ls250824/comfyui-runtime3 repository overview

⁠comfyui-runtime3

⁠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:22092026⁠.
• That base image already integrates the official pytorch/pytorch:2.12.1-cuda13.0-cudnn9-runtime image; do not install PyTorch or CUDA again in this image.
• Includes native llama.cpp b10218 and llama-cpp-python 0.3.35 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

The PyTorch, torchvision, torchaudio, Triton and CUDA versions above match the supplied base-image build log from 2026-09-22. ONNX Runtime is constrained to 1.22.*; its exact patch version is resolved during the build.

CUDA available: False in a Docker build without GPU access is expected. It does not establish whether GPU execution works in the deployed container. With the NVIDIA driver and NVIDIA Container Toolkit installed on the host, check the built image at runtime:

docker run --rm --gpus all ls250824/comfyui-runtime3:<tag> \
  python -c "import torch; print(torch.__version__, torch.version.cuda); assert torch.cuda.is_available(), 'CUDA is unavailable'; print(torch.cuda.get_device_name(0))"

⁠Wheels

⁠Native llama.cpp

The wheels and native archive are downloaded from LS110824/attentions-llama-cuda130⁠. The native archive is llama-cpp-b10218-cu130-linux-x86_64.tar.gz. The image 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 13.0 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.

⁠Build-time CUDA library detection

CUDA 13 Python wheels share nvidia/cu13/lib, while older wheels use per-package directories such as nvidia/cuda_runtime/lib and nvidia/cublas/lib. The Dockerfile registers discovered NVIDIA library directories with ldconfig and checks for libcudart, libcublas and libnccl files. This check does not load CUDA or require a GPU/driver.

⁠Optimised

⁠Build Constraints (/constraints.txt)

numpy<2
onnxruntime-gpu==1.22.*
onnxruntime==0
flash-attn==2.8.4
llama-cpp-python==0.3.35
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