OpenDPD Studio 2.2.12¶
Fixes issue #50: PA training could fail with Failed to find C compiler when DeltaGRU first initialized its optional Triton CUDA kernels.
- Local DeltaGRU training and inference now fall back to PyTorch on the same device when Triton cannot find a C compiler. The layer warns once and continues with the same weights, states and gradients. Unrelated CUDA and kernel failures still surface normally.
- The hosted GPU image includes GCC and the Python development headers needed by Triton. Generated shared libraries use a private, size-limited temporary cache; uploaded workspaces and the ordinary temporary directory remain mounted
noexec. - The GPU worker now verifies cold-cache DeltaGRU forward and backward execution under the production container restrictions before accepting jobs. This catches missing compilers, headers and unusable caches before a user starts training.
Install with uv pip install "opendpd==2.2.12" --torch-backend=auto, or open Studio on the web.