cuda.OutOfMemoryError: CUDA out of memory

[junzhoupro]

Dear Author, thanks for your work!
I'm running the training on my computer and had out of memory error.
I'm using 4090

Training with error:
torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 50.00 MiB (GPU 0; 23.64 GiB total capacity; 22.20 GiB already allocated; 70.75 MiB free; 22.38 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

I'm using 4090, my configs: anything I can do to train on my computer ?

batch_size = 1 #16
logger_freq = 1000
learning_rate = 1e-5
sd_locked = True #False
only_mid_control = True #False
n_gpus = 1
accumulate_grad_batches=1

[XavierCHEN34]

You could try "ddp_sharded", which requires smaller memories
trainer = pl.Trainer(gpus=1, strategy="ddp_sharded", precision=16, accelerator="gpu", callbacks=[logger], progress_bar_refresh_rate=1)

[PlanPersisitentPatient]

Dear Author, thanks for your work! I'm running the training on my computer and had out of memory error. I'm using 4090

Training with error: torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 50.00 MiB (GPU 0; 23.64 GiB total capacity; 22.20 GiB already allocated; 70.75 MiB free; 22.38 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

I'm using 4090, my configs: anything I can do to train on my computer ?

batch_size = 1 #16 logger_freq = 1000 learning_rate = 1e-5 sd_locked = True #False only_mid_control = True #False n_gpus = 1 accumulate_grad_batches=1

May I ask if you were able to train the model successfully? I am also using a graphics card with 24GB of VRAM and have tried the author's suggestions, but I still encounter a CUDA out of memory error.