Make force free GPU memory work in img2img

[addianto]

This PR attempts to ensure --free_gpu_mem option work in img2img.py.

The solution is similar to the procedure in txt2img.py. I added condition check to see if --free_gpu_mem is enabled. If so, then move the loaded diffusion model into the CPU (the RAM, I presume?).

[lstein]

It looks like the right approach. You tested this, right?

[addianto]

It looks like the right approach. You tested this, right?

I did. Let me explain how I tested the PR.

First, the following is my InvokeAI default setup defined at ~/.invokeai:

--root="D:\[REDACTED]\invokeai"
--outdir="D:\[REDACTED]\invokeai\outputs\all"
--no-nsfw_checker
--free_gpu_mem

The testing was done by monitoring GPU memory usage through Windows' Task Manager while generating img2img images.

Before the PR, the memory usage stayed at the top even though I have --free_gpu_mem enabled, as illustrated in the following screenshot. Sometimes it may resulted in CUDA OOM error in subsequent image generation.

After the PR, the memory usage was able to be reduced each time the img2img finished generating an image:

[lstein]

It looks like the right approach. You tested this, right?

I did. Let me explain how I tested the PR.

First, the following is my InvokeAI default setup defined at ~/.invokeai:

--root="D:\[REDACTED]\invokeai"
--outdir="D:\[REDACTED]\invokeai\outputs\all"
--no-nsfw_checker
--free_gpu_mem

The testing was done by monitoring GPU memory usage through Windows' Task Manager while generating img2img images.

Before the PR, the memory usage stayed at the top even though I have --free_gpu_mem enabled, as illustrated in the following screenshot. Sometimes it may resulted in CUDA OOM error in subsequent image generation.

After the PR, the memory usage was able to be reduced each time the img2img finished generating an image:

More than good enough for me. Thank you for the contribution!