Model stopped to do inference after some steps

I'm using SSD MobileNet v2 Open Images v4 (only works/train in tf1) then I created a conda env with tf1 to train

Setup to train

• Tensorflow-gpu 1.15
• Cuda 10.0
• Cudnn 7.6.0
• Python 3.7.9

To make inferences a use other conda env, with tf2

Setup to inference

• Tensorflow 2.4.0
• Cuda 11.1
• Cudnn 8.0.4
• Python 3.8.5

I have two env because some models only works in tf1 and other only in tf2, I use both to trains their respectively models, but I only use tf2 env to make inference, always worked in that way

The model was trained until 300k steps, I exported and model worked well, after some days, I got back to train and with 700k, I exported again and the model practically dont make any inference, in tensorboard the model was improving normally (total loss going down), but have a weird gap in loss graphics

Anyone have some suggestions or same issue?

I tried export the model version with 300k steps again to check if could be some setup problem but the model worked