inference speed
#2
by
hp47
- opened
as you report: Inference Speed: ~2-3 seconds per image (40 steps, RTX 4090)。in my test, even remove 'negative_prompt', the original qwen-edit-2509 takes ~40-50 seconds to edit an image (40steps, A100).
how could u be so fast?
+1
We were using our own inference framework, and kuan-wang has reported the speed on H100s.
Oh, thanks for pointing out the typo, will fix.
On a standard H100 GPU running in FP16 precision with DiffSynth, inference takes approximately 1.1 seconds per step, resulting in a total runtime of about 43 seconds for 40 steps.
jinleic
changed discussion status to
closed