🌟 Summary
v8.4.192 improves compatibility with PyTorch 2.15 and CUDA 13 for multi-GPU training and ONNX Runtime GPU inference, alongside a small update to Platform response documentation.
📊 Key Changes
- More reliable multi-GPU training: On PyTorch 2.8 and newer, each NCCL process group is explicitly bound to the GPU selected for its training rank. This helps prevent device-mismatch errors, especially when GPUs are selected in a non-default order.
- More dependable ONNX GPU inference: Linux CUDA setups now install ONNX Runtime’s CUDA dependencies and preload their libraries, helping prevent a missing cuRAND library from causing inference to fall back to the CPU.
- Updated CUDA tooling: The Docker training image now uses PyTorch 2.15, and the export image and GPU CI setup include the required ONNX Runtime CUDA package.
- Clearer Platform API docs: Prediction response examples no longer list
metadata.classNames. Class names remain available on individual detections; dataset and model class names are unchanged.
🎯 Purpose & Impact
- 🖥️ Fewer multi-GPU training failures for setups that use GPUs in a custom order or select GPUs dynamically.
- ⚡ More reliable GPU acceleration for ONNX Runtime on supported Linux CUDA environments, reducing unintended CPU fallback.
- 🧰 Better alignment across development, CI, and Docker environments when using newer PyTorch and CUDA dependencies.
- 📦 Slightly leaner, more accurate API examples that match the Platform prediction response.
No model architecture or prediction behavior changes are indicated in this release.
What's Changed
- Remove classNames from Platform predict response docs by @glenn-jocher in #26919
- Support torch 2.15 CUDA in DDP training and ONNX Runtime inference by @raimbekovm in #26928
Full Changelog: v8.4.191...v8.4.192