You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I'm uncertain whether it's feasible to bypass Ray when serving on a single machine with multiple GPUs. Ray introduces additional maintenance costs in this use case.
The text was updated successfully, but these errors were encountered:
At least ray is needed if you want to use tensor parallel with multiple GPUS, since each Worker instance should exist in a single process but not thread. However, we can just replace ray with multiprocess in this regard.
I haven't seen the other reason why we need ray in the code, maybe there are something, for example, memory issue, object sharing or some other stuff.
I'm uncertain whether it's feasible to bypass Ray when serving on a single machine with multiple GPUs. Ray introduces additional maintenance costs in this use case.
The text was updated successfully, but these errors were encountered: