Run Hecate's 9.6 µm 2D predictor on Windows Radeon GPUs through an optional DirectML helper.
I was using the released models to explore PHerc0139 and PHerc0813 on a Windows laptop with Radeon 740M integrated graphics. This adds a GPU path for that machine while keeping the original checkpoint, CPU/CUDA command and probability-image format.
On the same real PHerc0139 96×96 crop, the original CPU predictor took 13.02 s and DirectML took 3.54 s. All 9,216 image pixels matched exactly. Those timings cover prediction and blending after model loading. Reversed depth, padded edges and empty-column checks also matched exactly; the fresh Zarr-to-PNG command matched the array result. CPU fallback is disabled, and profiling showed only DirectML execution.
The change is separate from `hecate.py`: one helper, optional dependencies, a real-data verification script, and instructions. It supports 9.6 µm FP32 2D inference at batch size one. It reuses the existing volume reader, tile origins and PNG …
