v0.1.2

CPU wheels for CPython 3.10–3.12 (manylinux x86_64/aarch64, Windows AMD64, macOS ARM). Filenames are flowedge-0.1.2-*. Not on PyPI. CUDA stays a source backend (FLOWEDGE_BACKEND=cuda).

The repository now lives at https://github.com/reforcemind/FlowEdge. Docs: https://reforcemind.github.io/FlowEdge/

Published matched Diffusion Policy replay is unchanged: CPU 851 vs 1409 ms, CUDA GTX 1650 131 vs 345 ms. Not a 10 ms loop.

v0.1.1 Release assets remain flowedge-0.1.0-* (pyproject on that tag was 0.1.0).

v0.1.1

Patch on v0.1.0 so cibuildwheel can target macOS 14: floating from_chars is unavailable before macOS 26, so deployment-profile floats use strtof/snprintf.

CPU wheels (CPython 3.10-3.12, manylinux x86_64/aarch64, win_amd64, macosx_arm64) and the sdist are attached to this Release. The wheel PEP 440 version is 0.1.0 because pyproject.toml on this tag was not bumped; the GitHub tag is v0.1.1.

pip install example:

python -m pip install flowedge-0.1.0-cp310-cp310-win_amd64.whl

CUDA is not in these wheels. PyPI was not published. Source citation may still use v0.1.0.

v0.1.0

First tagged source release of FlowEdge, a C++23 inference runtime for real-time robotics policies (flow-matching, static memory, deterministic execution, deadline-aware local inference).

This tag is for citation. It is not a PyPI upload. cibuildwheel artifacts for CPython 3.10-3.12 are attached to the Build and publish wheels workflow on this release.

Included on this tag:

• Mamba + flow-matching path and a fixed LeRobot diffusion_pusht path
• Packed Conv1D, k-major ConvTranspose1D, and short-row GEMM on CPU
• Matched PushT replay on one Windows/Clang laptop: policy p50 851 ms vs LeRobot/PyTorch 1409 ms at threads=1 (0.60x). That is not a 10 ms control loop.
• Fake-robot loop at --period-ms 10 --on-miss hold: 20/20 misses (p50 655 ms, 4 threads)
• LeRobot plugin: --policy.type=flowedge and flowedge_smolvla; --on-miss hold|drop|raise
• Local Grok Bot verification gate; GitHub Actions is compile + ctest

Not in this release: CUDA/Metal/Vulkan, native SmolVLA VLM, PyPI wheels, Jetson or SO-100 hardware evidence.