invalid result in NVIDIA L4 (triton)

[humblemh]

Describe the bug
My fp16 model in triton server is working in nvidia a6000, t4, p4000.
but the same model is not worked in nvidia l4. (same hw, os)
the recogniton result is empty or wrong text.

To Reproduce
Steps to reproduce the behavior:

1. fp16 model loaded in triton server on nvidla l4

Expected behavior
recognition success

Desktop (please complete the following information):

• OS: Ubuntu 22.04