Drops to single thread for a long time after iterations
#73
by mtbc - opened
I run StableDiffusion3Pipeline from a checkout of stable-diffusion-3.5-large at commit ceddf0a7 with num_inference_steps=80 on CPU, not GPU. It loads the checkpoint shards and pipeline components then runs through the 80 iterations on a few CPU cores. However, after the 80 iterations, it then drops to a single core for a very long time before actually writing the image.
I don't see the same from FluxPipeline: that runs through the 80 iterations on a few cores then promptly writes the image (pipeline.images[0].save).
Any idea why stable diffusion is adding this extra single-core phase after the iterations are complete, and if there's anything I can do about that?
In case it matters, here are my current versions of some relevant packages:
accelerate==1.4.0
diffusers==0.32.2
huggingface-hub==0.29.1
mpmath==1.3.0
numpy==2.2.3
peft==0.14.0
protobuf==5.29.3
safetensors==0.5.2
sentencepiece==0.2.0
tokenizers==0.21.0
torch==2.6.0+cpu
transformers==4.49.0
xformers==0.0.29.post3