FLUX.2-klein-4B-ONNX-WebGPU

The transformer of black-forest-labs/FLUX.2-klein-4B, converted to ONNX for running in a web browser with ONNX Runtime Web's WebGPU backend, as used by Quick Inference.

Changes from the original

• The transformer only: one ONNX graph taking the noisy latents (x, one token per 16 × 16 pixels, 128 values each), their positions (x_ids), the timestep (0–1), the text features (ctx, three Qwen3 layers side by side) and their positions (ctx_ids), and returning the flow-matching velocity (pred). Its inputs and outputs match MarkShark2/flux2-klein-4b-onnx-webgpu-q4, whose text encoder and VAE decoder go with it.
• The linear layers in the blocks are quantised to 4 bits (round to nearest, block size 32); attention and the projections run in float16, the residual streams, norms, modulation and rotary embedding in float32.

These are modified files: their outputs are close to, but not identical to, the original model's.

Licence

Apache 2.0, as the original (LICENSE.md).