[dequantizeLinear] OV only support axis == 0

[lisa0314]

Phi-3-mini failed while creating session due to this validation code- https://github.com/openvinotoolkit/openvino/blob/master/src/frontends/onnx/frontend/src/op/dequantize_linear.cpp#L228.

For block-wise dequantizeLinear support, OV first reshape input to a shape [x.shape[0]/block_size, block_size, x.shape[1]]. Then do unsqueeze to zeroPoint and scale using axes = {1}. Finally do broadcasting for scale and zeroPoint. I suppose OV restricts the value of axis to 0 is to simplify the emulation.

Maybe we need to insert transpose before and after dequantizeLinear? @huningxin WDYT?

[lisa0314]

@Honry the axis of dequantizeLinear operator for phi3-mini always be 1. Am I right?

[lisa0314]

After inserting transpose before and after dequantizeLinear, this issue can be solved. However, an out-of-memory exception will be thrown during creating session for decode process. It seems there is not enough memory resource.

[huningxin]

@Honry the axis of dequantizeLinear operator for phi3-mini always be 1. Am I right?

Seems correct.

[Honry]

@Honry the axis of dequantizeLinear operator for phi3-mini always be 1. Am I right?

Seems correct.

Yes, according to the model builder tool of GenAI, the DequantizeLinear's axis always set to axis=-1, as its input is 2D, axis is always 1.

[huningxin]

@lisa0314 , please help file an OV issue.