[SDXL] Compel can't batch prompt (mat1 and mat2 shapes cannot be multiplied)

[Fqlox]

I wanted to make an inference with multiple prompt at one time and got:

mat1 and mat2 shapes cannot be multiplied (2x4096 and 2816x1280)

Here is a basic example:

pipe = StableDiffusion ....
compel = Compel(
  tokenizer=[pipe.tokenizer, pipe.tokenizer_2] ,
  text_encoder=[pipe.text_encoder, pipe.text_encoder_2],
  returned_embeddings_type=ReturnedEmbeddingsType.PENULTIMATE_HIDDEN_STATES_NON_NORMALIZED,
  requires_pooled=[False, True]
)

prompts_str = [
    "a Man wearing (sunglasses)1.2, close up, realistic, 4k",
    "a dog in a forest, long distance shot, photo, photography, highly detailed",
]

prompts = [compel(x) for x in prompts_str]

conditioning = torch.cat([e[0] for e in prompts], dim=1) 
pooled = torch.cat([e[1] for e in prompts], dim =1 ) 

image = pipe(
prompt_embeds=conditioning, 
pooled_prompt_embeds=pooled,
image_batch_size=2,
prompt_batch_size=2,
...
)

I tried to use list instead of torch.cat, I tried to pass on pooled prompt

Thanks in advance

[madriss]

Have you found a solution ?

I tried to concat with dim=0 and seemed to get it to work although with strange resutls.