Loading a converted model fails to recognize GPU

Help scvi-tools
1

I have an scvi model created with scvi version ‘0.14.6’ using GPUs. I am now working in an updated environment (also using GPUs) with scvi version ‘0.19.0’. I did the following to convert the old model into the new format and to try and load the model but got the following error.

scvi.model.SCVI.convert_legacy_save(working_path + 
                                    "/scvi_model_Immune_batch10x/scvi_model_Immune_batch10x/",
                                    working_path + 
                                    "/new_scvi_model_Immune_batch10x/")

model = scvi.model.SCVI.load(working_path + 
                             "/new_scvi_model_Immune_batch10x/", 
                             adata, 
                             use_gpu=True)

INFO     File /home/jupyter/new_scvi_model_Immune_batch10x/model.pt already downloaded                             
WARNING:jax._src.lib.xla_bridge:No GPU/TPU found, falling back to CPU. (Set TF_CPP_MIN_LOG_LEVEL=0 and rerun for more info.)`

I already confirmed that scvi, pytorch, etc. were properly installed, that I have the expected CUDA version, and I confirmed that the converted model has the expected format. Is there any way I can make sure that loading the converted model will recognize the GPU?

2

The message you’re seeing there is specific to jax. It’s ok to ignore this warning and your model should work just fine on GPU (as it’s using pytorch)