Skip to content

Fine-tuning the Mochi video generation model on GH200#

This guide helps you get started fine-tuning Genmo's Mochi video generation model using a Lambda On-Demand Cloud GH200 instance.

Launch your GH200 instance#

Begin by launching a GH200 instance:

  1. In the Lambda Cloud console, navigate to the SSH keys page, click Add SSH Key, and then add or generate a SSH key.
  2. Navigate to the Instances page and click Launch Instance.
  3. Follow the steps in the instance launch wizard.
    • Instance type: Select 1x GH200 (96 GB).
    • Region: Select an available region.
    • Filesystem: Don't attach a filesystem.
    • SSH key: Use the key you created in step 1.
  4. Click Launch instance.
  5. Review the EULAs. If you agree to them, click I agree to the above to start launching your new instance. Instances can take up to five minutes to fully launch.

Install dependencies#

  • Install the dependencies needed for this guide by running:

    git clone https://github.com/genmoai/mochi.git
cd mochi-tune
pip install --upgrade pip setuptools wheel packaging
pip install -e . --no-build-isolation
pip install moviepy==1.0.3 pillow==9.5.0 av==13.1.0
sudo apt -y install bc

Download the model weights#

  • Download the model weights by running:

    python3 ./scripts/download_weights.py weights/

Prepare your dataset#

Begin fine-tuning#

  • Begin fine-tuning by running:

    bash ./demos/fine_tuner/run.bash -c ./demos/fine_tuner/configs/lora.yaml -n 1

You should see output similar to:

Starting training with 1 GPU(s), mode: single_gpu
Using config: ./demos/fine_tuner/configs/lora.yaml
model=weights/dit.safetensors, optimizer=, start_step_num=0
Found 44 training videos in videos_prepared
Loaded 44/44 valid file pairs.
Loading model
Training type: LoRA
Attention mode: sdpa
Loading eval pipeline ...
Timing load_text_encoder
Timing load_vae
Stage                   Time(s)    Percent
load_text_encoder          0.21     17.34%
load_vae                   1.01     82.66%

[…]

Sampling: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 64/64 [03:33<00:00,  3.33s/it]
moving model from cpu -> cuda:0████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 64/64 [03:33<00:00,  1.09it/s]
moving model from cuda:0 -> cpu
Moviepy - Building video finetunes/my_mochi_lora/samples/0_200.mp4.
Moviepy - Writing video finetunes/my_mochi_lora/samples/0_200.mp4

Note

During the fine-tuning, you'll see messages similar to:

W1126 16:46:47.939000 265211801175072 torch/fx/experimental/symbolic_shapes.py:4449] [2/0_1] xindex is not in var_ranges, defaulting to unknown range.
W1126 16:46:51.271000 265211801175072 torch/fx/experimental/symbolic_shapes.py:4449] [2/1_1] xindex is not in var_ranges, defaulting to unknown range.
W1126 16:46:53.847000 265211801175072 torch/fx/experimental/symbolic_shapes.py:4449] [2/2_1] xindex is not in var_ranges, defaulting to unknown range.
W1126 16:46:56.411000 265211801175072 torch/fx/experimental/symbolic_shapes.py:4449] [2/3_1] xindex is not in var_ranges, defaulting to unknown range.

These messages can safely be disregarded.

Cleaning up#

When you're done with your instances, terminate them to avoid incurring unnecessary costs:

  1. In the Lambda Cloud console, navigate to the Instances page.
  2. Select the checkboxes of the instances you want to delete.
  3. Click Terminate. A dialog appears.
  4. Follow the instructions and then click Terminate instances to terminate your instances.

Next steps#