Semantic parsing using sequence-to-sequence models allows parsing of deeper\nrepresentations compared to traditional word tagging based models. In spite of\nthese advantages, widespread adoption of these models for real-time\nconversational use cases has been stymied by higher compute requirements and\nthus higher latency. In this work, we propose a non-autoregressive approach to\npredict semantic parse trees with an efficient seq2seq model architecture. By\ncombining non-autoregressive prediction with convolutional neural networks, we\nachieve significant latency gains and parameter size reduction compared to\ntraditional RNN models. Our novel architecture achieves up to an 81% reduction\nin latency on TOP dataset and retains competitive performance to non-pretrained\nmodels on three different semantic parsing datasets. Our code is available at\nhttps://github.com/facebookresearch/pytext\n