kopyl / -a1111-tensorrt-docker Goto Github PK

Docker image to play TensorRT technology with AUTOMATIC1111's web ui. Docker image: kopyl/a1111-tensorrt

• Speed of image generation is ≈75it/s for Stable Diffusion 1.5
• You can run it from Runpod. Video tutorial

• Expose ports: 8888, 3000
• Reserve disk space: 100 GB for contaier volume
• Configure right GPU: NVIDIA RTX 4090

• NVIDIA driver: 535.104.12
• CUDA driver: 11.8

• When you reuse one or some set of Stable Diffusion models a lot (not suitable for one-time generation)
• When you generate a lo of images for your videos (it's suitable cause when you save 1± second per image, it's a lot of time for like 2-3 minutes video)

• Convert SD model to fp16
• Use Oneflow (it's tricky, cause without any modification of the source code of the library you're going to get a huge cold start, so the speed evens out and also you can't use it with A1111, since it's Diffusers-like tech)

• You can freeze current commit of repositories (A1111 and TensorRT extension)
• You can add a serverless Dockerfile and a handler for Runpod