Python¶

CPU wheels for CPython 3.10–3.12 ship with GitHub Releases. They are not on PyPI yet. CUDA is a source build (FLOWEDGE_BACKEND=cuda), not a wheel extra.

Install a wheel¶

Open v0.1.2.
Download the .whl for your OS and CPython from that Release (stable URL, no Actions login). Filenames are flowedge-0.1.2-*. CUDA is not in the wheel.
Install it:

python -m pip install flowedge-0.1.2-*.whl

Platform	Wheel tag
Linux x86_64	`manylinux_2_28_x86_64`
Linux aarch64	`manylinux_2_28_aarch64`
Windows	`win_amd64`
macOS ARM	`macosx_arm64`

From a clone, python -m pip install . still compiles with a C++23 toolchain. That is the source path, not the first-run path.

Sample flow matching¶

mkdir -p models
curl -L "https://huggingface.co/ReForceMind/mamba_flow/resolve/main/mamba_flow.safetensors" \
  -o models/mamba_flow.safetensors
python examples/core/flow_sample.py models/mamba_flow.safetensors euler 10

import numpy as np, flowedge
e = flowedge.Engine("models/mamba_flow.safetensors")
a = e.sample(prefix=np.array([1, 2, 3, 4], dtype=np.int32),
             noise=np.random.randn(e.action_dim).astype("float32"),
             steps=10, method="euler")

That prefix is the ULP gate (~1e-6 rel vs PyTorch), not a policy p50.

Load may allocate once. Supported hot paths allocate nothing after init. One engine, one stream. The published wheel backend is CPU.

Zero-copy run_into / sample_into live in the Python API.

CUDA¶

Wheels do not include CUDA kernels. v0.1.2 is a CPU wheel, not PyPI. CUDA is a source build (FLOWEDGE_BACKEND=cuda). The first GPU user has to survive nvcc plus WSL or MSVC; a 4 GB card must not brick fe_engine_load.

On Windows, nvcc needs an MSVC-compatible host compiler. The MinGW/Clang tree used for CPU Release builds cannot host nvcc. Use WSL or MSVC.

cmake -S . -B build-cuda -DCMAKE_BUILD_TYPE=Release -DFLOWEDGE_BACKEND=cuda -DFLOWEDGE_PYTHON=ON
cmake --build build-cuda -j
python -c "import flowedge; e=flowedge.Engine('models/diffusion_pusht.flowedge.safetensors'); print(flowedge.cuda, e.cuda_resident)"

flowedge.cuda is compile-time. Engine.cuda_resident is runtime: 1 only if Mamba / flow / Diffusion Policy weights actually uploaded. A GTX 1650 4 GB WDDM/WSL host often cannot cudaMalloc the PushT U-Net (~959 MiB). Load then keeps CPU kernels; that is not an engine failure. FLOWEDGE_CUDA_REQUIRED=1 fails closed if attach did not happen. FLOWEDGE_CUDA_FORCE_HOST=1 skips the doomed upload.

--device cuda in the LeRobot adapter requires both a CUDA Core binary and cuda_resident. The published CUDA headline stays GTX 1650 matched replay 131 vs 345 ms. Fusion / BF16 stay ungated until isolate and native DDIM / policy p50 move on a card that can load the U-Net.

See CUDA.

PyPI (not published)¶

pip install flowedge is not live. On 2026-09-19 https://pypi.org/pypi/flowedge/json returned 404, so the name is still free. publish.yml already builds the same CPU wheels and can upload with Trusted Publishing when the repository variable FLOWEDGE_PUBLISH_PYPI is true. CUDA stays a source backend after that upload; do not advertise a CUDA extra that the wheel does not contain.

Maintainer checklist: PyPI Trusted Publisher for flowedge on reforcemind/FlowEdge → GitHub environment pypi → set FLOWEDGE_PUBLISH_PYPI=true → publish a tagged release.