dkeviv/Qwen3.5-0.8B-Text-ONNX-WebGPU-QDQ-bucket

Qwen3.5 0.8B Text ONNX — WebGPU QDQ repair

This bucket is derived from `onnx-community/Qwen3.5-0.8B-Text-ONNX`. It keeps the source q4f16 packed weights (the external data file remains approximately 469 MB) and rewrites the two custom operators that prevented standard WebGPU execution:

• —GatherBlockQuantized embedding lookup is expressed as standard Gather, nibble unpacking, and dequantization operators.
• —Qwen3.5 CausalConvWithState is expressed as state concat, depthwise causal convolution arithmetic, and SiLU using standard ONNX operators.

The source model's Apache-2.0 license and attribution are retained. This is a WebGPU compatibility derivative; validate the exact browser/runtime combination before using it in production.