[BUG] nnfusion jit data type mismatch with float16 precision

[LeiWang1999]

🐛 Bug. nnfusion jit data type mismatch with float16 precision

Use nnfusion python interface to compile and build a resnet50.float16.onnx model, I got several issues.

Firstly, nnfusion jit currently doesn't support float16 datatype in dtypes.py :

str2type = {
    "float":
    TypeObject._make(["float32", ctypes.c_float, torch.float32,
                      numpy.float32]),
    "float32":
    TypeObject._make(["float32", ctypes.c_float, torch.float32,
                      numpy.float32]),
    "double":
    TypeObject._make(
        ["float64", ctypes.c_double, torch.float64, numpy.float64]),
    "float64":
    TypeObject._make(["float64", ctypes.c_double, torch.float64,
                      numpy.float64]),
    "int8":
    TypeObject._make(["int8", ctypes.c_int8, torch.int8, numpy.int8]),
    "int16":
    TypeObject._make(["int16", ctypes.c_int16, torch.int16, numpy.int16]),
    "int32":
    TypeObject._make(["int32", ctypes.c_int32, torch.int32, numpy.int32]),
    "int64":
    TypeObject._make(["int64", ctypes.c_int64, torch.int64, numpy.int64]),
    "uint8":
    TypeObject._make(["uint8", ctypes.c_uint8, torch.uint8, numpy.uint8]),
    "uint16":
    TypeObject._make(["uint8", ctypes.c_uint16, None, numpy.uint16]),
    "uint32":
    TypeObject._make(["uint8", ctypes.c_uint32, None, numpy.uint32]),
    "uint64":
    TypeObject._make(["uint8", ctypes.c_uint64, None, numpy.uint64]),
}

I append float16 datatype map to solve this problem, but the followed issue comes up:

Traceback (most recent call last):
  File "resnet.py", line 53, in <module>
    nnf_execute()
  File "resnet.py", line 48, in nnf_execute
    executor({input_name: data_desc}, {output_name: nnf_out_desc})
  File "/workspace/v-leiwang3/nnfusion/src/python/nnfusion/executor.py", line 181, in __call__
    self.feed_data(*args, **kwargs)
  File "/workspace/v-leiwang3/nnfusion/src/python/nnfusion/executor.py", line 203, in feed_data
    raise Exception(
Exception: Shape or type mismatch for NNFusion model input input, expect [(64, 3, 224, 224), half], feed [(64, 3, 224, 224), float16]

This was caused by the para_info.json generated by nnfusion codegen, the input and output datatype is half while the nnfusion jit gave the datatype float16 and caused this mismatch exception.

{
    "input": {
        "input": {
            "id": "((int64_t*)(inputs[0]))",
            "name": "Parameter_351_0",
            "shape": [
                1,
                512
            ]
        }
    },
    "output": {
        "output": {
            "id": "((half*)(outputs[0]))",
            "name": "Result_1912_0",
            "shape": [
                1,
                100
            ]
        }
    }
}