Is your feature request related to a problem? Please describe.
I use tiler in a deployment pipeline to split large images into smaller patches before running model inference. As far as I understand, the current tiling and merging operations are handled on the CPU.
In my use case, the patches are eventually fed into a GPU model. This means data may need to move between CPU and GPU, and the tiling/merging steps can become a bottleneck, especially for large images or high-throughput inference workloads.
Describe the solution you'd like
It would be useful to support GPU-side tiling and merging, for example with PyTorch, CuPy, or another GPU array backend.
Ideally, the workflow could keep image tensors on the GPU throughout the pipeline:
large image tensor on GPU
→ split into patches on GPU
→ model inference on GPU
→ merge patches back on GPU
→ final output
This could reduce CPU-GPU data transfers and improve latency in deployment scenarios.
Additional context
This would be especially helpful for patch-based inference on large images, such as segmentation or restoration models, where overlapping patches need to be merged after inference.
Is your feature request related to a problem? Please describe.
I use
tilerin a deployment pipeline to split large images into smaller patches before running model inference. As far as I understand, the current tiling and merging operations are handled on the CPU.In my use case, the patches are eventually fed into a GPU model. This means data may need to move between CPU and GPU, and the tiling/merging steps can become a bottleneck, especially for large images or high-throughput inference workloads.
Describe the solution you'd like
It would be useful to support GPU-side tiling and merging, for example with PyTorch, CuPy, or another GPU array backend.
Ideally, the workflow could keep image tensors on the GPU throughout the pipeline:
This could reduce CPU-GPU data transfers and improve latency in deployment scenarios.
Additional context
This would be especially helpful for patch-based inference on large images, such as segmentation or restoration models, where overlapping patches need to be merged after inference.