StableDiffusion3Pipeline crashes with AttributeError: 'NoneType' object has no attribute 'ndim' when using num_images_per_prompt with batched prompts

[FredyRivera-dev]

Describe the bug

When using StableDiffusion3Pipeline with batched prompts (prompt as a list) and passing the parameter num_images_per_prompt (even when set to 1), the pipeline crashes with:

AttributeError: 'NoneType' object has no attribute 'ndim'

This error does not occur:

• when num_images_per_prompt is omitted
• when using other pipelines (SDXL, SD1.x, FLUX, etc.)
• when running single-prompt (non-batch) inference

This appears to be specific to Stable Diffusion 3 and its internal prompt/embedding handling.

Environment

• diffusers: 0.36.0
• torch: 2.8
• python: 3.12
• device: CUDA
• OS: Linux
• Pipeline: StableDiffusion3Pipeline
• Not using: prompt_embeds, custom embeddings, or LoRAs

Steps to reproduce

Minimal reproduction pattern:

from diffusers import StableDiffusion3Pipeline
import torch

pipe = StableDiffusion3Pipeline.from_pretrained(
    "stabilityai/stable-diffusion-3-medium",
    torch_dtype=torch.float16
).to("cuda")

prompts = [
    "A photo of a cat",
    "A photo of a dog",
]

images = pipe(
    prompt=prompts,                # batched prompts
    height=1024,
    width=1024,
    num_inference_steps=30,
    num_images_per_prompt=1,       # <- causes crash
).images

Removing num_images_per_prompt makes the same code run successfully.

Full traceback

0%|          | 0/30 [00:00<?, ?it/s]

--------------------------------------------------------------------------------
ERROR: 'NoneType' object has no attribute 'ndim'
--------------------------------------------------------------------------------

ERROR: <class 'AttributeError'>
ERROR: Traceback (most recent call last):
  File "/usr/local/lib/python3.12/site-packages/aquilesimage/runtime/batch_inf.py", line 294, in batch_infer
    return self.pipeline.generate_batch(
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/site-packages/aquilesimage/runtime/requestscopedpipeline.py", line 789, in generate_batch
    result = local_pipe(
             ^^^^^^^^^^^
  File "/usr/local/lib/python3.12/site-packages/torch/utils/_contextlib.py", line 120, in decorate_context
    return func(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/site-packages/diffusers/pipelines/stable_diffusion_3/pipeline_stable_diffusion_3.py", line 1064, in __call__
    noise_pred = self.transformer(
                 ^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1773, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1784, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/site-packages/diffusers/models/transformers/transformer_sd3.py", line 327, in forward
    encoder_hidden_states, hidden_states = block(
    ^^^^^^
  File "/usr/local/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1773, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1784, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/site-packages/diffusers/models/attention.py", line 715, in forward
    attn_output2 = self.attn2(hidden_states=norm_hidden_states2, **joint_attention_kwargs)
                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1773, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1784, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/site-packages/diffusers/models/attention_processor.py", line 605, in forward
    return self.processor(
           ^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/site-packages/diffusers/models/attention_processor.py", line 1849, in __call__
    context_input_ndim = encoder_hidden_states.ndim
                        ^^^^^^^^^^^^^^^^^^^^^^^^^^
AttributeError: 'NoneType' object has no attribute 'ndim'

ERROR: {'height': 1024, 'width': 1024, 'num_inference_steps': 30, 'device': 'cuda', 'num_images_per_prompt': 1}

Expected behavior

• num_images_per_prompt should work with batched prompts, or
• the pipeline should raise a clear validation error if this combination is unsupported

At minimum, the pipeline should not pass encoder_hidden_states=None into the attention processor.

Additional notes

• This issue only reproduces with StableDiffusion3Pipeline
• Other pipelines handle the same parameters correctly
• As a workaround, expanding prompts manually and not passing num_images_per_prompt avoids the crash

[FredyRivera-dev]

Hi, I found the error. It turns out that when using this:

from diffusers import StableDiffusion3Pipeline
import torch

pipe = StableDiffusion3Pipeline.from_pretrained(
    "stabilityai/stable-diffusion-3-medium",
    torch_dtype=torch.float16
).to("cuda")

pipe.transformer.fuse_qkv_projections() # <- THIS 
print("fuse_qkv_projections enabled")

prompts = [
    "A photo of a cat",
    "A photo of a dog",
]

images = pipe(
    prompt=prompts,
    height=1024,
    width=1024,
    num_inference_steps=30,
    num_images_per_prompt=1,
).images

This error occurs

--------------------------------------------------------------------------------
ERROR: 'NoneType' object has no attribute 'ndim'
--------------------------------------------------------------------------------

And by removing pipe.transformer.fuse_qkv_projections() it no longer happens and the inference occurs correctly

So that's strange behavior

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.