vLLM fixes GLM-5.2 shared-expert fusion and MTP on ROCm
September 29, 2026
vLLM developers fixed shared-expert fusion and Multi-Token Prediction (MTP) issues for GLM-5.2 on the ROCm DSA path. This improves model performance and compatibility for AMD hardware users.
HOW THIS AFFECTS YOU
●
builderYou can now run GLM-5.2 with optimized fusion and MTP on AMD ROCm platforms.