[Feature]: GLM 5.3 Performance Optimization
Author: yewentao256Created Sep 17, 2026Updated Sep 17, 2026
Labelsfeature requestglm
The feature, motivation and pitch
Following up to https://github.com/vllm-project/vllm/issues/46654
GLM-5.3 uses the same glm_moe_dsa base architecture as GLM-5.2 so the optimizations we did before should still apply, there are still some work we can do to optimize it further
CUDA
- https://github.com/vllm-project/vllm/pull/52164 @zobinHuang
- https://github.com/vllm-project/vllm/pull/53781 @MatthewBonanni
- https://github.com/vllm-project/vllm/pull/56629 @LucasWilkinson
ROCm
- https://github.com/vllm-project/vllm/pull/44527 @frida-andersson
- https://github.com/vllm-project/vllm/pull/45149 @tvirolai-amd
- https://github.com/vllm-project/vllm/pull/46832 @frida-andersson
- https://github.com/vllm-project/vllm/pull/47335 @xaguilar-amd
- https://github.com/vllm-project/vllm/pull/51314 @amd-sriram
- https://github.com/vllm-project/vllm/pull/51315 @amd-sriram
Source: vllm-project/vllm