Worked on the kvcache-ai/sglang repository to deliver variable-length attention support within a multimodal generation framework. Developed a custom operation in Python using PyTorch to enable robust handling of input sequences of varying lengths, addressing integration challenges with torch.compile and flash attention v4. This engineering effort improved the flexibility and reliability of the attention layer, ensuring stable builds and smoother deployment in production environments. By expanding the range of viable input lengths and reducing runtime failures, the work enhanced the framework’s ability to process diverse data. The focus remained on deep learning, attention mechanisms, and seamless integration with existing infrastructure.