Doesn't work boys - we'll get 'em next time. FIX INSIDE

#10
by mushroomfleet - opened

Also reported by people here:
https://github.com/comfyanonymous/ComfyUI/issues/5422

UPDOOT:

if you update Torch to 2.5.1, Torchvision to 0.20.1, Torchaudio to 2.5.1 and also Transformers to 4.46.1-py3-none-any.whl

  • it works
  • please include this in the model page instructions somewhere?

Error:

RuntimeError: Error(s) in loading state_dict for OpenAISignatureMMDITWrapper:
size mismatch for joint_blocks.0.x_block.adaLN_modulation.1.weight: copying a param with shape torch.Size([13824, 1536]) from checkpoint, the shape in current model is torch.Size([9216, 1536]).
size mismatch for joint_blocks.0.x_block.adaLN_modulation.1.bias: copying a param with shape torch.Size([13824]) from checkpoint, the shape in current model is torch.Size([9216]).
size mismatch for joint_blocks.1.x_block.adaLN_modulation.1.weight: copying a param with shape torch.Size([13824, 1536]) from checkpoint, the shape in current model is torch.Size([9216, 1536]).
size mismatch for joint_blocks.1.x_block.adaLN_modulation.1.bias: copying a param with shape torch.Size([13824]) from checkpoint, the shape in current model is torch.Size([9216]).
size mismatch for joint_blocks.2.x_block.adaLN_modulation.1.weight: copying a param with shape torch.Size([13824, 1536]) from checkpoint, the shape in current model is torch.Size([9216, 1536]).
size mismatch for joint_blocks.2.x_block.adaLN_modulation.1.bias: copying a param with shape torch.Size([13824]) from checkpoint, the shape in current model is torch.Size([9216]).
size mismatch for joint_blocks.3.x_block.adaLN_modulation.1.weight: copying a param with shape torch.Size([13824, 1536]) from checkpoint, the shape in current model is torch.Size([9216, 1536]).
size mismatch for joint_blocks.3.x_block.adaLN_modulation.1.bias: copying a param with shape torch.Size([13824]) from checkpoint, the shape in current model is torch.Size([9216]).
size mismatch for joint_blocks.4.x_block.adaLN_modulation.1.weight: copying a param with shape torch.Size([13824, 1536]) from checkpoint, the shape in current model is torch.Size([9216, 1536]).
size mismatch for joint_blocks.4.x_block.adaLN_modulation.1.bias: copying a param with shape torch.Size([13824]) from checkpoint, the shape in current model is torch.Size([9216]).
size mismatch for joint_blocks.5.x_block.adaLN_modulation.1.weight: copying a param with shape torch.Size([13824, 1536]) from checkpoint, the shape in current model is torch.Size([9216, 1536]).
size mismatch for joint_blocks.5.x_block.adaLN_modulation.1.bias: copying a param with shape torch.Size([13824]) from checkpoint, the shape in current model is torch.Size([9216]).
size mismatch for joint_blocks.6.x_block.adaLN_modulation.1.weight: copying a param with shape torch.Size([13824, 1536]) from checkpoint, the shape in current model is torch.Size([9216, 1536]).
size mismatch for joint_blocks.6.x_block.adaLN_modulation.1.bias: copying a param with shape torch.Size([13824]) from checkpoint, the shape in current model is torch.Size([9216]).
size mismatch for joint_blocks.7.x_block.adaLN_modulation.1.weight: copying a param with shape torch.Size([13824, 1536]) from checkpoint, the shape in current model is torch.Size([9216, 1536]).
size mismatch for joint_blocks.7.x_block.adaLN_modulation.1.bias: copying a param with shape torch.Size([13824]) from checkpoint, the shape in current model is torch.Size([9216]).
size mismatch for joint_blocks.8.x_block.adaLN_modulation.1.weight: copying a param with shape torch.Size([13824, 1536]) from checkpoint, the shape in current model is torch.Size([9216, 1536]).
size mismatch for joint_blocks.8.x_block.adaLN_modulation.1.bias: copying a param with shape torch.Size([13824]) from checkpoint, the shape in current model is torch.Size([9216]).
size mismatch for joint_blocks.9.x_block.adaLN_modulation.1.weight: copying a param with shape torch.Size([13824, 1536]) from checkpoint, the shape in current model is torch.Size([9216, 1536]).
size mismatch for joint_blocks.9.x_block.adaLN_modulation.1.bias: copying a param with shape torch.Size([13824]) from checkpoint, the shape in current model is torch.Size([9216]).
size mismatch for joint_blocks.10.x_block.adaLN_modulation.1.weight: copying a param with shape torch.Size([13824, 1536]) from checkpoint, the shape in current model is torch.Size([9216, 1536]).
size mismatch for joint_blocks.10.x_block.adaLN_modulation.1.bias: copying a param with shape torch.Size([13824]) from checkpoint, the shape in current model is torch.Size([9216]).
size mismatch for joint_blocks.11.x_block.adaLN_modulation.1.weight: copying a param with shape torch.Size([13824, 1536]) from checkpoint, the shape in current model is torch.Size([9216, 1536]).
size mismatch for joint_blocks.11.x_block.adaLN_modulation.1.bias: copying a param with shape torch.Size([13824]) from checkpoint, the shape in current model is torch.Size([9216]).
size mismatch for joint_blocks.12.x_block.adaLN_modulation.1.weight: copying a param with shape torch.Size([13824, 1536]) from checkpoint, the shape in current model is torch.Size([9216, 1536]).
size mismatch for joint_blocks.12.x_block.adaLN_modulation.1.bias: copying a param with shape torch.Size([13824]) from checkpoint, the shape in current model is torch.Size([9216]).

Prompt executed in 1.06 seconds
No pending upload

mushroomfleet changed discussion title from Doesn't work boys - we'll get 'em next time. to Doesn't work boys - we'll get 'em next time. FIX INSIDE

do we need to update this on the requirements.txt?

Working fine with torch 2.4.1 in Diffusers, perhaps the ComfyUI folks need to looks why their code isn't working.
Also in my experience Torch 2.5.1x is very broke, uses way more memory than 2.4.1.

try pip install -U diffusers

Sign up or log in to comment