Worked on the sglang repository to address a KV cache shape inconsistency encountered when using input embeddings in Python-based back end development. Implemented a targeted fix that resets output_ids for requests involving input_embeds during retraction, ensuring the KV cache maintains correct shapes and preventing token generation errors. This solution was validated against various input_embeds token generation scenarios to confirm reliability and robustness. By eliminating potential runtime errors caused by mis-shaped KV caches, the work improved the stability of embeddings-based prompting and retraction workflows. The approach demonstrated careful attention to runtime behavior and reliability in Python back end systems.