text
stringlengths
0
12.9k
This repository contains pure-Python wheels and compiled Windows MMCV wheels.
Each artifact is specific to its stated Python, PyTorch, CUDA, and platform
profile. Do not install every artifact together as a generic environment.
## Experimental Windows CUDA 12.8 / SM120 MMCV wheel
`mmcv-2.2.0-1cu128sm120-cp310-cp310-win_amd64.whl` is built from unmodified
MMCV 2.2.0 source for native Windows, 64-bit CPython 3.10, PyTorch 2.7.1+cu128,
and compute capability 12.0 only. It contains 58 CUDA ELF binaries targeting
SM120. It must not replace the separate CUDA 12.1 wheel or be used on Linux,
ARM64, Python 3.11, or other Blackwell compute capabilities.
A fresh Windows RTX 5050 Laptop GPU environment passed a synchronized Torch
matrix check, MMCV CUDA NMS, RoIAlign forward/backward against Torchvision,
deformable convolution forward/backward against ordinary convolution, and
modulated deformable convolution forward/backward. Its `pip check` passed.
Publication checks verified the exact SHA-256, ZIP/RECORD integrity, AMD64 PE
header, and embedded CUDA architecture flags. Reports and reproducible build
scripts are under [`builds/windows-cu128-sm120/`](builds/windows-cu128-sm120/).
**Full MAAP3D validation has NOT RUN:** the full installer, MMPose,
MMDetection, SAM3, and real checkpoint/video workflows remain pending.
This is an experimental component artifact, not a validated full environment.
The CUDA 12.8 MAAP3D installer downloads it with an immutable revision and
SHA-256 pin, installs it with `--no-deps`, and retains the full runtime checks.
It requires every visible NVIDIA GPU to have compute capability 12.0.
No compiler or full CUDA toolkit is needed to install this prebuilt MMCV
wheel. Its DLL imports require the matching Python/PyTorch CUDA runtime and
Microsoft Visual C++ 2015–2022 x64 runtime (MSVCP140/VCRUNTIME140).
SAM3 may have additional runtime compilation requirements.
MMCV's [`LICENSES.md`](licenses/mmcv/LICENSES.md) identifies NVIDIA-licensed
special operations. Complete [StyleGAN2](licenses/mmcv/NVIDIA-StyleGAN2-LICENSE.txt)
and [StyleGAN3](licenses/mmcv/NVIDIA-StyleGAN3-LICENSE.txt) license texts accompany
these artifacts; their research/evaluation use restrictions remain applicable.
The general Apache license does not override those operation-specific terms.
## What was changed
- **MMDetection:** widens the package's runtime MMCV version guard so the
validated MAAP3D stack can use MMCV 2.2.0, adds an inline downstream
modification notice, and uses a downstream version. The optional `mim`
dependency metadata retains upstream's `mmcv<2.2.0` declaration; this wheel
is therefore not intended to be installed through the `[mim]` extra.
- **SAM3:** relaxes the `ftfy` dependency for the BoxMOT-compatible stack,
declares an explicit PyTorch minimum, removes an incorrect MIT classifier,
adds a modification notice, and uses a downstream version. Version `.2`
lowers the declared minimum from PyTorch 2.7 to the validated PyTorch 2.4
runtime without replacing the immutable `.1` artifact. The SAM License is a
custom license with use and redistribution
conditions; read [`licenses/sam3/LICENSE`](licenses/sam3/LICENSE) in full.
- **Chumpy:** packages the pinned upstream Python 3.12-compatible revision
instead of patching Chumpy 0.70 in `site-packages`, and uses a downstream
version.
- **MMPose:** omits eager EDPose and RTMO registry imports needed by unsupported
GB10 paths and uses a downstream version. EDPose source and configuration
files remain in the wheel; the special-file licensing recorded in
[`licenses/mmpose_gb10/LICENSES.md`](licenses/mmpose_gb10/LICENSES.md) still
applies. Consult that notice together with the original
[ED-Pose license file](https://github.com/IDEA-Research/ED-Pose/blob/master/LICENSE);
this repository does not resolve or simplify those upstream terms.
Exact upstream repositories, commits, checksums, and modification summaries
are recorded in [`PROVENANCE.json`](PROVENANCE.json).
## Installation and profile restrictions
The preferred installation path is the matching MAAP3D profile installer,
which is responsible for selecting compatible PyTorch, CUDA, MMCV, and other
dependencies. These wheels are not complete environment bundles.
- The GB10 MMPose wheel must be installed with `--no-deps` after the
profile-specific `xtcocotools` build.
- Use SAM3 `.2` with the modern CUDA 12.1 / PyTorch 2.4 profile. The older
`.1` wheel retains its original PyTorch >=2.7 metadata for reproducibility.
- The MMPose wheel is limited to the GB10 `mmcv-lite` profile. Its disabled
eager registrations mean the EDPose and MMDetection-dependent RTMO paths
are not supported in that profile.
- The MMDetection wheel changes the runtime compatibility guard only; do not
request its optional `[mim]` extra.
For a manual download, select one artifact and pin a published tag or, for
maximum immutability, the full repository commit SHA:
```bash
hf download MAAP3D/dependency-wheels \
FILENAME.whl \
--repo-type dataset \
--revision TAG_OR_FULL_COMMIT_SHA \
--local-dir ./maap3d-wheels
```
Install the downloaded wheel only after preparing the matching profile. For
example, the GB10 MMPose artifact deliberately requires `--no-deps`:
```bash
python -m pip install --no-deps ./maap3d-wheels/mmpose-1.3.2+maap3d.gb10.5408bc76.1-py3-none-any.whl
```