@vinx13 The problem might be similar to the discussion in [Quantization] How to quantize transpose and nn.pad operators?
I am doing a quantize with Informer model, but after quantization, only conv1D operator was successfully quantized, other operators like transpose, nn.max_pool1d, add, mul, reshape which should have been supported in TVM quantization (tvm\python\tvm\relay\quantize_annotate.py) are not quantized.
The model IR code like below:
%12 = cast(%11, dtype="int8") /* ty=Tensor[(8, 7, 74), int8] */;
%13 = nn.conv1d(%12, meta[relay.Constant][1] /* ty=Tensor[(512, 7, 3), int8] */, channels=512, kernel_size=[3], out_dtype="float32") /* ty=Tensor[(8, 512, 72), float32] */;
%4 = transpose(%dec_inp3, axes=[0, 2, 1]) /* ty=Tensor[(8, 7, 72), float32] */;
...
%15 = multiply(%14, 0.000546359f /* ty=float32 */) /* ty=Tensor[(8, 512, 72), float32] */;
Could anyone help me out?Thanks a lot.