ir; the system/apt cuDNN then resolves that soname from
# /usr/lib, causing CUDNN_STATUS_SUBLIBRARY_VERSION_MISMATCH. A wheel *older* than the
# cuDNN torch was built against additionally aborts the process -- not an exception --
# with `Cannot load symbol cublasLtGetVersion` the first time torch selects its cuDNN
# SDPA backend. 9.24.0.43 ships tensor_ir and matches torch 2.13 / cuDNN 9.24;
# 9.25.0.15 is yanked. Linux-only: the wheel is CUDA and has no darwin build.
override-dependencies = ["nvidia-cudnn-cu13==9.24.0.43; sys_platform == 'linux'"]

[tool.uv.workspace]
members = ["packages/*"]
# ltx-kernels compiles CUDA extensions; keep it out of the workspace so a plain
# `uv sync` never forces a CUDA toolchain. Opt in via the `kernels` group below.
exclude = ["packages/ltx-kernels"]

[[tool.uv.index]]
name = "pypi"
url = "https://pypi.org/simple"

[dependency-groups]
dev = [
    "google-cloud-storage>=2.0",
    "matplotlib>=3.7",
    "pre-commit>=4.3.0",
    "ruff>=0.14.3",
    "pytest~=9.0",
    "pytest-xdist~=3.8",
    # Match torch's cu132 index (see packages/ltx-core pyproject): kernel builds must use
    # the venv nvcc, not a mismatched system toolkit (CI runners often ship CUDA 12.8).
    # `cccl` provides <nv/target> (pulled by cuda_fp16.h); `nvcc` alone is not enough.
    "cuda-toolkit[nvcc,cccl]==13.2.*; sys_platform == 'linux'",
    "nvidia-cuda-nvcc==13.2.*; sys_platform == 'linux'",
    "nvidia-cuda-cccl==13.2.*; sys_platform == 'linux'",
]
# Opt-in compiled CUDA kernels (blockwise FP8/FP6 GEMM + multi-GPU all2all), built
# editable from packages/ltx-kernels without build isolation (torch must already be
# installed), CUDA host required. Install with `uv sync --group kernels`.
kernels = ["ltx-kernels"]

[tool.ruff]
target-version = "py311"
line-length = 120
exclude = [
    # Vendored kernel sources, excluded from our lint/format rules.
    "packages/ltx-kernels/src/ltx_kernels/blockwise/functional.py",
    "packages/ltx-kernels/src/ltx_kernels/blockwise/linear.py",
    "packages/ltx-kernels/src/ltx_kernels/blockwise/triton_ops.py",
]

[tool.ruff.lint]
select = [
    "E", # pycodestyle
    "F", # pyflakes
    "W", # pycodestyle (warnings)
    "I", # isort
    "N", # pep8-naming
    "ANN", # flake8-annotations
    "B", # flake8-bugbear
    "A", # flake8-builtins
    "COM", # flake8-commas
    "C4", # flake8-comprehensions
    "DTZ", # flake8-datetimez
    "EXE", # flake8-executable
    "PIE", # flake8-pie
    "T20", # flake8-print
    "PT", # flake8-pytest
    "SIM", # flake8-simplify
    "ARG", # flake8-unused-arguments
    "PTH", # flake8--use-pathlib
    "ERA", # flake8-eradicate
    "RUF", # ruff specific rules
    "PL", # pylint
]
ignore = [
    "ANN002", # Missing type annotation for *args
    "ANN003", # Missing type annotation for **kwargs
    "ANN204", # Missing type annotation for special method
    "COM812", # Missing trailing comma
    "PTH123", # `open()` should be replaced by `Path.open()`
    "PLR2004", # Magic value used in comparison, consider replacing with a constant variable
    "N812", # Lowercase  imported as non-lowercase
]

[tool.ruff.lint.per-file-ignores]
# CI gates: stand-alone CLIs whose whole job is printing a report / GitHub annotations.
".github/actions/golden-pairing-gate/check_golden_pairing.py" = ["T201"]
# Emits `export …` lines for eval by kernels-build.
".github/actions/kernels-build/export_cuda_home.py" = ["T201"]
# setup.py must set CUDA_HOME before importing torch.utils.cpp_extension.
"packages/ltx-kernels/setup.py" = ["E402"]
# Tests for the CuTe DSL kernels. ltx-kernels is an optional, CUDA-host-only install,
# so importing it at module level would turn "skipped" into a collection error wherever
# it is absent -- which is most of CI.
"packages/ltx-kernels/tests/test_vae_*.py" = ["PLC0415"]
"packages/ltx-core/tests/**/test_dsl_kernels.py" = ["PLC0415"]
# NVFP4 tests ``importorskip("ltx_kernels")`` before importing convert/linear/fuse
# (which import the extension at module top level).
"packages/ltx-core/tests/ltx_core/quantization/nvfp4/test_*.py" = ["E402"]
# CuTe DSL kernels. Device-side code written in CUTLASS's own idiom: tensor-algebra
# names are conventionally uppercase (M, T, H, W, NH, TILER_N128), a `@cute.kernel`
# entry point takes one parameter per compile-time constant, and its arguments are
# DSL handles with no meaningful Python annotation. Formatting and everything else
# is still enforced.
"packages/ltx-kernels/src/ltx_kernels/vae/*.py" = [
    "ANN001", # Missing type annotation for function argument
    "ANN201", # Missing return type annotation for public function
    "ANN202", # Missing return type annotation for private function
    "N803",   # Argument name should be lowercase
    "N806",   # Variable in function should be lowercase
    "PLR0402", # `import cutlass.cute as cute` -- the conventional CuTe spelling
    "PLR0912", # Too many branches
    "PLR0913", # Too many arguments
    "PLR0915", # Too many statements
    "PLR1730", # `if a > b: a = b` -- a device-value clamp, not a Python min()
]

[tool.ruff.lint.pylint]
max-args = 10

[tool.ruff.lint.isort]
known-first-party = ["ltx_core", "ltx_pipelines", "ltx_trainer"]