.safetensors does not support a native int4 type. The strategy is to pack uint8 / uint32 and add quantization metadata to describe the actual packing. Investigate if there is an already existing quantization description used by other exporters which matches what .webnn requires and use it or invent a new one especially for webnn.
.safetensors does not support a native int4 type. The strategy is to pack uint8 / uint32 and add quantization metadata to describe the actual packing. Investigate if there is an already existing quantization description used by other exporters which matches what .webnn requires and use it or invent a new one especially for webnn.