Worked on performance and stability improvements for GLM workloads on ROCm, focusing on both feature development and runtime reliability. In the jeejeelee/vllm repository, implemented Fused Shared Expert support for GLM-4.5, 6, and 7, integrating AITER-based fusion logic and updating weight loading to efficiently handle widened expert tensors, which reduced inference overhead in Mixture-of-Experts models. In ROCm/aiter, tuned and stabilized BF16 GEMM configurations for GLM-4.7-FP8, removing unstable hipBLASLt entries and defaulting to dynamic configuration paths to ensure compatibility across builds. Utilized Python, PyTorch, and ROCm, emphasizing GPU programming and inference optimization.