Crash with ONNX Split operator

[abcdrm]

When I make a ONNX node Split:

node = onnx.helper.make_node(
    "Split",
    inputs=["input"],
    outputs=["output_1", "output_2", "output_3"],
    axis=1
)

The ONNX split doc says that the split input is optional. So I did not use this input.
Then I run this node with ppl.nn.llm, which gives:

[ERROR][2023-11-03 16:34:43.746][reshape_split.cc:51] ERROR: output_count[3] != split point size[2]
[ERROR][2023-11-03 16:34:43.746][kernel.cc:39] Reshape kernel[ppl_anonymous_node_0] failed: invalid value
[ERROR][2023-11-03 16:34:43.746][sequential_scheduler.cc:130] exec kernel[ppl_anonymous_node_0] of type[:Split:13] failed: invalid value
[ERROR][2023-11-03 16:34:43.746][runtime_impl.cc:333] Run() failed: invalid value
[ERROR][2023-11-03 16:34:43.746][pplnn_llm.cc:1213] Run() failed: invalid value

I check the code in pplnn/src/ppl/nn/oputils/onnx/reshape_split.cc

RetCode ReshapeSplit(InputOutputInfo* info, const ir::Attr* arg, const int64_t* split_point) {
    auto param = static_cast<const SplitParam*>(arg);
    auto split_size = info->GetInput<TensorImpl>(1)->GetShape()->GetDim(0);
    const TensorShape& shape = *info->GetInput<TensorImpl>(0)->GetShape();
    int dim_count = shape.GetDimCount();
    uint32_t split_axis = param->axis >= 0 ? param->axis : param->axis + dim_count;

I print info->GetInputCount(), which gives 1, means it has a single input. But split_size accesses info->GetInput<TensorImpl>(1), which is the second input. Plz check this issue.

Also in ppl.nn.llm src/ppl/nn/engines/llm_cuda/ops/onnx/split_op.cc.
Line 59-62:

auto split = info->GetInput<TensorImpl>(1);
if (!split->GetBufferPtr()) {
    return RC_NOT_FOUND;
}

This check seems not work.