[compile_fixture] Scale must be a constant tensor in quantize / dequantize op.

[markkraay]

Related to compile_fixture (#37). test_dequantize and test_quantize fail compiler.compile with the following: Example Code:

        data = [4, 8]
        input_tp = tp.Tensor(data, dtype=tp.int8)
        scale = torch.tensor(0.5, dtype=TORCH_DTYPES[dtype])
        scale_tp = tp.Tensor(scale, dtype=dtype)
        dequantized = compile_fixture(tp.dequantize, input_tp, scale_tp, dtype)

Error Msg:

E       tripy.common.exception.TripyException: 
E       
E       --> /tripy/tests/integration/conftest.py:36 in wrapper()
E             |
E          36 |             compiled_func = compiler.compile(*compile_args, **compile_kwargs)
E             | 
E       
E       Scale must be a constant tensor in dequantize op.

tripy/common/exception.py:195: TripyException

[yizhuoz004]

@markkraay Please provide more information to debug this issue.

• Your PR is not merged, can it be reproduced by existing code?
• What is dtype here?
• Scale must be a constant tensor in dequantize op. Then what is the op instead?

[yizhuoz004]

The root cause is in tripy/tripy/backend/compiler_api.py:L515, compiler uses full to create dummy input values for intial compilation. The fixes could be:

1. implement const fold optimization for Fill op when shape is constant.
2. The compile_fixture should add an argument to specify which tensors are inputs, while others are Parameters.