Unable to use text_encoder via PyTorch
#3
by
v12h
- opened
CyberRealistic Pony v12.5
I'm able to generate beautiful images out of the box with this model. But I run into issues when using the text_encoder at the pytorch level.
There appears to be NaN values in text_encoder 1.
NaN detection:
Logs of NaN detection:
text_encoder outputs tensor of NaN's:
Summary:
- I'm extending the 77 token limit by chunking the tokens and encoding per chunk.
- When I encode per chunk I get a tensor that is the correct shape, but filled with NaN's.
- I've replaced the encoder with the default one and the default one "works", but the output is bad.
- I'm using this model from: CyberrealisticPony from CivIt
via:
- text_encoder2 works as expected.
- I have tried using fp32, same result.
- I have double/triple/quadruple checked that my download is not corrupt.
I could be conceptually missing something. Maybe if text_encoder 1 is modified in a way that doesn't let me use it like the default text_encoder for the SDXL pipeline?
CORRUPTED: text_model.encoder.layers.11.self_attn.k_proj.weight has 589824 NaN values
Are NaN values common in models trained like CyberRealistic Pony v12.5?