Weekly SubMIT project team meeting

America/New_York
ZOOM

ZOOM

David Walter (Massachusetts Institute of Technology), Mariarosaria D'Alfonso (Massachusetts Institute of Technology)
    • 10:00 10:20
      Discussion 20m
      Speakers: Alexander Avdoshkin (MIT), David Walter (Massachusetts Institute of Technology), Luca Lavezzo (MIT), Mariarosaria D'Alfonso (Massachusetts Institute of Technology), Matthew Heine (Massachusetts Institute of Technology)

      present: Maria, David, Matt, CP

       

      Report of the week 34:

      • login nodes better behaving: not saturated by users, so the modifications from David were useful

       

      David:

      • updated Alarms of the certificate expiring (David)
        • ceph - latest fix vulnerability  (David)
        • needed update various keys (took ~ 1 day)
      • alarms updated
      • cleanup of various partitions 

       

      Matt:

      • jupiterHub: mathematica now on all login nodes available on drop down 

      • will commit the configuration on GitHub, here https://github.com/mit-submit/submit-admin/tree/main/servers

       

      Maria:

      • minor improvements in the user/admin guide

       

      Long term:

      • more modern GPU welcome: i.e. to host more complex model on ollama/VLLM resources 
        from Jason/Archi: "I got qwen 3.8 27b running on submit76. With all four v100s it's roughly 50 tok/s decode, 500/s prefill for a single request. Parallel requests it's about 100 tok/s decode, 1000/s prefill. Does submit have any other GPUs that are more modern? the tough thing with v100s is many kernels aren't supported "