Modules fail for Dreambooth example

[paulaserna16]

I'm trying to run the dreambooth tutorial, and when executing the dreambooth.py it raises an error related to the libraries, for example:

[NeMo W 2024-10-08 09:12:02 megatron_lm_encoder_decoder_model:78] Megatron num_microbatches_calculator not found, using Apex version.
Traceback (most recent call last):
File "/opt/NeMo/clip/convert_external_clip_to_nemo.py", line 53, in
from nemo.collections.multimodal.models.vision_language_foundation.clip.megatron_clip_models import MegatronCLIPModel
File "/usr/local/lib/python3.10/dist-packages/nemo/collections/multimodal/models/vision_language_foundation/clip/megatron_clip_models.py", line 311, in
class SiglipMHAPoolingHead(TransformerLayer):
NameError: name 'TransformerLayer' is not defined

I don't think you specify which NeMo image is compatible with this example. Two months ago it was working with the image nvcr.io/nvidia/nemo:24.02, but now it gives the error I put above.
Please could you tell me which image works with this tutorial? or which versions of the libraries are needed?

Thank you so much.

[paulaserna16]

Update on the procedure I'm following:

• I use the Nemo framework 24.07 image
• Launch a container with the image (witha couple of volumes): docker run --runtime=nvidia --gpus all -it --rm
  -v /mnt/shared_demos/dreambooth/nemo:/opt/NeMo
  -v /mnt/shared_models/huggingface/cache/hub/models--runwayml--stable-diffusion-v1-5/snapshots/1d0c4ebf6ff58a5caecab40fa1406526bca4b5b9:/opt/models
  --shm-size=8g -p 8888:8888
  --ulimit memlock=-1 --ulimit stack=67108864
  nvcr.io/nvidia/nemo:24.05
• I install the main branch from the nemo repository in order to get the nemo toolkit installed: python -m pip install git+https://github.com/NVIDIA/NeMo.git@main#egg=nemo_toolkit[all]
• The one above installs the megatron library in /opt/megatron-lm. However, this is version 0.8; and if I go to the path where it's located, the contents of /megatron/core are not updated, for example, there's no folder called extensions.

Thus, it installs something that cannot be used for the dreambooth example.

There's also a section (Megatron-LM) where it is specified to run the following in roder to use it:

• git clone https://github.com/NVIDIA/Megatron-LM.git && \
• cd Megatron-LM && \
• git checkout $mcore_commit && \
• pip install .

This has the versions updated but installs a version of megatron incompatible with dreambooth files (0.10) and image, I believe.

Could you help me with this issue please? is it related to the branch I'm installing? to the image?