> And we are using the older Tesla T4.

That's an undemanding and well-debugged chip by this point (6 years ago!). So you aren't experiencing any of the pain people using A100s or H100s (never mind people who have to stand up clusters with B100s soon) are going through now.