hi, thanks for your repo,
after read your code for traing with mmdetection-related framework, seems that you write those three fundamental code:

optimizer = HybridAdam(model.parameters())
model = zero_model_wrapper(model, zero_stage=2)
optimizer = zero_optim_wrapper(
    model, optimizer, optim_config=ZERO_OPTIM_CFG)

if this, we lose the general unitily of training of those specified frameworks, eg: those hooks.
maybe we can add those three lines of code to our framework inside, should that work?