[Feature Request] Add support for fp8 QDQ models.

[hopef]

In the latest version of the onnx-simplifier, I have met errors on fp8 QDQ models.

• Error 1: shape inference

$> onnxsim fp8.onnx fp8-sim.onnx
Simplifying...
Traceback (most recent call last):
  File "/opt/conda3/bin/onnxsim", line 8, in <module>
    sys.exit(main())
             ^^^^^^
  File "/opt/conda3/lib/python3.11/site-packages/onnxsim/onnx_simplifier.py", line 453, in main
    model_opt, check_ok = simplify(
                          ^^^^^^^^^
  File "/opt/conda3/lib/python3.11/site-packages/onnxsim/onnx_simplifier.py", line 187, in simplify
    model_opt_bytes = C.simplify(
                      ^^^^^^^^^^^
onnx.onnx_cpp2py_export.shape_inference.InferenceError: [ShapeInferenceError] (op_type:QuantizeLinear, node name: pts_bbox_head.transformer.decoder.layers.0.attentions.0.attn.query_quantizer/QuantizeLinear1): [TypeInferenceError] Inferred elem type differs from existing elem type: (17) vs (INT8)

• Error 2: CSETensorHash

$> onnxsim fp8.onnx fp8-sim.onnx --skip-shape-inference
Simplifying...
Traceback (most recent call last):
  File "/opt/conda3/bin/onnxsim", line 8, in <module>
    sys.exit(main())
             ^^^^^^
  File "/opt/conda3/lib/python3.11/site-packages/onnxsim/onnx_simplifier.py", line 453, in main
    model_opt, check_ok = simplify(
                          ^^^^^^^^^
  File "/opt/conda3/lib/python3.11/site-packages/onnxsim/onnx_simplifier.py", line 187, in simplify
    model_opt_bytes = C.simplify(
                      ^^^^^^^^^^^
RuntimeError: no supported data type: 17

Could you please add support for fp8 QDQ models?
The fp8 QDQ models are aligned with ModelOpt, which has the following structure. Compared to int8 QDQ models, the only difference is the data_type of zero points (int8 vs. float8_e4m3fn).

x_scales = torch.ones(1, dtype=torch.float32)
x_zero_points = torch.zeros(1, dtype=torch.float8_e4m3fn)
w_scales = torch.ones(32, dtype=torch.float32)
w_zero_points= torch.zeros(32, dtype=torch.float8_e4m3fn)

x = Q(x, x_scales, x_zero_points)
x = DQ(x, x_scales, x_zero_points)

quant_weights = Q(weights, w_scales, w_zero_points)
quant_weights = DQ(quant_weights , w_scales, w_zero_points)
y = Conv(x, quant_weights)

[OValery16]

I also have the same problem

[congyang12345]

You can try using onnxslim