Using negative prompt brakes SDXL txt2img

[FinecoFinit]

Steps to reproduce:

1. Set SDXL txt2img, use Euler A, prompt with negative prompt
2. Make image
3. Error happens:

[2023-12-14 19:14:50,365] WARNING: onnx-web worker: directml MainThread onnx_web.diffusers.load: error while enabling auto attention slicing: 'ORTStableDiffusionXLPipeline' object has no attribute 'enable_attention_slicing'
[2023-12-14 19:14:52,756] ERROR: onnx-web worker: directml MainThread onnx_web.chain.pipeline: error while running stage pipeline, 1 retries left
Traceback (most recent call last):
  File "D:\@home\fifi\Documents\onnx-web\api\onnx_web\chain\pipeline.py", line 227, in __call__
    stage_result = stage_pipe.run(
  File "D:\@home\fifi\Documents\onnx-web\api\onnx_web\chain\source_txt2img.py", line 144, in run
    result = pipe(
  File "D:\@home\fifi\Documents\onnx-web\api\onnx_env\lib\site-packages\optimum\onnxruntime\modeling_diffusion.py", line 590, in __call__
    return StableDiffusionXLPipelineMixin.__call__(self, *args, **kwargs)
  File "D:\@home\fifi\Documents\onnx-web\api\onnx_env\lib\site-packages\optimum\pipelines\diffusers\pipeline_stable_diffusion_xl.py", line 434, in __call__
    prompt_embeds = np.concatenate((negative_prompt_embeds, prompt_embeds), axis=0)
  File "<__array_function__ internals>", line 180, in concatenate
ValueError: all the input arrays must have same number of dimensions, but the array at index 0 has 2 dimension(s) and the array at index 1 has 3 dimension(s)

You will get same error even after removing negative prompt, restarting API helps to resolve error

[ssube]

I tested this with Euler A and a few other schedulers, but I'm not able to reproduce it:

{
  "input_size": {
    "width": 1024, 
    "height": 1024
  }, 
  "outputs": "txt2img_2045114381_b759c25db9a9fba9815472c4be2bc8e8eb29c2a9f0241fd66f4e30c9e049798b_1702606569_0.png", 
  "params": {
    "model": "diffusion-sdxl-cotton-candy-3d", 
    "pipeline": "txt2img-sdxl", 
    "scheduler": "euler-a", 
    "prompt": "a cute puppy dog", 
    "negative_prompt": "cat, cows, sheep, fish, zebras", 
    "cfg": 4.0, 
    "seed": 2045114381, 
    "steps": 50, 
    "eta": 0.0, 
    "batch": 1, 
    "control": "", 
    "input_prompt": "a cute puppy dog", 
    "input_negative_prompt": "cat, cows, sheep, fish, zebras", 
    "loopback": 0, 
    "tiled_vae": false, 
    "unet_overlap": 0.75, 
    "unet_tile": 1024, 
    "vae_overlap": 0.25, 
    "vae_tile": 512, 
    "denoise": 3
  }, 
  "highres": {
    "enabled": false, "iterations": 1, "method": "bilinear", "scale": 1, "steps": 1, "strength": 0.1
  }, 
  "upscale": {
    "upscale_model": "upscaling-swinir-real-x2", "correction_model": "correction-codeformer", "denoise": 0.5, "faces": false, "face_outscale": 1, "face_strength": 0.5, "format": "onnx", "outscale": 1, "pre_pad": 0, "scale": 1, "tile_pad": 10, "upscale_order": "correction-first"
  }, 
  "size": {
    "width": 1024, "height": 1024
  }
}

I do recall pushing a fix for a similar issue to optimum a while back (huggingface/optimum#1339), so it's possible something is out of date, possibly. I've updated the base requirements with the latest versions of diffusers and optimum, which are working well for me. Can you try updating to those versions and see if this is still happening?

[FinecoFinit]

@ssube

The conflict is caused by:
    The user requested huggingface-hub==0.16.4
    accelerate 0.25.0 depends on huggingface-hub
    controlnet-aux 0.0.2 depends on huggingface-hub
    datasets 2.15.0 depends on huggingface-hub>=0.18.0

Went smooth with huggingface-hub==0.19.4, problem fixed

[ssube]

I thought I had included huggingface-hub in the pip upgrade, but I'll bump that version as well before the release.