Alexzyx

/

lora-trained-xl-colab

like 0

When loading the state dict of this file, one can see that the weights are broken.

E.g. doing the following:

from safetensors.torch import load_file

sd = load_file("./pytorch_lora_weights.safetensors")
print(sd.values)

shows:

...
        [ 0.4089, -0.4093,  0.0417,  ...,  0.3376, -0.0011, -0.3596]]), tensor([[nan, nan, nan, nan],
        [nan, nan, nan, nan],
        [nan, nan, nan, nan],
        ...,
        [nan, nan, nan, nan],
        [nan, nan, nan, nan],
        [nan, nan, nan, nan]]), tensor([[nan, nan, nan,  ..., nan, nan, nan],
        [nan, nan, nan,  ..., nan, nan, nan],
        [nan, nan, nan,  ..., nan, nan, nan],
        [nan, nan, nan,  ..., nan, nan, nan]]), tensor([[nan, nan, nan, nan],
        [nan, nan, nan, nan],
        [nan, nan, nan, nan],
        ...,
        [nan, nan, nan, nan],
        [nan, nan, nan, nan],
        [nan, nan, nan, nan]]), tensor([[nan, nan, nan,  ..., nan, nan, nan],
        [nan, nan, nan,  ..., nan, nan, nan],
        [nan, nan, nan,  ..., nan, nan, nan],
        [nan, nan, nan,  ..., nan, nan, nan]]), tensor([[nan, nan, nan, nan],
        [nan, nan, nan, nan],
        [nan, nan, nan, nan],
        ...,
        [nan, nan, nan, nan],
        [nan, nan, nan, nan],
        [nan, nan, nan, nan]]), tensor([[ 0.3282, -0.1082,  0.1383,  ..., -0.2528, -0.2138, -0.3022],
        [ 0.2439,  0.0952,  0.1135,  ...,  0.2367, -0.1400,  0.0796],
        [-0.1357,  0.2351, -0.4367,  ..., -0.5447, -0.3476, -0.0175],
        [-0.2212,  0.2084, -0.5737,  ..., -0.0218, -0.1498,  0.6869]]), tensor([[nan, nan, nan, nan],
        [nan, nan, nan, nan],
        [nan, nan, nan, nan],
        ...,
        [nan, nan, nan, nan],
        [nan, nan, nan, nan],
        [nan, nan, nan, nan]])])

E.g. for the key: sd['unet.up_blocks.1.attentions.2.transformer_blocks.1.attn2.processor.to_v_lora.up.weight']