[Discussion] A Technical Approach to LLMs with TVM Unity

Thanks TQ for the great question. We are working on dlight, a lightweight auto-scheduler for dynamic shape workloads. After that, users are able to define their own models with different architectures.