non-docker AF2 on HPC cannot use GPU xla_bridge.py:257] No GPU/TPU found, falling back to CPU. (Set TF_CPP_MIN_LOG_LEVEL=0 and rerun for more info.)

Hi I am planning to use non-docker version on the HPC with GPU
I use https://github.com/amorehead/alphafold_non_docker

I found the `run_alphafold.sh` is nor working well so I checked this one https://sbgrid.org/wiki/examples/alphafold2
and seems working using the `run_alphafold.py` script.

Here is what I have done:

As the website indicating, I use:
`
TF_FORCE_UNIFIED_MEMORY=1
`
`
XLA_PYTHON_CLIENT_MEM_FRACTION=0.5
`
`
XLA_PYTHON_CLIENT_ALLOCATOR=platform
`

and then:

`nvidia-smi`

In order to use all GPUs, I use:
`
export CUDA_VISIBLE_DEVICES=0,1,2
`

here is a screenshot of my GPU HPC:
`htop`

```
python run_alphafold.py \
--data_dir=/export/home2/ql9f/download/ \
--output_dir=/export/III-data/waters/ql9f/RESC6/T1083outdir \
--fasta_paths=/export/III-data/waters/ql9f/RESC6/T1083.fasta \
--max_template_date=2020-05-14 \
--db_preset=reduced_dbs \
--model_preset=monomer \
--uniref90_database_path=/export/home2/ql9f/download/uniref90/uniref90.fasta \
--mgnify_database_path=/export/home2/ql9f/download/mgnify/mgy_clusters_2018_12.fa \
--template_mmcif_dir=/export/home2/ql9f/download/pdb_mmcif/mmcif_files \
--obsolete_pdbs_path=/export/home2/ql9f/download/pdb_mmcif/obsolete.dat \
--small_bfd_database_path=/export/home2/ql9f/download/small_bfd/bfd-first_non_consensus_sequences.fasta \
--pdb70_database_path=/export/home2/ql9f/download/pdb70/pdb70 \
--use_gpu_relax=True
```

The program is running but it did not use GPU as I expected, and the RAM only uses like 10GB which I think is too low.
So my question is: how can I use all the GPU and RAM and possibly all cores/threads on the HPC with GPU and make my prediction faster?
Why does the program CANNOT find GPU?

any suggestion would be great.
Qiushi