fix check params

[satpalsr]

Check params was removed for Bloom.
Not evident why.

[jeffra]

Good catch, this was accidentally added. Thank you @satpalsr

[jeffra]

Removing this return seems to cause failures on some of our unit tests, need to investigate further before merging.

[lekurile]

@satpalsr Looks like this check was removed in GH-2083 and was likely done due to certain models having mixed dtypes for various parameters. I think this likely warrants a deeper discussion about what purpose __check_params() serves in the runtime engine and if we should look to remove it altogether.

In the meantime, I think we can close this PR since simply removing the return in the function will result in errors in the unit tests. E.g. TestSaveTensorClone.test_save_tensor_clone[False-2]:

ValueError: torch.float32 is enabled but the following parameters have dtype that is not torch.float32: [('linears.0.weight', torch.float16), ('linears.0.bias', torch.float16), ('linears.1.weight', torch.float16), ('linears.1.bias', torch.float16), ('linears.2.weight', torch.float16), ('linears.2.bias', torch.float16), ('linears.3.weight', torch.float16), ('linears.3.bias', torch.float16)]