I'm seeing a 12 second run time that isn't there for other save nodes, and that is really hurting me when using comfy with an API. It seems like the SD PromptSaver node might actually be flushing the flux model and reloading it into memory; I'm guessing that if not using the SD ParameterGenerator, this is what you're doing to get model name, which seems really inneficient. if model name could just be a string that would be ideal, or really anything that wouldn't trigger a reload
Reproduction steps
run any basic workflow using the SD PromptSaver, take a look at the inference time reported for that node on the first run from a cold boot
Image file
first run:
second run:
first run with a dummy empty model nullModel.ckpt:
Description
I'm seeing a 12 second run time that isn't there for other save nodes, and that is really hurting me when using comfy with an API. It seems like the SD PromptSaver node might actually be flushing the flux model and reloading it into memory; I'm guessing that if not using the SD ParameterGenerator, this is what you're doing to get model name, which seems really inneficient. if model name could just be a string that would be ideal, or really anything that wouldn't trigger a reload
Reproduction steps
run any basic workflow using the SD PromptSaver, take a look at the inference time reported for that node on the first run from a cold boot
Image file
first run: second run: first run with a dummy empty model nullModel.ckpt: