Method _infer_concrete_shape_from_args

python/tvm/relax/base_py_module.py:306–346

(self, shape, in_args)

        return out_tensors

    def _infer_concrete_shape_from_args(self, shape, in_args):
        concrete = []
        symbolic_positions = []
        for idx, dim in enumerate(shape):
            if isinstance(dim, int | np.integer):
                concrete.append(int(dim))
            elif isinstance(dim, tirx.IntImm):
                concrete.append(int(dim.value))
            else:
                concrete.append(None)
                symbolic_positions.append(idx)

        if not symbolic_positions:
            return concrete

        candidates = []
        if in_args is not None:
            if not isinstance(in_args, list | tuple):
                in_args = [in_args]
            for obj in in_args:
                if hasattr(obj, "shape") and isinstance(obj.shape, tuple | list):
                    try:
                        candidates.append(tuple(int(x) for x in obj.shape))
                        continue
                    except (ValueError, TypeError):
                        # Skip objects with invalid shapes
                        pass

        target_ndim = len(shape)
        for cand in candidates:
            if len(cand) == target_ndim:
                for pos in symbolic_positions:
                    concrete[pos] = cand[pos]
                if all(x is not None for x in concrete):
                    return concrete

        raise ValueError(
            "Cannot infer concrete output shape from symbolic shape and inputs. "
            "Please provide a concrete `out_sinfo` (e.g., a tuple/list of ints) "
            "or ensure input tensors carry shapes that determine output extents."
        )

    def _convert_tvm_dtype_to_torch(self, tvm_dtype: str) -> "torch.dtype":
        """Convert TVM dtype string to PyTorch dtype."""