Nablo404/birefnet-lite-webgpu

BiRefNet_lite (fp16), patched to run fully on WebGPU

This is onnx-community/BiRefNet_lite-ONNX (onnx/model_fp16.onnx), an ONNX export of ZhengPeng7/BiRefNet_lite. The graph is rewritten so it runs entirely on WebGPU in the browser. The weights are unchanged.

Why

On some GPUs the original fails in ONNX Runtime Web's WebGPU backend. It was measured on Intel Iris Xe in Chrome, where three problems showed up:

1. Concat/Split nodes with up to 1024 inputs or 32 outputs exceed the per-shader storage-buffer limit: "Too many storage buffers in shader. Current: 17, Max is 16".
2. The deformable-convolution index math is int64. WebGPU has no int64, so those nodes run on the CPU.
3. There is no WebGPU Sum kernel. Large fp16 tensors were therefore summed on the CPU, which went past the browser's 4 GB memory limit (std::bad_alloc).

What changed

• Each oversized Concat or Split is now a tree of the same op, each using at most 8 buffers. 8 is WebGPU's guaranteed minimum.
• The int64 index chains are int32, with a cast back to int64 right before each GatherND.
• Each Sum is now a chain of Adds.

The output is bit-identical to the original: all 1,048,576 output values matched on a test image (onnxruntime-node, CPU).

Usage (Transformers.js)

import { AutoModel, AutoProcessor, RawImage } from "@huggingface/transformers";
const model = await AutoModel.from_pretrained("<this repo>", { dtype: "fp16", device: "webgpu" });
const processor = await AutoProcessor.from_pretrained("<this repo>");
const { pixel_values } = await processor(await RawImage.fromURL(url));
const { output_image } = await model({ input_image: pixel_values });

Licence

MIT, the same as the original. See LICENSE. The original copyright notice is kept as the licence requires.