Checklist

• 1. If the issue you raised is not a feature but a question, please raise a discussion at https://github.com/sgl-project/sglang/discussions/new/choose Otherwise, it will be closed.
• 2. Please use English, otherwise it will be closed.

Motivation

We find out during GLM4.5 onboarding that it is very difficult to tune mem-fraction-static, implying potential bugs in memory estimation logic when handling spec decoding / mtp models. Creating a ticket for tracking. (cc @zhyncs @hebiao064 )

Behaviors:

• None of the GLM4.5 model can run without explicitly specifying mem-fraction-static when MTP is enabled.
• The full weight (300B) model can start successfully with 0.8, but the lighter weight Air version (100B) instead will run into OOM unless we adjust mem-fraction-static to 0.5 or below.

Related resources

No response