slow perf with heavy text encoder

[lo-th]

i try with last flux-2-klein-9b-kv-fp8
with heavy text_encoders : qwen_3_8b_fp8mixed (8.06 Go)
if text encoder is in normal Chord seam's lost and take very long time, sometime crash
but if text encoder is in cpu is very fast

Normal map is really good best of all method i see before

[alexisrolland]

Hi there and thank you for sharing about your test. Do you mean that you modified the Chord nodes to use qwen_3_8b_fp8mixed as a text encoder? Or do you mean you use qwen_3_8b_fp8mixed upstream from the Chord nodes in your workflow?