ls250824/comfyui-runtime2 repository overview

⁠comfyui-runtime2

⁠Information

• Docker base image for ComfyUI inference with GPU (CUDA) acceleration.
• This image does not start any services; use ls250824/run-x for that.
• Based on ls250824/pytorch-cuda-ubuntu-runtime:09092026⁠.
• That base image already integrates the official pytorch/pytorch:2.10.0-cuda12.8-cudnn9-runtime image; do not install PyTorch or CUDA again in this image.
• Includes native llama.cpp b10218 and llama-cpp-python 0.3.34 as separate installations so their shared libraries cannot override each other.

⁠Websites

• ComfyUI⁠
• Code server⁠
• HuggingFace cli⁠
• Flash attention⁠
• Sage attention⁠
• Onnxruntime-gpu⁠
• Triton⁠
• torch_generic_nms⁠
• llama.cpp⁠
• llama-cpp-python⁠

⁠Images on Docker

• If the image is less than one day old, it might not be tested yet or might still be updated.

⁠Latest Image Setup

⁠Image

⁠Wheels

⁠Native llama.cpp

The image downloads and installs it under /opt/llama.cpp. The included commands, such as llama-cli, llama-server, and llama-bench, are available through PATH.

The archive is built for CUDA 12.8 on Linux x86_64. Its binaries use an embedded relative RPATH for the libraries in /opt/llama.cpp/lib. llama-cpp-python continues to use its package-local libraries. Do not set LD_LIBRARY_PATH=/opt/llama.cpp/lib or LLAMA_CPP_LIB_PATH globally and do not add this directory to ld.so.conf.

⁠Optimised

⁠Build Constraints (/constraints.txt)

numpy<2
onnxruntime-gpu==1.22.*
onnxruntime==0
flash-attn==2.8.4
llama-cpp-python==0.3.34
sageattention==2.2.0
typer==0.21.1
click==8.*
huggingface_hub>=1.24.0

⁠Adding ComfyUI-Manager Internal Version Instead of Legacy

WORKDIR /ComfyUI
RUN --mount=type=cache,target=/root/.cache/pip \
    python -m pip install --no-cache-dir --root-user-action ignore -c /constraints.txt \
    matrix-nio \
    -r manager_requirements.txt

⁠Docker Speedup

export DOCKER_BUILDKIT=1
export COMPOSE_DOCKER_CLI_BUILD=1