[Bug]: RuntimeError: "slow_conv2d_cpu" not implemented for 'Half' about stable-diffusion-webui

[zero41120]

Here is what I did to make CLIP interrogation work on my M3 Mac:

1. First, I encountered the same error as you, seeing "Half" in the error message. My assumption is that something in the Mac configuration causes the system to use fp16 instead of fp32. Although I don't fully understand what fp16 and fp32 mean, here is the Mac config that gets loaded by default in A1111:

   stable-diffusion-webui/webui-macos-env.sh

   Line 13 in 1c0a0c4

   export COMMANDLINE_ARGS="--skip-torch-cuda-test --upcast-sampling --no-half-vae --use-cpu interrogate"

2. I learned that using the --no-half flag prevents the system from using fp16. So, following the instruction in the Mac file, go to this file:

   stable-diffusion-webui/webui-user.sh

   Line 13 in 1c0a0c4

   #export COMMANDLINE_ARGS=""

3. Uncomment this line and use these flags:

export COMMANDLINE_ARGS="--skip-torch-cuda-test --upcast-sampling --no-half-vae --use-cpu interrogate --no-half"

This makes the interrogation work, and SD1.5 models generate images fine

However, on my 18GB M3 Mac, loading my favorite SDXL model causes the Python memory usage to increase from the usual 10GB to over 20GB, halting image generation. To use the SDXL model again, I have to remove the --no-half flag.

---

From my understanding, the M3 Mac uses a shared memory strategy, allowing both regular RAM and VRAM to use the same 18GB of RAM. However, when total usage exceeds 18GB, the system starts writing data to disk and swapping data between disk and RAM. This process is expensive and, since the SDXL Python process exceeds the 18GB supported by my Mac, even if the request is received, the swapping makes image generation impossible to complete in a reasonable time.

It's likely that the A1111 developers can fix this issue easily by making the interrogation use --no-half by default in the Mac environment, regardless of the flag. This change can be updated in this file:

stable-diffusion-webui/modules/interrogate.py

Line 122 in 1c0a0c4

if not shared.cmd_opts.no_half and not self.running_on_cpu:

.