ImportError: cannot import name 'randn_tensor' from 'diffusers.utils'

[Zerg-Overmind]

Describe the bug

In the newest version, whenever you run: from diffusers.utils import randn_tensor
you will get the error: importError: cannot import name 'randn_tensor' from 'diffusers.utils'
But you can fix this by adding "from .torch_utils import randn_tensor" to your pythonxx/site-packages/diffusers/utils/init.py" after installing the diffusers....

Reproduction

run "from diffusers.utils import randn_tensor" after installing the latest version

Logs

No response

System Info

diffusers 0.21.0

Who can help?

No response

[DN6]

Hi @Zerg-Overmind. We moved randn tensors out of utils to enable lazy imports. You can now access it via

from diffusers.utils.torch_utils import randn_tensor

[sayakpaul]

@DN6 seems to me we can close this issue no?

[johnowhitaker]

@DN6 @sayakpaul FYI this breaks a few things - for e.g. I just tried to use TRL for something and get an error thanks to this:

----> [4](vscode-notebook-cell://ssh-remote%2Brig2/home/jonathan/kaggle_llm/sft_letter.ipynb#W0sdnNjb2RlLXJlbW90ZQ%3D%3D?line=3) from trl import SFTTrainer, DataCollatorForCompletionOnlyLM
      [5](vscode-notebook-cell://ssh-remote%2Brig2/home/jonathan/kaggle_llm/sft_letter.ipynb#W0sdnNjb2RlLXJlbW90ZQ%3D%3D?line=4) from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, TrainingArguments

File [~/miniconda3/envs/cuda11-7/lib/python3.8/site-packages/trl/__init__.py:7](https://vscode-remote+ssh-002dremote-002brig2.vscode-resource.vscode-cdn.net/home/jonathan/kaggle_llm/~/miniconda3/envs/cuda11-7/lib/python3.8/site-packages/trl/__init__.py:7)
      5 from .core import set_seed
      6 from .environment import TextEnvironment, TextHistory
----> 7 from .extras import BestOfNSampler
      8 from .import_utils import is_diffusers_available, is_peft_available
      9 from .models import (
     10     AutoModelForCausalLMWithValueHead,
     11     AutoModelForSeq2SeqLMWithValueHead,
     12     PreTrainedModelWrapper,
     13     create_reference_model,
     14 )

File [~/miniconda3/envs/cuda11-7/lib/python3.8/site-packages/trl/extras/__init__.py:16](https://vscode-remote+ssh-002dremote-002brig2.vscode-resource.vscode-cdn.net/home/jonathan/kaggle_llm/~/miniconda3/envs/cuda11-7/lib/python3.8/site-packages/trl/extras/__init__.py:16)
      1 # flake8: noqa
      2 
      3 # Copyright 2022 The HuggingFace Team. All rights reserved.
   (...)
     14 # See the License for the specific language governing permissions and
     15 # limitations under the License.
---> 16 from .best_of_n_sampler import BestOfNSampler

File [~/miniconda3/envs/cuda11-7/lib/python3.8/site-packages/trl/extras/best_of_n_sampler.py:7](https://vscode-remote+ssh-002dremote-002brig2.vscode-resource.vscode-cdn.net/home/jonathan/kaggle_llm/~/miniconda3/envs/cuda11-7/lib/python3.8/site-packages/trl/extras/best_of_n_sampler.py:7)
      4 from transformers import GenerationConfig, PreTrainedTokenizer, PreTrainedTokenizerFast
      6 from ..core import set_seed
----> 7 from ..models import SUPPORTED_ARCHITECTURES, PreTrainedModelWrapper
     10 class BestOfNSampler(object):
     11     def __init__(
     12         self,
     13         model: PreTrainedModelWrapper,
   (...)
     20         generation_config: Optional[GenerationConfig] = None,
     21     ) -> None:

File [~/miniconda3/envs/cuda11-7/lib/python3.8/site-packages/trl/models/__init__.py:29](https://vscode-remote+ssh-002dremote-002brig2.vscode-resource.vscode-cdn.net/home/jonathan/kaggle_llm/~/miniconda3/envs/cuda11-7/lib/python3.8/site-packages/trl/models/__init__.py:29)
     25 from ..import_utils import is_diffusers_available
     28 if is_diffusers_available():
---> 29     from .modeling_sd_base import (
     30         DDPOPipelineOutput,
     31         DDPOSchedulerOutput,
     32         DDPOStableDiffusionPipeline,
     33         DefaultDDPOStableDiffusionPipeline,
     34     )

File [~/miniconda3/envs/cuda11-7/lib/python3.8/site-packages/trl/models/modeling_sd_base.py:27](https://vscode-remote+ssh-002dremote-002brig2.vscode-resource.vscode-cdn.net/home/jonathan/kaggle_llm/~/miniconda3/envs/cuda11-7/lib/python3.8/site-packages/trl/models/modeling_sd_base.py:27)
     25 from diffusers.models.attention_processor import LoRAAttnProcessor
     26 from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion import rescale_noise_cfg
---> 27 from diffusers.utils import randn_tensor
     30 @dataclass
     31 class DDPOPipelineOutput(object):
     32     """
     33     Output class for the diffusers pipeline to be finetuned with the DDPO trainer
     34 
   (...)
     42 
     43     """

ImportError: cannot import name 'randn_tensor' from 'diffusers.utils'

Might be worth looking through at least the HF libs looking for from diffusers.utils import randn_tensor and adding a fix that checks the diffusers version or something?

[osanseviero]

Similar behaviour for me

from trl import SFTTrainer

File [~/miniconda3/envs/book/lib/python3.10/site-packages/trl/__init__.py:7](https://file+.vscode-resource.vscode-cdn.net/home/osanseviero/Desktop/workspace/genai/nbs/~/miniconda3/envs/book/lib/python3.10/site-packages/trl/__init__.py:7)
      5 from .core import set_seed
      6 from .environment import TextEnvironment, TextHistory
----> 7 from .extras import BestOfNSampler
      8 from .import_utils import is_diffusers_available, is_peft_available
      9 from .models import (
     10     AutoModelForCausalLMWithValueHead,
     11     AutoModelForSeq2SeqLMWithValueHead,
     12     PreTrainedModelWrapper,
     13     create_reference_model,
     14 )

File [~/miniconda3/envs/book/lib/python3.10/site-packages/trl/extras/__init__.py:16](https://file+.vscode-resource.vscode-cdn.net/home/osanseviero/Desktop/workspace/genai/nbs/~/miniconda3/envs/book/lib/python3.10/site-packages/trl/extras/__init__.py:16)
      1 # flake8: noqa
      2 
      3 # Copyright 2022 The HuggingFace Team. All rights reserved.
   (...)
...
   (...)
     42 
     43     """

ImportError: cannot import name 'randn_tensor' from 'diffusers.utils' (/home/osanseviero/miniconda3/envs/book/lib/python3.10/site-packages/diffusers/utils/__init__.py)

cc @lvwerra as well

[dage0127]

pip install diffusers==0.18.0 works fine.