About save checkpoint

[c-junhao]

Thank you for your interesting work on PA-MoE. I could not find explicit checkpoint-saving logic in the released training script, so I would really appreciate it if you could clarify how the reported results were obtained. Thank you very much for your time and help.

[tinyzqh]

Regarding your question about not seeing explicit checkpoint-saving logic in the released training script: in this release, we primarily provide the code for PA-MoE’s forward network architecture. The results reported under the setting where QoE remains unchanged (steady-state QoE) were obtained by replacing the forward inference process in the Pensieve framework with the forward process provided in our code; under the same QoE constraint, you can reproduce PA-MoE’s corresponding results.

If you are interested in the non–steady-state QoE setting (i.e., QoE varies over time), then the full implementation needed for training and reproduction is included in the source code, covering the required components and workflow.