Does it works with cpu too?

[andreae293]

Hi, has anyone ever tried to train with cpu?
i know it will be super slow but im tried for the fun of it

i currently disabled my gpu by setting this line in image_train_stable.py
torch.cuda.is_available = lambda : False

Traceback (most recent call last):
File "scripts\image_train_stable.py", line 157, in
main()
File "scripts\image_train_stable.py", line 85, in main
TrainLoop(
File "c:\users\andre\desktop\ml\glid-3\guided_diffusion\train_util.py", line 194, in run_loop
self.run_step(batch, cond)
File "c:\users\andre\desktop\ml\glid-3\guided_diffusion\train_util.py", line 208, in run_step
self.forward_backward(batch, cond)
File "c:\users\andre\desktop\ml\glid-3\guided_diffusion\train_util.py", line 236, in forward_backward
losses = compute_losses()
File "c:\users\andre\desktop\ml\glid-3\guided_diffusion\respace.py", line 96, in training_losses
return super().training_losses(self._wrap_model(model), *args, **kwargs)
File "c:\users\andre\desktop\ml\glid-3\guided_diffusion\gaussian_diffusion.py", line 1137, in training_losses
model_output = model(x_t, self._scale_timesteps(t), **model_kwargs)
File "c:\users\andre\desktop\ml\glid-3\guided_diffusion\respace.py", line 133, in call
return self.model(x, new_ts, **kwargs)
File "C:\Users\andre\anaconda3\envs\ldm\lib\site-packages\torch\nn\modules\module.py", line 1110, in _call_impl
return forward_call(*input, **kwargs)
File "c:\users\andre\desktop\ml\glid-3\guided_diffusion\unet.py", line 880, in forward
h = module(h, emb, context)
File "C:\Users\andre\anaconda3\envs\ldm\lib\site-packages\torch\nn\modules\module.py", line 1110, in _call_impl
return forward_call(*input, **kwargs)
File "c:\users\andre\desktop\ml\glid-3\guided_diffusion\unet.py", line 217, in forward
x = layer(x)
File "C:\Users\andre\anaconda3\envs\ldm\lib\site-packages\torch\nn\modules\module.py", line 1110, in _call_impl
return forward_call(*input, **kwargs)
File "C:\Users\andre\anaconda3\envs\ldm\lib\site-packages\torch\nn\modules\conv.py", line 447, in forward
return self._conv_forward(input, self.weight, self.bias)
File "C:\Users\andre\anaconda3\envs\ldm\lib\site-packages\torch\nn\modules\conv.py", line 443, in _conv_forward
return F.conv2d(input, weight, bias, self.stride,
RuntimeError: "slow_conv2d_cpu" not implemented for 'Half'

sorry for bothering with useless question but am i doing something wrong?
thanks

edit:
nevermind
i removed both .half() from the image_train_stable.py and deleted --use_fp16 from the training arguments

this way i was able to train on cpu

[timotheecour4]

@andreae293 can you please provide more details?
I tried:

export CUDA_VISIBLE_DEVICES=-1
MODEL_FLAGS="--actual_image_size 512 --lr_warmup_steps 10000 --ema_rate 0.9999 --attention_resolutions 64,32,16 --class_cond False --diffusion_steps 1000 --image_size 64 --learn_sigma False --noise_schedule linear --num_channels 320 --num_heads 8 --num_res_blocks 2 --resblock_updown False --use_fp16 False --use_scale_shift_norm False "
TRAIN_FLAGS="--lr 5e-5 --batch_size 32 --log_interval 10 --save_interval 5000 --kl_model kl.pt --resume_checkpoint diffusion.pt"
export OPENAI_LOGDIR=./logs/
python scripts/image_train_stable.py --data_dir /path/to/image-and-text-files $MODEL_FLAGS $TRAIN_FLAGS

where I change the default from README from --use_fp16 True to --use_fp16 False (and IIUC no need to remove .half() from the image_train_stable.py with this flag), but it gives:

RuntimeError: CUDA out of memory. Tried to allocate 4.00 GiB (GPU 0; 15.78 GiB total capacity; 10.52 GiB already allocated; 3.86 GiB free; 10.77 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

ie, still seems to use gpu instead of cpu

[ghost]

Stable Diffusion requires CUDA to run the AI, as it is the language for communicating with the GPU and preforming the necessary calculations. Using with the CPU would require a complete rewrite or virtualization which requires more RAM and money than it would take to go and buy a supported CUDA GPU. Although if anyone who is reading this is willing, would it be possible to utilize a TPU from Kaggle or Google Collab instead? I feel like it might be more efficient than a GPU or CPU, as it is meant for processing Tensors directly.

[andreae293]

@timotheecour4 sorry for the late response
if you dont have enough RAM you have to dedicate a lot of GB to the virtual memory (paging file or swap memory)
i dont know the minimum required by this repo,but i did dedicate 100 GB of virtual memory
also if you are trying to finetune i suggest you to look for dreambooth for stable diffusion

@TheRealUnBot stable diffusion does not necessarily requires CUDA-supporting hardware to run
since its based on pytorch you can run just fine in CPU with the downside of being x50 slower