Three solutions for insufficient video memory
For Step1X-Edit runtime video memory requirements (42GB of video memory is required for the base version), the following are specific solutions:
- Use of FP8 quantization version: The official FP8 quantization model reduces the video memory requirement to 16GB for consumer graphics cards such as the NVIDIA 3090 Ti. Installation requires the selection of the model when downloadingstep1x-edit-i1258-fp8.safetensorsweights file
- Adjusting the resolution parameter: By
--size_levelThe parameter forces the use of 512×512 resolution (default), which saves about 8GB of video memory compared to 1024×1024 - Distributed reasoning scheme: For specialized research scenarios, officially available model parallelism techniques can be used by modifying the
scripts/dist_inference.pyScripts distribute compute load to multiple GPUs
Additional suggestions: close unnecessary background programs to free up video memory, or rent A100/H800 instances in cloud service platforms for temporary computing (~$1.5/hour)
This answer comes from the articleStep1X-Edit: An Open Source Tool for Editing Images with Natural Language InstructionsThe
May not be reproduced without permission:AI productivity tools " How to solve the problem of insufficient video memory when running Step1X-Edit?
