Sign inSign up

ls250824/run-diffusion-pipe

By ls250824

•Updated 3 months ago

Training image and video models on runpod.io with diffusion pipe

Image
Machine learning & AI
Developer tools
0

1.4K

ls250824/run-diffusion-pipe repository overview

⁠run-diffusion-pipe

⁠Synopsis

  • Models are automatically downloaded based on the specified paths in the environment configuration.
  • /workspace is persistent on RunPod. On first start the container moves diffusion-pipe, README.md, and docs/ into /workspace.
  • Code Server starts on port 9000 and TensorBoard starts on port 6006 when CUDA is available.
  • Authentication credentials can be set via secrets for:
    • Code server authentication (not possible to switch off)
    • Hugging Face token for model access.
    • SSH/SCP access when PUBLIC_KEY is set.

Ensure that the required environment variables and secrets are correctly set before running the pod. See below for options.

⁠Hardware Requirements

  • Volume: Models + dataset (3000 images/frames) + output (10 epochs)
ModelGPUContainer DiskVolume Disk
WAN 2.2RTX A5000 (24 Gb)15 Gb80 Gb
LTX 2.3L40S (48 Gb)15 Gb100 Gb
ZIBRTX 4090 (24 Gb)15 Gb40 Gb
FLUX-Klein-9BRTX 4090 (24 Gb)15 Gb60 Gb
Krea-2RTX 4090 (24 Gb)15 Gb60 Gb

⁠Image setup

ComponentVersion
OSUbuntu 22.04 x86_64
Python3.11.x
PyTorch2.9.0
CUDA12.8.1
Triton3.4.0
nvcc12.8.x
diffusion pipecloned from upstream main during image build
code serverinstalled from official install script during image build

⁠Installed Attentions

⁠Wheels
PackageVersion
flash_attn2.8.3
sageattention2.2.0
⁠Build for
ProcessorCompute CapabilitySM
RTX A50008.6sm_86
RTX 40908.9sm_89

⁠Environment Variables

⁠Authentication Tokens
TokenEnvironment Variable
HuggingfaceHF_TOKEN
Code ServerPASSWORD
SSH/SCPPUBLIC_KEY

⁠Huggingface model configuration

TypeModelFile or directory inside repo/workspace/models/<Directory>Include patternsExclude patterns
PartialHF_MODEL[1-20]HF_MODEL_NAME[1-20]HF_MODEL_LOCAL_DIR[1-20]
FullHF_FULL_MODEL[1-20]HF_FULL_MODEL_LOCAL_DIR[1-20]HF_FULL_MODEL_INCLUDE[1-20]HF_FULL_MODEL_EXCLUDE[1-20]
⁠Workspace move
Environment VariableDefaultDescription
MOVE_STATUS_INTERVAL5Seconds between /diffusion-pipe to /workspace/diffusion-pipe progress checks.
MOVE_STALL_TIMEOUT300Seconds without size changes before the move is treated as stalled.

⁠Connection options

⁠Services
ServicePort
Tensorboard6006 (HTTP)
Code Server9000 (HTTP)
SSH/SCP22 (TCP)

⁠Websites

⁠Supported models

Tag summary

Content type

Image

Digest

sha256:8f682b072…

Size

11.6 GB

Last updated

3 months ago

docker pull ls250824/run-diffusion-pipe:27062026