pyproject.toml

[build-system]
requires = [
  "setuptools>=61.0",
  "setuptools-rust>=1.11",
  "setuptools-scm>=8.0",
  "torch==2.14.1",
  "wheel",
]
build-backend = "setuptools.build_meta"

[project]
name = "sglang"
dynamic = ["version"]
description = "QSA HiSparse: SGLang with request-scoped CPU KV offload for Qwen Sparse Attention."
readme = "README.md"
requires-python = ">=3.10"
license = { file = "LICENSE" }
classifiers = [
  "Programming Language :: Python :: 3",
  "License :: OSI Approved :: Apache Software License",
]

# Please keep dependency lists in this file sorted alphabetically by package name.
dependencies = [
  "aiohttp",
  "anthropic>=0.20.0",
  "apache-tvm-ffi==0.1.11",
  "av==16.1.0 ; sys_platform == 'linux' and (platform_machine == 'aarch64' or platform_machine == 'arm64' or platform_machine == 'armv7l')",
  "blobfile==3.0.0",
  "build",
  "compressed-tensors==0.18.0",
  "cuda-python>=13.0",
  # Transitive dep of flashinfer-python, pinned because 1.6.0rc6 shipped no cp310 linux x86_64 wheel.
  "cuda-tile==1.6.0rc5",
  "datasets",
  "decord2 ; sys_platform == 'linux' and (platform_machine == 'aarch64' or platform_machine == 'arm64' or platform_machine == 'armv7l')",
  "distro",
  "easydict",  # Required by remote model code (e.g. DeepSeek-OCR) loaded via trust_remote_code; validated by transformers 5.4+ check_imports
  "einops",
  "fastapi",
  "flash-attn-4>=4.0.0b18",
  "flashinfer_python[cu13]==0.7.0.post1", # keep it aligned with jit-cache version in Dockerfile
  "gguf",
  "humming-kernels[cu13]==0.1.12",
  "interegular",
  "IPython",
  "kernels>=0.14.1,<0.15",
  "llguidance>=1.7.6,<2.0.0",
  "mistral_common>=1.11.5",
  "modelscope",
  "msgspec",
  "ninja",
  "numba==0.65.1",
  "numpy",
  "nvidia-cutlass-dsl[cu13]==4.8.0",
  "nvidia-mathdx==25.6.0",
  "nvidia-ml-py",
  "nvshmem4py-cu13",
  "openai==2.6.1",
  "openai-harmony==0.0.4",
  "orjson",
  "outlines==0.1.11",
  "packaging",
  "partial_json_parser",
  "pillow",
  "prometheus-client>=0.20.0",
  "psutil",
  "py-spy",
  "pybase64",
  "pydantic",
  "python-multipart",
  "pyzmq>=25.1.2",
  "quack-kernels==0.6.5",
  "regex",
  "requests",
  "scipy",
  "sentencepiece==0.2.1",  # 0.2.2 rejects null pieces in InternVL tokenizers.
  "setproctitle",
  "sgl-deep-ep==0.1.2.post1",
  "sgl-deep-gemm==0.2.1.post1",
  "sglang-kernel==0.4.9",
  "smg-grpc-servicer>=0.9.0",
  "soundfile==0.13.1",
  "tiktoken",
  "tilelang==0.1.12",
  "timm==1.0.16",
  "tokenizers==0.23.2",
  "tokenspeed_mla==0.1.8",
  "tomli ; python_version < '3.11'",
  "torch==2.14.1",
  "torch_memory_saver>=0.0.9.post1",
  "torchaudio==2.11.0",
  "torchcodec==0.17.0 ; sys_platform != 'linux' or (sys_platform == 'linux' and platform_machine != 'aarch64' and platform_machine != 'arm64' and platform_machine != 'armv7l')", # Not available on Linux ARM.
  "torchvision",
  "tqdm",
  "transformers==5.17.0",
  "uvicorn",
  "uvloop",
  "watchfiles",
  "xgrammar==0.2.7",
  "xxhash",
  "zstandard",
]

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

[project.optional-dependencies]
checkpoint-engine = ["checkpoint-engine==0.1.2"]
runai = ["runai-model-streamer[s3,gcs,azure]>=0.15.7"]
diffusion = [
  "addict==2.4.0",
  "av==16.1.0",
  "cache-dit==1.5.1",
  "cloudpickle==3.1.2",
  "diffusers==0.37.0",
  "imageio==2.36.0",
  "imageio-ffmpeg==0.5.1",
  "moviepy>=2.0.0",
  "msgpack",
  "nvidia-modelopt",
  "opencv-python-headless==4.10.0.84",
  "PyYAML==6.0.1",
  "remote-pdb==2.1.0",
  "runai_model_streamer>=0.15.7",
  "scikit-image==0.25.2",
  "st_attn==0.0.7 ; platform_machine != 'aarch64' and platform_machine != 'arm64'",
  "trimesh>=4.0.0",
  "vsa==0.0.4 ; platform_machine != 'aarch64' and platform_machine != 'arm64'",
  "websockets",
  "xatlas",
]

diffusion-qvg = [
  # quant-videogen 0.1.0 pins an incompatible Torch version, so it is
  # installed separately with --no-deps; this extra provides its missing
  # import-time dependencies.
  "loguru>=0.7",
  "termcolor>=2.3",
]

ray = [
  "ray[default]>=2.55.1",
]

tracing = [
  "opentelemetry-api",
  "opentelemetry-exporter-otlp",
  "opentelemetry-exporter-otlp-proto-grpc",
  "opentelemetry-sdk",
]

http2 = [
  "granian>=2.6.0",
]

fastokens = [
  "fastokens>=0.1.1,<0.2.0",
]

test = [
  "accelerate",
  "addict",
  "auto-round>=0.13.1",
  "av==16.1.0",
  "bitsandbytes",
  "pymupdf",
  "diff-cover",
  "expecttest",
  "granian>=2.6.0",
  "jsonlines",
  "lm-eval[api]>=0.4.9.2",
  "matplotlib",
  # 4.9.3 is what sgl-eval's math_verify grader needs (latex2sympy2_extended
  # raises ImportError on 4.7.x); pinned here so other test deps cannot pull
  # it down.
  "antlr4-python3-runtime==4.9.3",
  "pandas",
  "parameterized",
  "peft>=0.18.0",
  "polars",
  "pytest",
  "pytest-cov",
  "sentence_transformers",
  # Pinned: a bump moves scoring for every eval test at once, so re-baseline
  # MODEL_SCORE_THRESHOLDS in test/registered/eval/test_text_models_gsm8k_eval.py
  # and the mmlu thresholds of run_eval's other callers before changing it.
  "sgl-eval==0.1.2",
  "sglang[fastokens]",
  "tabulate",
]

dev = ["sglang[test]"]

all = [
  "sglang[diffusion]",
  "sglang[http2]",
  "sglang[tracing]",
]

[tool.uv.extra-build-dependencies]
st-attn = ["setuptools", "torch"]
vsa = ["setuptools", "torch"]

[project.urls]
"Homepage" = "https://github.com/mochgolf/qsa-hisparse"
"Bug Tracker" = "https://github.com/mochgolf/qsa-hisparse/issues"
"Upstream" = "https://github.com/sgl-project/sglang"
"Documentation" = "https://docs.sglang.io"
"Cookbook" = "https://docs.sglang.io/cookbook"

[project.scripts]
sglang = "sglang.cli.main:main"
killall_sglang = "sglang.cli.killall:main"

[tool.setuptools.package-data]
"sglang" = [
  "srt/**/*",
  "kernels/**/*",
  "multimodal_gen/apps/realtime_webui/**/*"
]

[tool.setuptools.exclude-package-data]
"sglang" = [
  "kernels/aot/*",
  "kernels/aot/**/*",
  "srt/mem_cache/rust_tree_core/mem_cache_inspection*.so",
]
# .claude/skills is a directory symlink; setuptools build_py cannot copy it.
"sglang.multimodal_gen" = [".claude/*"]

[tool.setuptools.packages.find]
exclude = [
  "assets*",
  "benchmark*",
  "docs*",
  "dist*",
  "playground*",
  "scripts*",
  "sglang.kernels.aot*",
  "tests*",
]

[tool.wheel]
exclude = [
  "assets*",
  "benchmark*",
  "docs*",
  "dist*",
  "playground*",
  "scripts*",
  "sglang/kernels/aot*",
  "tests*",
]

[tool.setuptools_scm]
root = ".."
version_file = "sglang/_version.py"
git_describe_command = ["python3", "scripts/release/get_version_tag.py"]
# Allow editable installs even when .git metadata is not available.
fallback_version = "0.0.0.dev0"

# Rust extension modules are auto-discovered by setup.py from the Cargo
# workspace in ../rust and its declared extension manifests
# ([package.metadata.sglang] python-module in each crate).
# This CUDA pyproject builds all of them; platform variants restrict the set
# via [tool.sglang] rust-extensions (see pyproject_other.toml).

[tool.kernels.dependencies]
"kernels-community/sgl-flash-attn3" = 1