Should we extend qnn.dense to qnn.matmul to handle the transpose in quantized models? We are facing similar issues in quantized bert models.
cc @comaniac @jcf94