How are beams separated in Uplink?

I think CPE.

Where does it says that this is used in CPE and not in gNodeB?
Yes, it is used in CPE.

But mmWave gNB almost no support UL MU-MIMO?

I think at Huawei UL MU-MIMO is possible only for low frequency TDD cells with 32T32R or 64T64R or higher.

In case of UL, RBs will be allocated to the paired UEs by the eNB using the DCI format 0_1.
In MU MIMO, the same set of RBs will be allocated to every paired UEs.
So, I think there is a DMRS sequence in DCI format 0_1 which will be used by every UE when sending the UL data.
From this DMRS sequence, the eNB will decode the PUSCH for every UE.

Don’t know whether I am correct or wrong…

I have to accept that the uplink massive Mimo “works somehow” in real life but after reading all these posts I have serious doubts if we really understand what is actually happening. I have been through loads of videos, posts, forums, websites, books and no one is clear about it.
The best way is to model it but this could drag us into modelling problems.