FLUX Depth FP8 (Blockwise)
FP8 Conversion of Flux Dev Depth
Load Cover image for workflow, Example Depth image does not have workflow.
20-40% speed up over full model for older cards or if loading TE's into GPU
Note that this model is in FP8 E4M3 it is possible that a scaled conversion with mixed E5M2 and E4M3 would be beneficial but based off of the info from Blackforest I stuck with the higher precision.
BLOCKWISE
The following blocks where not converted
"time_in.in_layer.bias",
"time_in.in_layer.weight",
"time_in.out_layer.bias",
"time_in.out_layer.weight",
"txt_in.bias",
"txt_in.weight",
"vector_in.in_layer.bias",
"vector_in.in_layer.weight",
"vector_in.out_layer.bias",
"vector_in.out_layer.weight",
"final_layer.adaLN_modulation.1.bias",
"final_layer.adaLN_modulation.1.weight",
"final_layer.linear.bias",
"final_layer.linear.weight",
"guidance_in.in_layer.bias",
"guidance_in.in_layer.weight",
"guidance_in.out_layer.bias",
"guidance_in.out_layer.weight",
"img_in.bias", "img_in.weight"
Description
FAQ
Comments (5)
Do you have this as a safetensor file??
Yes, just converted and everything looks to be in order BF16, FP8E4M3 will run a test then upload
@Felldude Great thanks
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.