I am trying to run Edit Mind with GPU acceleration on a Windows host using Docker Desktop (WSL2 backend). My physical GPU is an NVIDIA GeForce GTX 1080 (Pascal Architecture / Compute Capability 6.1).
Here is the exact timeline of what I did and the errors I encountered:
-
Initial Attempt & Driver Block: I first tried running docker compose up following the advanced mode steps exactly as I was not in linux so Easy Mode's | sh didn't work. I made sure to use docker-compose.cuda.yml command for Nvidia GPU. The container deployment immediately crashed with a requirement error from the Nvidia container utility:
nvidia-container-cli: requirement error: unsatisfied condition: cuda>=12.8, please update your driver to a newer version
-
Host Driver Upgrade: To fix this, I upgraded my host machine's CUDA driver environment from version 12.7 to CUDA 12.9. Running nvidia-smi on my Windows host now successfully reflects CUDA 12.9.
-
Fallback to CPU: After updating the host driver, running the compose environment caused this:
{"level":30,"time":1785045277017,"pid":156,"hostname":"0e9ee30f546d","msg":"GPU unavailable — fallback to CPU"}
This appears to be due to the legacy driver branch limits of the GTX 1080 trying to communicate with the modern container runtime dependencies.
-
Bypassing the fallback? EasyOCR fail: To bypass this, I tried to manually intercept the environment configuration, forcing the base runtime target down to nvidia/cuda:12.4.1-runtime-ubuntu22.04 and adding NVIDIA_DISABLE_REQUIRE=1. While the container now boots, it crashes as soon as the background Python machine learning engines attempt to initialize the text recognition features:
edit-mind-ml | 2026-07-26 06:07:08,254 - INFO - BLIP model loaded successfully on device: cuda
edit-mind-ml | 2026-07-26 06:08:59,227 - ERROR - Failed to initialize EasyOCR reader: CUDA error: no kernel image is available for execution on the device
edit-mind-ml | Search for `cudaErrorNoKernelImageForDevice' in https://nvidia.com for more information.
Environment Details
- OS: Windows (WSL2 Backend)
- GPU: NVIDIA GeForce GTX 1080 (Pascal / Compute Capability 6.1)
- Host Driver: Updated and currently sitting at CUDA 12.9 via
nvidia-smi
- Docker Setup: Deploying via the standalone
docker-compose.cuda.yml configuration
Request
Because the pre-compiled wheel files inside the project's default GPU images lack the compiled compute capability paths for older sm_61 cards, the environment drops dead when EasyOCR calls the device.
I tried dropping docker altogether and installing pytorch to run it natively/directly but I don't know how to make it work or if it's possible at all. Since copying individual configuration files via curl leaves the local project directory without the necessary source subfolders (like apps/ml) to run a complete docker compose build --no-cache, is there any recommended workflow that exists to force a local image rebuild with TORCH_CUDA_ARCH_LIST="6.1" and a -devel- base image to make this compatible with a GTX 1080?
Thanks. I really dont want to buy a new GPU just to do this.
I am trying to run Edit Mind with GPU acceleration on a Windows host using Docker Desktop (WSL2 backend). My physical GPU is an NVIDIA GeForce GTX 1080 (Pascal Architecture / Compute Capability 6.1).
Here is the exact timeline of what I did and the errors I encountered:
Initial Attempt & Driver Block: I first tried running
docker compose upfollowing the advanced mode steps exactly as I was not in linux so Easy Mode's| shdidn't work. I made sure to usedocker-compose.cuda.ymlcommand for Nvidia GPU. The container deployment immediately crashed with a requirement error from the Nvidia container utility:nvidia-container-cli: requirement error: unsatisfied condition: cuda>=12.8, please update your driver to a newer versionHost Driver Upgrade: To fix this, I upgraded my host machine's CUDA driver environment from version 12.7 to CUDA 12.9. Running
nvidia-smion my Windows host now successfully reflects CUDA 12.9.Fallback to CPU: After updating the host driver, running the compose environment caused this:
{"level":30,"time":1785045277017,"pid":156,"hostname":"0e9ee30f546d","msg":"GPU unavailable — fallback to CPU"}This appears to be due to the legacy driver branch limits of the GTX 1080 trying to communicate with the modern container runtime dependencies.
Bypassing the fallback? EasyOCR fail: To bypass this, I tried to manually intercept the environment configuration, forcing the base runtime target down to
nvidia/cuda:12.4.1-runtime-ubuntu22.04and addingNVIDIA_DISABLE_REQUIRE=1. While the container now boots, it crashes as soon as the background Python machine learning engines attempt to initialize the text recognition features:Environment Details
nvidia-smidocker-compose.cuda.ymlconfigurationRequest
Because the pre-compiled wheel files inside the project's default GPU images lack the compiled compute capability paths for older
sm_61cards, the environment drops dead when EasyOCR calls the device.I tried dropping docker altogether and installing pytorch to run it natively/directly but I don't know how to make it work or if it's possible at all. Since copying individual configuration files via curl leaves the local project directory without the necessary source subfolders (like
apps/ml) to run a completedocker compose build --no-cache, is there any recommended workflow that exists to force a local image rebuild withTORCH_CUDA_ARCH_LIST="6.1"and a-devel-base image to make this compatible with a GTX 1080?Thanks. I really dont want to buy a new GPU just to do this.