Change in checkpoint sizes?

Since about a week, I notice the size of my checkpoints to differ significantly compared to previous runs. I am training the same model on the same data, only on another GPU (A100 80GB vs A100 40GB). Previously, the optimizer states were way larger than the model states, now it is inversed and the inference results are nonsense, in contrast to the results of previous runs. I have to clue what might cause this, is this a common problem?

ds_report output on the new GPU (same as on the old GPU):

JIT compiled ops requires ninja
ninja .................. [OKAY]
--------------------------------------------------
op name ................ installed .. compatible
--------------------------------------------------
async_io ............... [NO] ....... [OKAY]
fused_adam ............. [NO] ....... [OKAY]
cpu_adam ............... [NO] ....... [OKAY]
cpu_adagrad ............ [NO] ....... [OKAY]
cpu_lion ............... [NO] ....... [OKAY]
 [WARNING]  Please specify the CUTLASS repo directory as environment variable $CUTLASS_PATH
evoformer_attn ......... [NO] ....... [NO]
fused_lamb ............. [NO] ....... [OKAY]
fused_lion ............. [NO] ....... [OKAY]
inference_core_ops ..... [NO] ....... [OKAY]
cutlass_ops ............ [NO] ....... [OKAY]
transformer_inference .. [NO] ....... [OKAY]
quantizer .............. [NO] ....... [OKAY]
ragged_device_ops ...... [NO] ....... [OKAY]
ragged_ops ............. [NO] ....... [OKAY]
random_ltd ............. [NO] ....... [OKAY]
 [WARNING]  sparse_attn requires a torch version >= 1.5 and < 2.0 but detected 2.2
 [WARNING]  using untested triton version (2.2.0), only 1.0.0 is known to be compatible
sparse_attn ............ [NO] ....... [NO]
spatial_inference ...... [NO] ....... [OKAY]
transformer ............ [NO] ....... [OKAY]
stochastic_transformer . [NO] ....... [OKAY]

Furthermore, the output folders do not contain a config.json anymore, which seems to be renamed to adapter_config.json

Start by comparing the checkpoint contents and the presence or absence of config.json versus adapter_config.json across the two GPU runs, then verify the PyTorch and DeepSpeed environments against the reported ds_report output. Done means identifying a reproducible cause for the changed optimizer/model state sizes and incorrect inference results, with a confirmed explanation or fix.