Pointcept

Perceive the world with sparse points, a codebase for point cloud perception research. Latest works: Concerto (NeurIPS'25), Sonata (CVPR'25 Highlight), PTv3 (CVPR'24 Oral)

PTv3, Sonata, Concerto 등을 통합 지원하는 point cloud perception 연구 프레임워크

Implementations

It is also an official implementation of the following paper:

🚀 Utonia - Toward One Encoder for All Point Clouds
Concerto - Joint 2D-3D Self-Supervised Learning Emerges Spatial Representations
Sonata - Self-Supervised Learning of Reliable Point Representations
Point Transformer V3 - Simpler, Faster, Stronger
OA-CNNs - Omni-Adaptive Sparse CNNs for 3D Semantic Segmentation
Towards Large-scale 3D Representation Learning with Multi-dataset Point Prompt Training
Masked Scene Contrast - A Scalable Framework for Unsupervised 3D Representation Learning
Learning Context-aware Classifier for Semantic Segmentation (3D Part)
Point Transformer V2 - Grouped Vector Attention and Partition-based Pooling
Point Transformer

Additionally, Pointcept integrates the following excellent work (contain above):

Backbone: MinkUNet, SpUNet, SPVCNN, OACNNs, PTv1, PTv2, PTv3, StratifiedFormer, OctFormer, Swin3D
Semantic Segmentation: Mix3d, CAC
Instance Segmentation: PointGroup
Pre-training: PointContrast, Contrastive Scene Contexts, Masked Scene Contrast, Point Prompt Training, Sonata, Concerto
Datasets: ScanNet, ScanNet200, ScanNet++, S3DIS, ArkitScene, HM3D, Matterport3D, Structured3D, SemanticKITTI, nuScenes, ModelNet40, Waymo

Requirements

Ubuntu: 18.04 and above.
CUDA: 11.3 and above.
PyTorch: 1.10.0 and above.

Blackwell Architecture 지원

CUDA Toolkit 은 12.8 이상을 사용해야 한다.

CUDA 툴킷 12.8 이하 버전을 사용하여 빌드된 애플리케이션

1. Blackwell Architecture Compatibility — Blackwell Compatibility Guide 13.2 documentation

CUDA 툴킷 버전 2.1부터 12.8까지를 사용하여 빌드된 CUDA 애플리케이션은 커널의 PTX 버전을 포함하도록 빌드된 경우 Blackwell GPU와 호환됩니다. 다음 단계를 따라 애플리케이션 로드 시 PTX를 JIT 컴파일하도록 강제하여 이를 테스트할 수 있습니다.

Download and install the latest driver from https://www.nvidia.com/drivers.
Set the environment variable CUDA_FORCE_PTX_JIT=1
Launch the application.

CUDA_FORCE_PTX_JIT=1로 설정하면 애플리케이션 바이너리에 포함된 GPU 바이너리 코드는 무시됩니다.

대신 각 커널의 PTX 코드가 JIT 컴파일되어 GPU 바이너리 코드가 생성됩니다. PTX 코드가 포함되지 않은 애플리케이션은 실행되지 않습니다. 이는 해당 애플리케이션이 Blackwell 아키텍처와 호환되지 않으므로 호환성을 위해 다시 빌드해야 함을 의미합니다.

반대로 이 환경 변수를 설정한 상태에서 애플리케이션이 정상적으로 작동하면 Blackwell 아키텍처와 호환되는 것입니다.

필수 패키지 설치

sudo apt install libsparsehash-dev gcc-13 g++-13

spconv 설치

Because spconv-cu128 was not found in the package registry and you require spconv-cu128, we can conclude that your requirements are unsatisfiable.

spconv-cu128 wheel 이 PyPI 에 아직 없는 것이 원인입니다. 최신은 spconv-cu126 (2.3.8).

3가지 대응 방법:

옵션	내용	장단점
spconv-cu126 사용	cu126 wheel 을 CUDA 12.8 런타임에서 실행	✅ 즉시 설치. CUDA ABI 호환. ⚠️ Blackwell 네이티브 커널 X → PTX JIT fallback (첫 실행 지연, 약간의 성능 손실)
spconv 소스 빌드	pip install spconv + cumm 수동	✅ 네이티브 sm_120. ❌ 빌드 시간·설정 복잡, cumm 도 같이 빌드 필요
spconv 제외	requirements 에서 삭제	✅ 간단. ❌ SparseUNet/MinkUNet 백본 사용 불가 (PTv3 는 무관)

uv 초기화

내가 conda 계열을 싫어함. 그래서 uv로 설정한다.

사용하는 Python Version 3.10 (상세 버전은 3.10.19)

requirements.cu128.txt

requirements.cu124.txt 파일을 만들고 다음과 같이 추가한다:

# Pointcept - GPU requirements (CUDA 12.8 / Blackwell sm_120)
# Usage:
#   ./uv-install-cu128.sh
# or manually:
#   uv venv --python 3.10 && source .venv/bin/activate
#   uv pip install -r requirements.cu128.txt
#   TORCH_CUDA_ARCH_LIST="12.0" uv pip install ./libs/pointops
#
# System prerequisites:
#   - CUDA 12.8 toolkit + cuDNN 9.x (nvcc must know sm_120)
#     https://developer.nvidia.com/cuda-12-8-0-download-archive
#   - gcc-13 / g++-13
#   - libsparsehash-dev (apt install libsparsehash-dev)
#   - NVIDIA driver >= 570 (Blackwell consumer cards require driver 572+)

--extra-index-url https://download.pytorch.org/whl/cu128

# PyTorch (CUDA 12.8, Blackwell officially supported from 2.7.0)
# NOTE: +cu128 local version identifier 를 명시해야 PyPI 의 CPU 빌드로 대체되지 않음.
torch==2.7.0+cu128
torchvision==0.22.0+cu128
torchaudio==2.7.0+cu128

# PyG extensions (torch 2.7 + cu128 wheels)
--find-links https://data.pyg.org/whl/torch-2.7.0+cu128.html
torch-cluster==1.6.3
torch-scatter==2.1.2
torch-sparse==0.6.18
torch-geometric==2.7.0

# spconv — SparseUNet / MinkUNet 백본용
# NOTE: spconv-cu128 wheel 은 아직 PyPI 에 미공개.
#       cu126 wheel 을 CUDA 12.8 런타임에서 사용 (CUDA major ABI 호환).
#       Blackwell 전용 cubin 이 없어 첫 실행 시 PTX JIT 지연이 발생할 수 있음.
#       PTv3 only 사용자는 이 줄을 주석 처리해도 무방.
spconv-cu126==2.3.8

# Core
ninja
h5py
pyyaml
sharedarray
tensorboard
tensorboardx
wandb
yapf
addict
einops
scipy
plyfile
termcolor
timm
peft
open3d

# CLIP
ftfy
regex
tqdm
clip @ git+https://github.com/openai/CLIP.git

# (Optional) Blackwell 대응 flash-attn — 필요 시 별도 설치 권장
# flash-attn>=2.7.4

설치 스크립트

uv-install-cu128.sh 파일을 만들어서 실행하자:

#!/bin/bash
set -euo pipefail

PYTHON_VERSION="3.10"
VENV_DIR=".venv"
MIN_CUDA_MAJOR=12
MIN_CUDA_MINOR=8

echo "=== Pointcept GPU (CUDA 12.8 / Blackwell sm_120) Installation ==="

# --- [1/6] System prerequisite checks ---
echo "[1/6] Checking system prerequisites..."

if ! command -v uv &> /dev/null; then
    echo "ERROR: 'uv' not found."
    echo "  Install: curl -LsSf https://astral.sh/uv/install.sh | sh"
    exit 1
fi

if ! command -v nvcc &> /dev/null; then
    echo "ERROR: nvcc not found."
    echo "  Blackwell (sm_120) requires CUDA >= ${MIN_CUDA_MAJOR}.${MIN_CUDA_MINOR} toolkit."
    echo "  Install: https://developer.nvidia.com/cuda-12-8-0-download-archive"
    exit 1
fi

NVCC_VER=$(nvcc --version | grep -oP 'release \K[0-9]+\.[0-9]+' | head -1)
NVCC_MAJOR=${NVCC_VER%.*}
NVCC_MINOR=${NVCC_VER#*.}
if [ "$NVCC_MAJOR" -lt "$MIN_CUDA_MAJOR" ] || \
   { [ "$NVCC_MAJOR" -eq "$MIN_CUDA_MAJOR" ] && [ "$NVCC_MINOR" -lt "$MIN_CUDA_MINOR" ]; }; then
    echo "ERROR: nvcc ${NVCC_VER} found, but Blackwell (sm_120) requires >= ${MIN_CUDA_MAJOR}.${MIN_CUDA_MINOR}."
    echo "  Current nvcc: $(which nvcc)"
    echo "  Install CUDA 12.8 toolkit and ensure it is first in PATH."
    echo "    export PATH=/usr/local/cuda-12.8/bin:\$PATH"
    echo "    export LD_LIBRARY_PATH=/usr/local/cuda-12.8/lib64:\$LD_LIBRARY_PATH"
    exit 1
fi
echo "  nvcc ${NVCC_VER} OK"

if command -v nvidia-smi &> /dev/null; then
    GPU_CC=$(nvidia-smi --query-gpu=compute_cap --format=csv,noheader | head -1)
    echo "  GPU compute capability: ${GPU_CC}"
fi

if ! command -v g++-13 &> /dev/null; then
    echo "WARNING: g++-13 not found."
    echo "  CUDA 12.8 nvcc requires gcc-13/g++-13 on Ubuntu 22.04+."
    echo "  Install: sudo apt install gcc-13 g++-13"
fi

if ! dpkg -s libsparsehash-dev &> /dev/null 2>&1; then
    echo "WARNING: libsparsehash-dev not found."
    echo "  Install: sudo apt install libsparsehash-dev"
    if [[ -t 0 ]]; then
        read -p "Continue anyway? [y/N] " -r
        [[ $REPLY =~ ^[Yy]$ ]] || exit 1
    else
        echo "  Non-interactive shell — aborting."
        exit 1
    fi
fi

# --- [2/6] Create / activate venv ---
if [ ! -d "$VENV_DIR" ]; then
    echo "[2/6] Creating virtual environment (Python $PYTHON_VERSION)..."
    uv venv --python "$PYTHON_VERSION"
else
    echo "[2/6] Virtual environment already exists, reusing..."
fi
source "$VENV_DIR/bin/activate"

# --- [3/6] Install Python dependencies ---
echo "[3/6] Installing Python dependencies (PyTorch cu128 + PyG + spconv)..."
# --reinstall: 기존 venv 재사용 시 torch 등이 상위 버전으로 남아 있을 수 있으므로
#              requirements pin 으로 강제 복구.
uv pip install --reinstall -r requirements.cu128.txt

# --- [4/6] Verify PyTorch sees Blackwell ---
echo "[4/6] Verifying PyTorch / CUDA runtime..."
python - <<'PY'
import torch
print(f"  torch {torch.__version__}")
print(f"  torch.version.cuda = {torch.version.cuda}")
print(f"  cuda available     = {torch.cuda.is_available()}")
if torch.cuda.is_available():
    cc = torch.cuda.get_device_capability(0)
    print(f"  GPU                = {torch.cuda.get_device_name(0)}")
    print(f"  compute capability = sm_{cc[0]}{cc[1]}")
    arches = torch.cuda.get_arch_list()
    print(f"  supported arches   = {arches}")
    if not any(a.endswith("_120") or a.endswith("_100") for a in arches):
        print("  WARNING: this PyTorch build has no Blackwell (sm_100/sm_120) kernels.")
PY

# --- [5/6] Build C++/CUDA extensions with Blackwell arch ---
echo "[5/6] Building libs/pointops (and siblings) for sm_120..."
# Ship kernels for Ampere + Ada + Hopper + Datacenter-Blackwell + Consumer-Blackwell.
export TORCH_CUDA_ARCH_LIST="8.0;8.6;8.9;9.0;10.0;12.0"
echo "  TORCH_CUDA_ARCH_LIST=${TORCH_CUDA_ARCH_LIST}"

EXT_LIBS=("pointops" "pointops2" "pointgroup_ops" "pointrope" "pointseg")
for lib in "${EXT_LIBS[@]}"; do
    if [ -d "libs/$lib" ] && [ -f "libs/$lib/setup.py" ]; then
        echo "  -> building libs/$lib"
        (cd "libs/$lib" && rm -rf build/ dist/ *.egg-info && uv pip install --no-build-isolation .)
    fi
done

# --- [6/6] Smoke test ---
echo "[6/6] Running smoke test..."
python - <<'PY'
import torch
assert torch.cuda.is_available(), "CUDA not available"
x = torch.randn(1024, 3, device="cuda")
y = (x @ x.T).sum()
torch.cuda.synchronize()
print(f"  CUDA matmul OK  (sum={y.item():.3f})")
try:
    import pointops
    print("  pointops import OK")
except Exception as e:
    print(f"  pointops import FAILED: {e}")
try:
    import spconv.pytorch as spconv
    print("  spconv import OK  (cu126 wheel on CUDA 12.8 runtime; PTX JIT on sm_120)")
except Exception as e:
    print(f"  spconv import FAILED: {e}")
PY

echo ""
echo "=== Installation complete ==="
echo "Activate env: source ${VENV_DIR}/bin/activate"

실행하면 대략 다음과 같이 출력된다:

$ ./uv-install-cu128.sh
=== Pointcept GPU (CUDA 12.8 / Blackwell sm_120) Installation ===
[1/6] Checking system prerequisites...
  nvcc 12.8 OK
  GPU compute capability: 12.0
[2/6] Creating virtual environment (Python 3.10)...
Using CPython 3.10.19
Creating virtual environment at: .venv
Activate with: source .venv/bin/activate
[3/6] Installing Python dependencies (PyTorch cu128 + PyG + spconv)...
Resolved 164 packages in 6.47s
Prepared 164 packages in 14ms
Installed 164 packages in 109ms
 + absl-py==2.4.0
 + accelerate==1.13.0
 + addict==2.4.0
 + aiohappyeyeballs==2.6.1
 + aiohttp==3.13.5
 + aiosignal==1.4.0
 + annotated-doc==0.0.4
 + annotated-types==0.7.0
 + anyio==4.13.0
 + asttokens==3.0.1
 + async-timeout==5.0.1
 + attrs==26.1.0
 + blinker==1.9.0
 + ccimport==0.4.4
 + certifi==2022.12.7
 + charset-normalizer==2.1.1
 + click==8.3.2
 + clip==1.0 (from git+https://github.com/openai/CLIP.git@d05afc436d78f1c48dc0dbf8e5980a9d471f35f6)
 + comm==0.2.3
 + configargparse==1.7.5
 + contourpy==1.3.2
 + cumm-cu126==0.7.11
 + cycler==0.12.1
 + dash==4.1.0
 + decorator==5.2.1
 + einops==0.8.2
 + exceptiongroup==1.3.1
 + executing==2.2.1
 + fastjsonschema==2.21.2
 + filelock==3.25.2
 + fire==0.7.1
 + flask==3.1.3
 + fonttools==4.62.1
 + frozenlist==1.8.0
 + fsspec==2026.2.0
 + ftfy==6.3.1
 + gitdb==4.0.12
 + gitpython==3.1.46
 + grpcio==1.80.0
 + h11==0.16.0
 + h5py==3.16.0
 + hf-xet==1.4.3
 + httpcore==1.0.9
 + httpx==0.28.1
 + huggingface-hub==1.11.0
 + idna==3.4
 + importlib-metadata==7.1.0
 + ipython==8.39.0
 + ipywidgets==8.1.8
 + itsdangerous==2.2.0
 + jedi==0.19.2
 + jinja2==3.1.6
 + joblib==1.5.3
 + jsonschema==4.26.0
 + jsonschema-specifications==2025.9.1
 + jupyter-core==5.9.1
 + jupyterlab-widgets==3.0.16
 + kiwisolver==1.5.0
 + lark==1.3.1
 + markdown==3.10.2
 + markdown-it-py==4.0.0
 + markupsafe==3.0.3
 + matplotlib==3.10.8
 + matplotlib-inline==0.2.1
 + mdurl==0.1.2
 + mpmath==1.3.0
 + multidict==6.7.1
 + narwhals==2.19.0
 + nbformat==5.10.4
 + nest-asyncio==1.6.0
 + networkx==3.4.2
 + ninja==1.13.0
 + numpy==2.2.6
 + nvidia-cublas-cu12==12.8.3.14
 + nvidia-cuda-cupti-cu12==12.8.57
 + nvidia-cuda-nvrtc-cu12==12.8.61
 + nvidia-cuda-runtime-cu12==12.8.57
 + nvidia-cudnn-cu12==9.7.1.26
 + nvidia-cufft-cu12==11.3.3.41
 + nvidia-cufile-cu12==1.13.0.11
 + nvidia-curand-cu12==10.3.9.55
 + nvidia-cusolver-cu12==11.7.2.55
 + nvidia-cusparse-cu12==12.5.7.53
 + nvidia-cusparselt-cu12==0.6.3
 + nvidia-nccl-cu12==2.26.2
 + nvidia-nvjitlink-cu12==12.8.61
 + nvidia-nvtx-cu12==12.8.55
 + open3d==0.19.0
 + packaging==24.1
 + pandas==2.3.3
 + parso==0.8.6
 + pccm==0.4.16
 + peft==0.19.1
 + pexpect==4.9.0
 + pillow==12.1.1
 + platformdirs==4.9.6
 + plotly==6.7.0
 + plyfile==1.1.3
 + portalocker==2.10.1
 + prompt-toolkit==3.0.52
 + propcache==0.4.1
 + protobuf==7.34.1
 + psutil==7.2.2
 + ptyprocess==0.7.0
 + pure-eval==0.2.3
 + pybind11==3.0.4
 + pydantic==2.13.2
 + pydantic-core==2.46.2
 + pygments==2.20.0
 + pyparsing==3.3.2
 + pyquaternion==0.9.9
 + python-dateutil==2.9.0.post0
 + pytz==2026.1.post1
 + pyyaml==6.0.3
 + referencing==0.37.0
 + regex==2026.4.4
 + requests==2.28.1
 + retrying==1.4.2
 + rich==15.0.0
 + rpds-py==0.30.0
 + safetensors==0.7.0
 + scikit-learn==1.7.2
 + scipy==1.15.3
 + sentry-sdk==2.58.0
 + setuptools==70.2.0
 + sharedarray==3.2.4
 + shellingham==1.5.4
 + six==1.17.0
 + smmap==5.0.3
 + spconv-cu126==2.3.8
 + stack-data==0.6.3
 + sympy==1.14.0
 + tensorboard==2.20.0
 + tensorboard-data-server==0.7.2
 + tensorboardx==2.6.5
 + termcolor==3.3.0
 + threadpoolctl==3.6.0
 + timm==1.0.26
 + tokenizers==0.22.2
 + tomli==2.4.1
 + torch==2.7.0+cu128
 + torch-cluster==1.6.3+pt27cu128
 + torch-geometric==2.7.0
 + torch-scatter==2.1.2+pt27cu128
 + torch-sparse==0.6.18+pt27cu128
 + torchaudio==2.7.0+cu128
 + torchvision==0.22.0+cu128
 + tqdm==4.66.5
 + traitlets==5.14.3
 + transformers==5.5.4
 + triton==3.3.0
 + typer==0.24.1
 + typing-extensions==4.15.0
 + typing-inspection==0.4.2
 + tzdata==2026.1
 + urllib3==1.26.13
 + wandb==0.26.0
 + wcwidth==0.6.0
 + werkzeug==3.1.8
 + widgetsnbextension==4.0.15
 + xxhash==3.6.0
 + yapf==0.43.0
 + yarl==1.23.0
 + zipp==3.19.2
[4/6] Verifying PyTorch / CUDA runtime...
  torch 2.7.0+cu128
  torch.version.cuda = 12.8
  cuda available     = True
  GPU                = NVIDIA RTX PRO 2000 Blackwell Generation Laptop GPU
  compute capability = sm_120
  supported arches   = ['sm_75', 'sm_80', 'sm_86', 'sm_90', 'sm_100', 'sm_120', 'compute_120']
[5/6] Building libs/pointops (and siblings) for sm_120...
  TORCH_CUDA_ARCH_LIST=8.0;8.6;8.9;9.0;10.0;12.0
  -> building libs/pointops
Using Python 3.10.19 environment at: /home/zer0/Projects/Pointcept/.venv
Resolved 27 packages in 1.87s
      Built pointops @ file:///home/zer0/Projects/Pointcept/libs/pointops
Prepared 1 package in 1m 33s
Installed 1 package in 0.83ms
 + pointops==1.0 (from file:///home/zer0/Projects/Pointcept/libs/pointops)
  -> building libs/pointops2
Using Python 3.10.19 environment at: /home/zer0/Projects/Pointcept/.venv
Resolved 27 packages in 1.58s
      Built pointops2 @ file:///home/zer0/Projects/Pointcept/libs/pointops2
Prepared 1 package in 1m 16s
Installed 1 package in 0.96ms
 + pointops2==1.0 (from file:///home/zer0/Projects/Pointcept/libs/pointops2)
  -> building libs/pointgroup_ops
Using Python 3.10.19 environment at: /home/zer0/Projects/Pointcept/.venv
Resolved 1 package in 1.41s
      Built pointgroup-ops @ file:///home/zer0/Projects/Pointcept/libs/pointgroup_ops
Prepared 1 package in 16.03s
Installed 1 package in 0.81ms
 + pointgroup-ops==0.0.0 (from file:///home/zer0/Projects/Pointcept/libs/pointgroup_ops)
  -> building libs/pointrope
Using Python 3.10.19 environment at: /home/zer0/Projects/Pointcept/.venv
Resolved 1 package in 2.17s
      Built pointrope @ file:///home/zer0/Projects/Pointcept/libs/pointrope
Prepared 1 package in 1m 39s
Installed 1 package in 0.69ms
 + pointrope==0.0.0 (from file:///home/zer0/Projects/Pointcept/libs/pointrope)
  -> building libs/pointseg
Using Python 3.10.19 environment at: /home/zer0/Projects/Pointcept/.venv
Resolved 26 packages in 2.29s
      Built pointseg @ file:///home/zer0/Projects/Pointcept/libs/pointseg
Prepared 1 package in 31.73s
Installed 1 package in 0.44ms
 + pointseg==1.0 (from file:///home/zer0/Projects/Pointcept/libs/pointseg)
[6/6] Running smoke test...
  CUDA matmul OK  (sum=1048.214)
  pointops import OK
  spconv import OK  (cu126 wheel on CUDA 12.8 runtime; PTX JIT on sm_120)

=== Installation complete ===
Activate env: source .venv/bin/activate

Favorite site

GitHub - Pointcept/Pointcept