[Build] Migrate vendored DeepGEMM from pybind to TORCH_LIBRARY (abi3) (#48962)
Signed-off-by: Chris Leonard <chleonar@redhat.com> Co-authored-by: Shengqi Chen <harry-chen@outlook.com>
This commit is contained in:
co-authored by
Shengqi Chen
parent
cbe1f9740d
commit
cc73cca9c8
+34
-66
@@ -1,85 +1,53 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""Build DeepGEMM's `_C` pybind11 extension for <TARGET_PY>.
|
||||
"""Build DeepGEMM's TORCH_LIBRARY extension and copy vendored artifacts.
|
||||
|
||||
Driven from cmake/external_projects/deepgemm.cmake. The driver runs against
|
||||
the build interpreter's torch; <TARGET_PY> is only consulted for INCLUDEPY
|
||||
and SOABI, so target venvs don't need torch installed.
|
||||
DeepGEMM now registers ops via TORCH_LIBRARY into ``deep_gemm._C_extension``
|
||||
(abi3) and exposes the legacy API through ``deep_gemm/_C.py``. This driver
|
||||
delegates to DeepGEMM's ``setup.py build_ext --inplace`` and copies the shim
|
||||
plus extension into the cmake output directory.
|
||||
|
||||
Usage: python build_deepgemm_C.py <DEEPGEMM_SRC_DIR> <OUTPUT_DIR> <TARGET_PY>
|
||||
Usage: python build_deepgemm_C.py <DEEPGEMM_SRC_DIR> <OUTPUT_DIR>
|
||||
"""
|
||||
|
||||
import json
|
||||
import os
|
||||
import shutil
|
||||
import subprocess
|
||||
import sys
|
||||
from pathlib import Path
|
||||
|
||||
import torch
|
||||
from torch.utils import cpp_extension
|
||||
|
||||
if len(sys.argv) != 4:
|
||||
sys.exit(f"usage: {sys.argv[0]} <SRC> <OUT> <TARGET_PY>")
|
||||
if len(sys.argv) != 3:
|
||||
sys.exit(f"usage: {sys.argv[0]} <SRC> <OUT>")
|
||||
|
||||
src = Path(sys.argv[1]).resolve()
|
||||
out = Path(sys.argv[2]).resolve()
|
||||
target_py = sys.argv[3]
|
||||
_pkg = src / "deep_gemm"
|
||||
out.mkdir(parents=True, exist_ok=True)
|
||||
|
||||
info = json.loads(
|
||||
subprocess.check_output(
|
||||
[
|
||||
target_py,
|
||||
"-c",
|
||||
"import sysconfig, json; "
|
||||
"print(json.dumps({k: sysconfig.get_config_var(k) "
|
||||
"for k in ('EXT_SUFFIX', 'INCLUDEPY')}))",
|
||||
]
|
||||
).decode()
|
||||
if not (_pkg / "_C.py").is_file():
|
||||
sys.exit(
|
||||
f"DeepGEMM source at {src} is missing deep_gemm/_C.py; "
|
||||
"expected TORCH_LIBRARY migration layout"
|
||||
)
|
||||
|
||||
# Avoid DeepGEMM's clean-git assertion when vendoring a local dirty tree.
|
||||
env = os.environ.copy()
|
||||
env.pop("DG_SKIP_CUDA_BUILD", None)
|
||||
|
||||
print(f"[build_deepgemm_C] building in {src} with {sys.executable}", flush=True)
|
||||
subprocess.check_call(
|
||||
[sys.executable, "setup.py", "build_ext", "--inplace"],
|
||||
cwd=src,
|
||||
env=env,
|
||||
)
|
||||
|
||||
cuda_home = cpp_extension.CUDA_HOME
|
||||
if cuda_home is None:
|
||||
sys.exit("CUDA_HOME not found; cannot build DeepGEMM _C")
|
||||
# CCCL lives outside the standard CUDAToolkit search (mirrors DeepGEMM's setup.py).
|
||||
includes = [
|
||||
info["INCLUDEPY"],
|
||||
f"{cuda_home}/include",
|
||||
f"{cuda_home}/include/cccl",
|
||||
str(src / "csrc"),
|
||||
str(src / "deep_gemm/include"),
|
||||
str(src / "third-party/cutlass/include"),
|
||||
str(src / "third-party/cutlass/tools/util/include"),
|
||||
str(src / "third-party/deep_jit/include"),
|
||||
*cpp_extension.include_paths(device_type="cuda"),
|
||||
]
|
||||
shim = _pkg / "_C.py"
|
||||
shutil.copy2(shim, out / shim.name)
|
||||
|
||||
cmd = [
|
||||
os.environ.get("CXX", "g++"),
|
||||
"-shared",
|
||||
"-fPIC",
|
||||
"-std=c++20",
|
||||
"-O3",
|
||||
"-g0",
|
||||
"-Wno-psabi",
|
||||
"-Wno-deprecated-declarations",
|
||||
"-DTORCH_API_INCLUDE_EXTENSION_H",
|
||||
"-DTORCH_EXTENSION_NAME=_C",
|
||||
f"-D_GLIBCXX_USE_CXX11_ABI={int(torch.compiled_with_cxx11_abi())}",
|
||||
*(f"-I{p}" for p in includes),
|
||||
str(src / "csrc/python_api.cpp"),
|
||||
*(f"-L{p}" for p in cpp_extension.library_paths(device_type="cuda")),
|
||||
f"-L{cuda_home}/lib64",
|
||||
"-ltorch",
|
||||
"-ltorch_python",
|
||||
"-ltorch_cpu",
|
||||
"-ltorch_cuda",
|
||||
"-lc10",
|
||||
"-lc10_cuda",
|
||||
"-lcudart",
|
||||
"-lnvrtc",
|
||||
"-o",
|
||||
str(out / f"_C{info['EXT_SUFFIX']}"),
|
||||
]
|
||||
print("[build_deepgemm_C] " + " ".join(cmd), flush=True)
|
||||
subprocess.check_call(cmd)
|
||||
so = _pkg / "_C_extension.abi3.so"
|
||||
if not so.is_file():
|
||||
sys.exit(f"DeepGEMM build did not produce {so}")
|
||||
for stale in out.glob("_C_extension*.so"):
|
||||
stale.unlink()
|
||||
shutil.copy2(so, out / so.name)
|
||||
print(f"[build_deepgemm_C] installed {so.name} -> {out}", flush=True)
|
||||
|
||||
@@ -1,41 +1,45 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
|
||||
"""Assert the installed vLLM has a `_C.cpython-X.Y-*.so` for every CPython
|
||||
covered by `requires-python`. Fails closed if a Python's `.so` is missing
|
||||
from the wheel — i.e. the regression that surfaced in #41476/#41512.
|
||||
"""Assert the vendored DeepGEMM package has the TORCH_LIBRARY binding layout.
|
||||
|
||||
Run from a CI test job after vLLM is installed, e.g. the H100 deepgemm
|
||||
kernel tests in .buildkite/test_areas/kernels.yaml.
|
||||
Expects ``deep_gemm/_C.py`` plus a single ``_C_extension*.so`` (abi3) under
|
||||
``vllm.third_party.deep_gemm``. Run after vLLM is installed, e.g. the H100
|
||||
deepgemm kernel tests in .buildkite/test_areas/kernels.yaml.
|
||||
"""
|
||||
|
||||
import importlib.util
|
||||
import os
|
||||
import sys
|
||||
from pathlib import Path
|
||||
|
||||
import regex as re
|
||||
import tomllib
|
||||
|
||||
SO_RE = re.compile(r"^_C\.cpython-(\d)(\d+)-")
|
||||
def main() -> int:
|
||||
spec = importlib.util.find_spec("vllm.third_party.deep_gemm")
|
||||
if spec is None or spec.origin is None:
|
||||
print(
|
||||
"vllm.third_party.deep_gemm not importable; is vllm installed?",
|
||||
file=sys.stderr,
|
||||
)
|
||||
return 1
|
||||
pkg_dir = Path(spec.origin).parent
|
||||
|
||||
shim = pkg_dir / "_C.py"
|
||||
so_files = sorted(pkg_dir.glob("_C_extension*.so"))
|
||||
missing = []
|
||||
if not shim.is_file():
|
||||
missing.append("_C.py")
|
||||
if len(so_files) != 1:
|
||||
missing.append("exactly one _C_extension*.so")
|
||||
|
||||
print(
|
||||
f"deepgemm vendored binding: shim={shim.is_file()}, "
|
||||
f"extensions={[p.name for p in so_files]}"
|
||||
)
|
||||
if missing:
|
||||
print(f"missing: {missing}", file=sys.stderr)
|
||||
return 1
|
||||
return 0
|
||||
|
||||
|
||||
def required_pythons() -> list[str]:
|
||||
pyproject = Path(__file__).resolve().parent.parent / "pyproject.toml"
|
||||
spec = tomllib.loads(pyproject.read_text())["project"]["requires-python"]
|
||||
m = re.match(r">=3\.(\d+),<3\.(\d+)", spec)
|
||||
if not m:
|
||||
sys.exit(f"unexpected requires-python format: {spec!r}")
|
||||
return [f"3.{v}" for v in range(int(m[1]), int(m[2]))]
|
||||
|
||||
|
||||
spec = importlib.util.find_spec("vllm.third_party.deep_gemm")
|
||||
if spec is None or spec.origin is None:
|
||||
sys.exit("vllm.third_party.deep_gemm not importable; is vllm installed?")
|
||||
pkg_dir = Path(spec.origin).parent
|
||||
|
||||
found = {f"{m[1]}.{m[2]}" for f in os.listdir(pkg_dir) if (m := SO_RE.match(f))}
|
||||
required = required_pythons()
|
||||
missing = [v for v in required if v not in found]
|
||||
print(f"deepgemm _C: found {sorted(found)}, required {required}, missing {missing}")
|
||||
sys.exit(1 if missing else 0)
|
||||
if __name__ == "__main__":
|
||||
raise SystemExit(main())
|
||||
|
||||
@@ -10,7 +10,7 @@ DEEPGEMM_GIT_REPO="https://github.com/vllm-project/DeepGEMM.git"
|
||||
# NOTE: This targets the vLLM fork's dev branch, which carries the sm120
|
||||
# and sm90 paged-MQA ports plus the SwiGLU alpha/beta and SiTU Mega MoE
|
||||
# activations, plus the CUDA 12.x layout header fix from vllm-project/DeepGEMM#12.
|
||||
DEEPGEMM_GIT_REF="e1f418c2a4f20818221f6b0e578b4c2f634d4c3f"
|
||||
DEEPGEMM_GIT_REF="1e1842a833699298f7afc02eefb2ed168fff6938"
|
||||
WHEEL_DIR=""
|
||||
|
||||
# Parse command line arguments
|
||||
@@ -93,6 +93,8 @@ pushd "$INSTALL_DIR/deepgemm"
|
||||
|
||||
# Checkout the specific reference
|
||||
git checkout "$DEEPGEMM_GIT_REF"
|
||||
git submodule sync --recursive
|
||||
git submodule update --init --recursive
|
||||
|
||||
# Clean previous build artifacts
|
||||
# (Based on https://github.com/deepseek-ai/DeepGEMM/blob/main/install.sh)
|
||||
|
||||
Reference in New Issue
Block a user