mirror of
https://github.com/deepseek-ai/DeepInfer.git
synced 2026-10-31 00:36:20 +08:00
b4f7a42652
`h20_ops_ab.py` assumed the TileLang backend is importable, so on a vendor image it could not start at all -- which is exactly where the Triton backend needs checking. Two modes fix that: * `--dump DIR` saves the selected backend's output per case. * `--compare DIR` checks against dumps written elsewhere. Together they answer "do triton 3.5.1 and 3.7.1 agree" without either host having TileLang. Case builders that genuinely need a package the host lacks now report a skip instead of aborting the run, so partial coverage is visible rather than fatal; the import of the sampler's C++-backed inputs moved into the one builder that uses it for the same reason. This is what showed the 19 ported kernels produce element-wise identical output across triton 3.5.1, 3.6.0 and 3.7.1 -- 115 cases, no differences.
92 lines
2.8 KiB
TOML
92 lines
2.8 KiB
TOML
[build-system]
|
|
requires = [
|
|
"ninja",
|
|
"setuptools >= 77.0.3",
|
|
"torch >= 2.7",
|
|
]
|
|
build-backend = "setuptools.build_meta"
|
|
|
|
[project]
|
|
name = "deepinfer"
|
|
version = "0.1.0"
|
|
description = "DeepInfer: High-performance, large-scale inference framework for DeepSeek models"
|
|
readme = "README.md"
|
|
license = "Apache-2.0"
|
|
license-files = ["LICENSE"]
|
|
requires-python = ">=3.12"
|
|
authors = [{ name = "DeepSeek-AI" }]
|
|
keywords = ["llm", "inference", "deepseek", "cuda", "ascend"]
|
|
classifiers = [
|
|
"Development Status :: 4 - Beta",
|
|
"Intended Audience :: Developers",
|
|
"Programming Language :: Python :: 3.12",
|
|
"Topic :: Scientific/Engineering :: Artificial Intelligence",
|
|
]
|
|
dynamic = ["dependencies"]
|
|
urls = { Repository = "https://github.com/deepseek-ai/deepinfer", Issues = "https://github.com/deepseek-ai/deepinfer/issues" }
|
|
|
|
[tool.pytest.ini_options]
|
|
# Use importlib import mode so that collecting test files inside the deepinfer
|
|
# package (e.g. deepinfer/model_executor/layers/tilelang_ops/*.py) does NOT
|
|
# trigger the package __init__.py chain (which imports heavy deps like
|
|
# deep_gemm).
|
|
addopts = ["--import-mode=importlib"]
|
|
testpaths = ["tests"]
|
|
markers = [
|
|
"ci: core tests that run on every CI pipeline",
|
|
"ci_npu: core tests that run on the NPU/Ascend CI pipeline",
|
|
"benchmark: performance tests selected by benchmark jobs",
|
|
"long_context: long-input/output decode PPL regression",
|
|
"gpu: unit tests requiring a working GPU",
|
|
]
|
|
|
|
[tool.ruff.lint]
|
|
select = [
|
|
# pycodestyle
|
|
"E",
|
|
# Pyflakes
|
|
"F",
|
|
# pyupgrade
|
|
"UP",
|
|
# flake8-bugbear
|
|
"B",
|
|
# flake8-simplify
|
|
"SIM",
|
|
# isort
|
|
"I",
|
|
# type hint
|
|
"ANN",
|
|
# disallow specific packages
|
|
"TID251",
|
|
]
|
|
ignore = [
|
|
# missing return type for private function; dynamically typed Any
|
|
"ANN202", "ANN401",
|
|
# star imports
|
|
"F405", "F403",
|
|
# lambda expression assignment
|
|
"E731",
|
|
# f-string without any placeholders
|
|
"F541",
|
|
# line too long, handled by black formatting
|
|
"E501",
|
|
# module level import not at top of file
|
|
"E402",
|
|
# function definition does not bind loop variable
|
|
"B023",
|
|
]
|
|
|
|
[tool.ruff.lint.per-file-ignores]
|
|
# Triton kernel bodies: parameters are pointers and scalars that Triton itself types, and the
|
|
# kernels are launched through `kernel[grid](...)` rather than called directly, so the usual
|
|
# signature annotations would be either wrong or rejected by the JIT.
|
|
"deepinfer/model_executor/layers/tilelang_ops/triton/*.py" = ["ANN001", "ANN201", "ANN202"]
|
|
# Same reason, for the environment probe: its kernels exist to be compiled by whatever Triton the
|
|
# image ships, and annotating their pointers would change what the JIT sees.
|
|
"scripts/flagtree_triton_probe.py" = ["ANN001", "ANN201", "ANN202"]
|
|
|
|
[tool.ruff]
|
|
line-length = 140
|
|
include = ["./**/*.py"]
|
|
extend-exclude = ["vendor"]
|