[Bug]: Vitis quantization doesn't work with ORT 1.16.

[guotuofeng]

What happened?

The following error happens with the test case if the ORT is 1.16.0. The root cause is the calibrator API is changed by PR
microsoft/onnxruntime@d0316ee#diff-118260dbcc5091cfd805764c7a1a71be2727d89f118ae1bc9f6c31d5fa75c2a1

2023-09-19T23:04:00.2420563Z =================================== FAILURES ===================================
2023-09-19T23:04:00.2421087Z _______________________ test_vitis_ai_quantization_pass ________________________
2023-09-19T23:04:00.2421343Z 
2023-09-19T23:04:00.2433300Z tmp_path = PosixPath('/tmp/pytest-of-cloudtest/pytest-0/test_vitis_ai_quantization_pas0')
2023-09-19T23:04:00.2434332Z 
2023-09-19T23:04:00.2434920Z     def test_vitis_ai_quantization_pass(tmp_path):
2023-09-19T23:04:00.2435303Z         # setup
2023-09-19T23:04:00.2435642Z         input_model = get_onnx_model()
2023-09-19T23:04:00.2436074Z         dummy_user_script = str(tmp_path / "dummy_user_script.py")
2023-09-19T23:04:00.2436476Z         dummy_data = str(tmp_path / "dummy_data")
2023-09-19T23:04:00.2437015Z         with open(dummy_user_script, "w") as f:
2023-09-19T23:04:00.2437342Z             f.write(" ")
2023-09-19T23:04:00.2456174Z         if not os.path.exists(dummy_data):
2023-09-19T23:04:00.2457734Z             os.mkdir(dummy_data)
2023-09-19T23:04:00.2458345Z     
2023-09-19T23:04:00.2458831Z         config = {"user_script": dummy_user_script, "data_dir": dummy_data, "dataloader_func": dummy_calibration_reader}
2023-09-19T23:04:00.2459324Z         output_folder = str(tmp_path / "vitis_ai_quantized")
2023-09-19T23:04:00.2459701Z     
2023-09-19T23:04:00.2460062Z         # create VitisAIQuantization pass
2023-09-19T23:04:00.2460521Z         p = create_pass_from_dict(VitisAIQuantization, config, disable_search=True)
2023-09-19T23:04:00.2461028Z         # execute
2023-09-19T23:04:00.2461414Z >       quantized_model = p.run(input_model, None, output_folder)
2023-09-19T23:04:00.2461644Z 
2023-09-19T23:04:00.2462266Z test/unit_test/passes/vitis_ai/test_vitis_ai_quantization.py:53: 
2023-09-19T23:04:00.2462983Z _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
2023-09-19T23:04:00.2463480Z ***/passes/***_pass.py:400: in run
2023-09-19T23:04:00.2464201Z     output_model = self._run_for_config(model, data_root, config, output_model_path)
2023-09-19T23:04:00.2464637Z ***/passes/onnx/vitis_ai_quantization.py:347: in _run_for_config
2023-09-19T23:04:00.2464969Z     quantize_static(
2023-09-19T23:04:00.2465304Z ***/passes/onnx/vitis_ai/quantize.py:234: in quantize_static
2023-09-19T23:04:00.2465654Z     quantizer = VitisQDQQuantizer(
2023-09-19T23:04:00.2466173Z ***/passes/onnx/vitis_ai/quantizer.py:471: in __init__
2023-09-19T23:04:00.2466534Z     ONNXQuantizer.__init__(
2023-09-19T23:04:00.2467046Z _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
2023-09-19T23:04:00.2467253Z 
2023-09-19T23:04:00.2467583Z self = <***.passes.onnx.vitis_ai.quantizer.VitisQDQQuantizer object at 0x7fb2987f98b0>
2023-09-19T23:04:00.2467963Z model = ir_version: 5
2023-09-19T23:04:00.2468259Z producer_name: "pytorch"
2023-09-19T23:04:00.2468759Z producer_version: "2.0.0"
2023-09-19T23:04:00.2469229Z graph {
2023-09-19T23:04:00.2469775Z   node {
2023-09-19T23:04:00.2470043Z     input: "input"
2023-09-19T23:04:00.2470420Z     input: "fc1.w....infer"
2023-09-19T23:04:00.2470787Z   value: "onnxruntime.quant"
2023-09-19T23:04:00.2471100Z }
2023-09-19T23:04:00.2471683Z metadata_props {
2023-09-19T23:04:00.2472149Z   key: "onnx.quant.pre_process"
2023-09-19T23:04:00.2472604Z   value: "onnxruntime.quant"
2023-09-19T23:04:00.2473005Z }
2023-09-19T23:04:00.2473247Z 
2023-09-19T23:04:00.2473661Z per_channel = False, reduce_range = False
2023-09-19T23:04:00.2474079Z mode = <QuantizationMode.QLinearOps: 1>, static = True
2023-09-19T23:04:00.2474535Z weight_qType = <QuantType.QInt8: 0>, activation_qType = <QuantType.QUInt8: 1>
2023-09-19T23:04:00.2475314Z tensors_range = {'/fc1/Gemm_output_0': (-0.890625, 1.1015625), 'input': (0.0, 0.2490234375), 'output': (0.0, 0.99609375)}
2023-09-19T23:04:00.2475850Z nodes_to_quantize = None, nodes_to_exclude = None
2023-09-19T23:04:00.2476460Z op_types_to_quantize = ['Relu', 'Add', 'Clip', 'AveragePool', 'Transpose', 'ConvTranspose', ...]
2023-09-19T23:04:00.2477336Z extra_options = {'ActivationSymmetric': False, 'AddQDQPairToWeight': False, 'WeightSymmetric': True}
2023-09-19T23:04:00.2477907Z 
2023-09-19T23:04:00.2478172Z     def __init__(
2023-09-19T23:04:00.2478433Z         self,
2023-09-19T23:04:00.2478684Z         model,
2023-09-19T23:04:00.2478958Z         per_channel,
2023-09-19T23:04:00.2479462Z         reduce_range,
2023-09-19T23:04:00.2479725Z         mode,
2023-09-19T23:04:00.2480314Z         static,
2023-09-19T23:04:00.2480574Z         weight_qType,
2023-09-19T23:04:00.2480856Z         activation_qType,
2023-09-19T23:04:00.2481119Z         tensors_range,
2023-09-19T23:04:00.2481397Z         nodes_to_quantize,
2023-09-19T23:04:00.2481667Z         nodes_to_exclude,
2023-09-19T23:04:00.2481939Z         op_types_to_quantize,
2023-09-19T23:04:00.2482232Z         extra_options=None,
2023-09-19T23:04:00.2482488Z     ):
2023-09-19T23:04:00.2482774Z         if not model_has_infer_metadata(model):
2023-09-19T23:04:00.2483103Z             model = save_and_reload_model_with_shape_infer(model)
2023-09-19T23:04:00.2483452Z         self.value_infos = {vi.name: vi for vi in model.graph.value_info}
2023-09-19T23:04:00.2483830Z         self.value_infos.update({ot.name: ot for ot in model.graph.output})
2023-09-19T23:04:00.2484188Z         self.value_infos.update({it.name: it for it in model.graph.input})
2023-09-19T23:04:00.2484495Z     
2023-09-19T23:04:00.2484938Z         self.model = ONNXModel(model)
2023-09-19T23:04:00.2485259Z         if not static:
2023-09-19T23:04:00.2485560Z             self.model.replace_gemm_with_matmul()
2023-09-19T23:04:00.2485849Z     
2023-09-19T23:04:00.2486289Z         self.per_channel = per_channel  # weight-pack per channel
2023-09-19T23:04:00.2486648Z         self.reduce_range = reduce_range
2023-09-19T23:04:00.2486989Z         self.mode = mode  # QuantizationMode.Value
2023-09-19T23:04:00.2487336Z         self.static = static  # use static quantization for inputs.
2023-09-19T23:04:00.2487682Z         self.fuse_dynamic_quant = False
2023-09-19T23:04:00.2488114Z     
2023-09-19T23:04:00.2488617Z         self.extra_options = extra_options if extra_options else {}
2023-09-19T23:04:00.2488954Z         self.enable_subgraph_quantization = (
2023-09-19T23:04:00.2489297Z             "EnableSubgraph" in self.extra_options and self.extra_options["EnableSubgraph"]
2023-09-19T23:04:00.2489625Z         )
2023-09-19T23:04:00.2489895Z         self.force_quantize_no_input_check = (
2023-09-19T23:04:00.2490274Z             "ForceQuantizeNoInputCheck" in self.extra_options and self.extra_options["ForceQuantizeNoInputCheck"]
2023-09-19T23:04:00.2490625Z         )
2023-09-19T23:04:00.2490967Z         self.q_matmul_const_b_only = "MatMulConstBOnly" in self.extra_options and self.extra_options["MatMulConstBOnly"]
2023-09-19T23:04:00.2491338Z         self.is_weight_symmetric = (
2023-09-19T23:04:00.2491657Z             weight_qType in (QuantType.QInt8, QuantType.QFLOAT8E4M3FN)
2023-09-19T23:04:00.2492005Z             if "WeightSymmetric" not in self.extra_options
2023-09-19T23:04:00.2492320Z             else self.extra_options["WeightSymmetric"]
2023-09-19T23:04:00.2492610Z         )
2023-09-19T23:04:00.2492874Z         self.is_activation_symmetric = (
2023-09-19T23:04:00.2493234Z             False if "ActivationSymmetric" not in self.extra_options else self.extra_options["ActivationSymmetric"]
2023-09-19T23:04:00.2493577Z         )
2023-09-19T23:04:00.2493809Z     
2023-09-19T23:04:00.2494119Z         self.activation_qType = activation_qType.tensor_type
2023-09-19T23:04:00.2494447Z         self.weight_qType = weight_qType.tensor_type
2023-09-19T23:04:00.2494738Z         """
2023-09-19T23:04:00.2495058Z             Dictionary specifying the min and max values for tensors. It has following format:
2023-09-19T23:04:00.2495371Z                 {
2023-09-19T23:04:00.2495644Z                     "param_name": [min, max]
2023-09-19T23:04:00.2495902Z                 }
2023-09-19T23:04:00.2496161Z             example:
2023-09-19T23:04:00.2496403Z                 {
2023-09-19T23:04:00.2496810Z                     'Conv_3:0': [np.float32(0), np.float32(0.5)],
2023-09-19T23:04:00.2497277Z                     'Conv_4:0': [np.float32(1), np.float32(3.5)]
2023-09-19T23:04:00.2497971Z                 }
2023-09-19T23:04:00.2498396Z         """
2023-09-19T23:04:00.2498809Z         if tensors_range is not None and any(map(lambda t: not isinstance(t, TensorData), tensors_range.values())):
2023-09-19T23:04:00.2499723Z >           raise TypeError(
2023-09-19T23:04:00.2500181Z                 f"tensors_range contains unexpected types {set(type(v) for v in tensors_range.values())}, not TensorData."
2023-09-19T23:04:00.2500613Z             )
2023-09-19T23:04:00.2501417Z E           TypeError: tensors_range contains unexpected types {<class 'tuple'>}, not TensorData.
2023-09-19T23:04:00.2501659Z 
2023-09-19T23:04:00.2502160Z /opt/hostedtoolcache/Python/3.8.18/x64/lib/python3.8/site-packages/onnxruntime/quantization/onnx_quantizer.py:129: TypeError
2023-09-19T23:04:00.2502726Z ------------------------------ Captured log call -------------------------------
2023-09-19T23:04:00.2503230Z INFO     ***.passes.onnx.vitis_ai_quantization:vitis_ai_quantization.py:290 Preprocessing model for quantization

Version?

ORT 1.16.0