保存paddle模型时infer_slots_data插入位置导致预估出core

请提出你的问题 Please ask your question

将infer_slots_data 放到sequence_pool 之前，预估出core，放到sequence_pool 之后，则不出core，求问原因

for data in slots:
emb = fluid.layers.embedding(input=data, size=[self._dict_dim, self._emb_dim], is_sparse=True, is_distributed=True, param_attr=fluid.ParamAttr(name="embedding"))
self.infer_slots_data.append(emb)
bow = fluid.layers.sequence_pool(input=emb, pool_type='sum')
#self.infer_slots_data.append(emb)

core信息：
#0 0x00007f1fba52422b in paddle::operators::ConcatOp::GetExpectedKernelType(paddle::framework::ExecutionContext const&) const ()

Start with the reported Python snippet, comparing the graph when infer_slots_data is appended before versus after sequence_pool. Then inspect the ConcatOp::GetExpectedKernelType frame from the core output and reproduce the model-save or inference path. Done means identifying why the insertion order triggers the crash and documenting a confirmed fix or reproducible diagnosis.