Dataflow management provides limited performance improvement to the transformer model due to its lesser weight reuse than the convolution neural network. The cosFormer reduced computational complexity while achieving comparable performance to the vanilla transformer for natural language processing tasks. However, the unstructured sparsity in the cosFormer makes it a challenge to be implemented efficiently. This brief proposes a parallel unstructured sparsity handling (PUSH) scheme to compute sparse-dense matrix multiplication (SDMM) efficiently. It transforms unstructured sparsity into structured sparsity and reduces the total memory access by balancing the memory accesses of the sparse and dense matrices in the SDMM. We also employ unstructured weight pruning cooperating with PUSH to further increase the structured sparsity of the model. Through verification on an FPGA platform, the proposed accelerator achieves a throughput of 2.82 TOPS and an energy efficiency of 144.8 GOPs/W for HotpotQA dataset with long sequences.