libreyolo

LibreYOLO v1.7.3, a patch release that makes LibreYOLO("owner/repo") load Hugging Face Hub models on a plain pip install libreyolo (#1020).

Fixed

• LibreYOLO("owner/repo") works on a plain pip install libreyolo. huggingface_hub is now a core dependency, so loading a model by its Hugging Face repo id no longer needs the hf extra. The extra still exists for pushing to the Hub and gated downloads.

• Loading LibreYOLO/LibreYOLO9t, LibreYOLO/LibreYOLOXs and the other YOLO9 and YOLOX repos by repo id. These repos also hold a legacy lowercase checkpoint, so the loader stopped with "multiple checkpoints". When a repo holds several checkpoints, the one named after the repo is now chosen.

• A libreyolo/ folder in the working directory no longer hides Hub repo ids on macOS and Windows. The check for a local LibreYOLO/ folder ignored case on those filesystems, so running from a folder containing a LibreYOLO clone made LibreYOLO("LibreYOLO/<repo>") look for a local file.

Changed

• TFLite INT8 export quantizes once instead of twice. onnx2tf -oiqt builds the integer model twice — float32 I/O, then int8 I/O — and only the second is shipped, although the I/O dtype does not change the calibrated scales. export(format="tflite", int8=True, ...) now builds the SavedModel with onnx2tf and runs a single in-process TFLiteConverter pass; int8_backend="onnx2tf" keeps the previous route, and int8_output_dtype="float32" writes float outputs for host-dequantizing runtimes (in-process backend only).

LibreYOLO v1.7.2, a patch release that makes pip install "libreyolo[all]" and the other extras install cleanly with pip and uv on Linux, Windows and macOS (#989).

Changed

• libreyolo[all] no longer includes paddle. (#989) X2Paddle needs ONNX 1.17 or earlier while the TFLite exporter in all needs ONNX 1.20, so the two cannot share an environment. Install libreyolo[paddle] on its own; Paddle export needs Python 3.10 or 3.11.

Fixed

• pip install "libreyolo[all]" works on Python 3.12 and 3.13, Apple Silicon and Linux ARM. (#989) On Python 3.12 and newer the paddle and tflite extras in all pinned incompatible ONNX versions, so pip failed or fell back to an older LibreYOLO release. On macOS and Linux ARM, vlm, vlm-train, ground and all failed on decord, which ships wheels for Linux x86_64 and Windows only; it is now installed only there, the platforms where LocateAnything can run. tensorrt-cu12, which has no ARM wheels, is skipped on Linux ARM, where JetPack provides TensorRT. Loading LocateAnything where decord is not available now says so before downloading its weights, and Core AI export on Python 3.10 says it needs Python 3.11 instead of repeating the install command. Thanks @gboeer.

• Every extra installs without a compiler on Linux x86_64 and Windows, and on Apple Silicon with Python 3.12+. (#989) pip compiled three packages from source, which fails on a machine without the matching toolchain. tensorrt (and so all) pulled pycuda, which needs the CUDA toolkit headers; it is gone. TensorRT export and inference never needed it, and INT8 calibration uses cuda.bindings, which PyPI torch installs on Linux (elsewhere, pip install cuda-python). tflite pulled onnx-simplifier, a C++ build that duplicated the onnxsim module the onnx extra already provides; it is gone. coreml pulled coremltools on Windows and Linux ARM, where it has no wheels; it is now installed on macOS and Linux x86_64 only, and CoreML export elsewhere says so. Remaining source builds, all upstream: on Linux ARM, tflite (and so all) on Python 3.12+ compiles two packages onnx2tf pins exactly, and executorch compiles coremltools; triton (and so all) compiles gevent on macOS with Python 3.10 or 3.11 (needs the Xcode command line tools) and on Linux ARM with Python 3.10.

• pip install "libreyolo[executorch]" gets a runtime that loads. (#989) ExecuTorch 1.2's compiled runtime is built against torch 2.11 but accepts any newer torch, so pip paired it with torch 2.15 and loading a .pte failed with Symbol not found in libc10. The extra now pins torch>=2.11,<2.12.

• MNN and ExecuTorch export work from a virtualenv that is not activated. (#989) Running the venv's python directly (an IDE run button, a Jupyter kernel, venv/bin/python script.py) left mnnconvert and ExecuTorch's flatc off PATH, so MNN export said mnnconvert was missing and ExecuTorch export could not start flatc. Both are now found in the environment's own scripts directory.

• A failing import MNN on current Linux says why. (#989) MNN's Linux wheel asks for an executable stack, which glibc 2.41 and newer refuse (Ubuntu 25.04+, Fedora 42+), so MNN export and inference failed with "Install with: pip install libreyolo[mnn]" although MNN was installed. The error now names the cause and the workarounds: GLIBC_TUNABLES=glibc.rtld.execstack=2 (glibc 2.42+) or patchelf --clear-execstack.

• uv sync resolves from a source checkout. (#989) The dev group asked for pytest 9 while ExecuTorch 1.2 requires pytest below 9, coreai-torch was requested on Python 3.10 although it needs 3.11, and extras that cannot share an environment (vla with molmo2 or marigold, executorch or paddle with tflite) were not declared as conflicts. A new Dependency resolution workflow resolves uv lock and every extra on Python 3.10 to 3.13 for Linux x86_64, Linux

LibreYOLO v1.7.1, a patch release: unguarded training scripts run on macOS and Windows, and export fixes for TorchScript, OpenVINO, Paddle and ExecuTorch from the 1.7.0 test campaign.

Changed

• RF-DETR PyTorch boxes are clipped to the image. (#980) Every exported backend (ONNX, TensorRT, OpenVINO, ExecuTorch, ...) clips RF-DETR boxes to the image, and upstream RF-DETR does since rf-detr#1168; PyTorch predict and val did not, so a box could extend past the picture. COCO val2017 mAP50-95 moves by +0.00003 to +0.00008 (n/s/m/l).

• YOLO9-E2E TFLite export is validated. (#983) With onnx2tf 2.6.9 or newer and LiteRT 2.1.2, the trained LibreYOLO9E2Et checkpoint passes raw and public predict parity at a 128x128 canvas, so its tier moves from blocked to validated. Older onnx2tf versions, which change its predicted classes, are refused with an upgrade hint. YOLO9-P2 TFLite stays blocked.

• The executorch extra installs ExecuTorch 1.2. (#983) pip install libreyolo[executorch] could resolve ExecuTorch 1.0 with an older Torch; it now pins the validated executorch>=1.2,<1.3.

Fixed

• Unguarded train() and val() scripts run on macOS and Windows. (#971) A call at the top level of a script, outside if __name__ == "__main__":, now runs with workers=0 and logs one line on how to get parallel data loading back, instead of stopping with an error (#954). This applies under the spawn and forkserver start methods; guarded scripts, notebooks, the CLI and Linux fork keep their workers.

• TorchScript exports of YOLO9, YOLO9-E2E and YOLO9-P2 run on any device. (#974) An export traced on CPU, the TorchScript default, failed on CUDA with "Expected all tensors to be on the same device": the head built its anchor grid inside the traced graph, which kept the trace device. The exporter now pins the grid as constants for the fixed export canvas, and torch.jit.load(map_location=...) moves them with the weights. FP16 exports included.

• OpenVINO FP32 and INT8 exports keep full-precision weights. (#975) openvino.save_model compresses weights to FP16 unless told otherwise, so half=False exports wrote FP16 weights. FP16 exports are unchanged.

• A failed Paddle export raises an error instead of ending the script. (#975) X2Paddle calls exit(0) when its static-graph conversion fails, which ended the calling process as a success with no model written. On Python 3.12 and later, where paddlepaddle 2.6.2 cannot build that static graph, Paddle export now says so before converting.

• RF-DETR ExecuTorch programs load on Linux. (#979) The Linux ExecuTorch 1.2 runtime has no aten::grid_sampler_2d kernel (the macOS one does), so RF-DETR detect, segment, pose and OBB .pte files failed to load with "Missing operator" (error 0x14). The exporter now decomposes grid sampling into primitives every runtime registers, as the Core AI exporter already does.

• A missing image directory says so. (#979) A data yaml whose train/val folder does not exist (a dataset that was never downloaded, a typo) raised ValueError: Unsupported path format; it now raises FileNotFoundError: Image directory not found.

• LibreYOLO("LibreRFDETRn-sem.pt") loads by name. (#981) The only published DINOv2 semantic weight keeps its pre-split RF-DETR name; it loaded from a path, but auto-download stopped with "unable to determine size from filename". It now downloads and loads as LibreDINOv2 semantic.

• Two instances of one LibreGen model can load at the same time. (#986) They share one weights directory, and the first loads could race on the download's completion marker: one instance read it half written, removed it and raised FileNotFoundError although the download had succeeded. Snapshot preparation is now serialized per directory and the marker is written atomically.

• The 3D generation knob error no longer lists imgsz. (#970) Speridl

LibreYOLO v1.7.0, image-to-3D and text-to-image generation with LibreGen, the WeVisOne-VL VLM family, validation that explains more of its results, and YOLO9 training fixes.

116 commits, 27 merged pull requests. Several changes alter training results, metrics or output paths you may compare against v1.6.0: read Changed and Breaking changes before upgrading. Full details in CHANGELOG.md and the upgrade guide.

Highlights

• Image to 3D with Speridlabs Idol (#964): LibreGen("speridlabs-idol")("chair.jpg").save("chair.ply") turns one photo of an object into the whole object as 3D Gaussian splats (.ply or .splat). Ported from the SDK of Speridlabs, with Apache-2.0 code and weights; nothing gated, no login. New generate3d task.
• Text to image (#933): LibreGen("flux2-klein-4b")("a red fox"), with FLUX.2 klein 4B and SANA-Sprint 0.6B. New generate task, CLI libreyolo generate, extra libreyolo[gen].
• WeVisOne-VL (#957): Tencent's 2B and 4B VLMs detect, point, segment, matte and read text through LibreVLM.
• Validation (#931): a confusion matrix on detect, segment and classify, per-class P/R/F1 curves on detect and segment, class-agnostic NMS in predict and val, and top1/top5 on classify results.
• YOLO9 training fixes (#930, #939): fine-tunes no longer start with a randomly initialised PGI branch, gradients are clipped, labels are assigned in fp32, and label-assignment memory no longer grows with max_labels. Existing checkpoints load and give bit-identical float outputs.
• CLI (#930): libreyolo train clip_max_norm= for the families whose recipe clips gradients.

Changed

• YOLO9-family training (#930): the gradient norm is clipped at 10, labels are assigned in fp32, and fine-tunes from the published weights no longer attach a randomly initialised PGI branch. clip_max_norm=0 turns clipping off.
• YOLO9-m/c PGI branch (#930): the upstream CBLinear/CBFuse branch, so PGI training uses more memory; inference is unchanged.
• perspective (#930) warps by corner displacement in every family with the shared affine warp, so the same value gives a different warp.
• Mixup ratio (#930) is drawn from Beta(1, 1): default YOLOv7 training changes.
• CLI run folders (#946): EC, D-FINE, SegFormer, LingBot-Vision, V-JEPA 2 and YOLO9-P2 use their family run name. name=exp (or name=yolo9_p2_exp) gives the old folder.
• Confusion matrix plots (#931): counts in confusion_matrix.png, the normalized plot in confusion_matrix_normalized.png, detections counted at confidence 0.25.
• Numbers that move: RTMDet-Ins keeps up to max_det masks, 300 by default, instead of at most 100 (#953); segmentation val treats box-only label rows as rectangle masks (#945); val(single_cls=True) scores a multiclass model's detections as one class (#938).

Breaking changes

• YOLO9-family checkpoints saved by 1.7.0 do not load in 1.6.x (#930). Older checkpoints load in 1.7.0.
• YOLO9 module, attribute and ONNX node names changed, with no aliases (#930): update code that reaches into model.model and Hailo end-node configs.
• libreyolo.validation.ConfusionMatrix is transposed (#931): matrix[predicted, true], default conf_thres=0.25.
• predict(agnostic_nms=True) raises where it is not supported (#931), and an unknown classifier interpolation raises (#942).

Fixed

• A missing libreyolo[rfdetr] extra is named instead of a misleading error (#965).
• CLI training uses the SegFormer, LingBot-Vision and D-FINE recipes and loads RF-DETR segment checkpoints (#955).
• Unguarded train()/val() scripts stop with one clear line instead of multiprocessing tracebacks (#954).
• libreyolo train device=0,1 trains on several GPUs (#91