Why relax onnx frontend can not support dynamic slice

irasin · November 19, 2024, 12:55pm

It seems only constant slice is supported now

apache/tvm/blob/main/python/tvm/relax/frontend/onnx/onnx_frontend.py#L1742


steps = get_constant(inputs[4], params)
if not all(
    [
        (
            isinstance(param, (relax.Constant, relax.ShapeExpr, relax.PrimValue))
            or param is None
        )
        for param in [starts, ends, axes, steps]
    ]
):
    raise ValueError("Only constant Slice parameters are currently supported.")
# Convert parameters to constant lists.
starts = get_prim_expr_list(starts)
ends = get_prim_expr_list(ends)
if axes is not None:
    axes = get_prim_expr_list(axes)
else:
    axes = list(range(len(starts)))
# Convert negative axis to positive if needed.
for i, axis in enumerate(axes):
    if axis < 0:

I wonder why can not support dynamic slice in onnx_frontend, since we already have the relax.dynamic_strided_slice op here

github.com

apache/tvm/blob/main/python/tvm/relax/op/index.py#L102


        The sliced result.


    Note
    ----
    strided_slice require the input `begin`, `end` and `strides` to have the
    same length as `axes`.
    """
    return _ffi_api.strided_slice(x, axes, begin, end, strides, assume_inbound)  # type: ignore




def dynamic_strided_slice(
    x: Expr,
    begin: Expr,
    end: Expr,
    strides: Expr,
) -> Expr:
    """Dynamic strided slice of a tensor. `begin`, `end`, `strides` can be computed at runtime.


    Parameters
    ----------
    x : Expr