fix sd2 switching

[light-and-ray]

Closes: #13763
I start webui with not sd 2.1 model, then I try to load sd 2.1 checkpoint, and I get "NotImplementedError"
If I've started webui with sd2 model in --ckpt flag, it works, even if I switch to other and back to sd2

This bug appeared in October, now I think it's connected with device = devices.cpu in is_using_v_parameterization_for_sd2 inside this commit d04e3e9#diff-b710a9b8e9fbcc5bc5a014f938c9c74564c1dcfc86929f0dc9ff643ba3fe7873R30

NotImplementedError: No operator found for memory_efficient_attention_forward with inputs:
query : shape=(1, 64, 5, 64) (torch.float32)
key : shape=(1, 64, 5, 64) (torch.float32)
value : shape=(1, 64, 5, 64) (torch.float32)
attn_bias : <class 'NoneType'>
p : 0.0

decoderF is not supported because:
device=cpu (supported: {'cuda'})
attn_bias type is <class 'NoneType'>

flshattF@v2.3.6 is not supported because:
device=cpu (supported: {'cuda'})
dtype=torch.float32 (supported: {torch.float16, torch.bfloat16})

tritonflashattF is not supported because:
device=cpu (supported: {'cuda'})
dtype=torch.float32 (supported: {torch.float16, torch.bfloat16})
operator wasn't built - see python -m xformers.info for more info
triton is not available
Only work on pre-MLIR triton for now

cutlassF is not supported because:
device=cpu (supported: {'cuda'})

smallkF is not supported because:
max(query.shape[-1] != value.shape[-1]) > 32
device=cpu (supported: {'cuda'})
unsupported embed per head: 64

After this patch the bug is gone for me

Checklist:

• I have read contributing wiki page
• I have performed a self-review of my own code
• My code follows the style guidelines
• My code passes tests

[AUTOMATIC1111]

The reason I didn't want to run the calculation on GPU, is that it takes substantially longer to transfer the model to GPU than to just do one inference on CPU. I merged the other one.

[light-and-ray]

I merged the other one.

Which? The hcl's solves the other issue

[light-and-ray]

The reason I didn't want to run the calculation on GPU

But it doesn't work, at least not for everyone. Read the description