Slow Generation Times Under Linux vs Windows

[NeonTropics]

Hello all!

I have been using Fooocus for years now under Windows and have to say it's been a pretty easy and fun experience. That said I'm having a really weird issue under Linux that I can't seem to figure out after moving my daily workflow to a Ubuntu 24.04 install.

My System:
RAM: 64 GB
CPU: AMD Ryzen 7 2700X Eight-Core Processor
GPU: NVIDIA GeForce GTX 1070/PCIe/SSE2 (8GB)

Kernel Linux 6.8.0-52-generic x86_64
Release 24.04.1 LTS (Noble Numbat) 64-bit

Under Windows using default config and models I would get about 3.60s/it. Using the same defaults in Linux I get about 20.07s/it. That's a huge performance hit.

I am using the nvidia-driver-550 (proprietary) with the CUDA and CuDNN toolkit installed. I verified CUDA by getting a sample CUDA application and compiling it with NVCC - Success.

Looking at the Fooocus Log it does show my GPU detected as CUDA device 0 and when running a generation I can see the GPU Memory fill and the utilization hit 98%. (same as Windows).

I have tried following the install instructions from GitHub (Conda and Local Python methods) as well as via Pinokio. Everything has the same performance hit.

Any ideas would really be appreciated at this point as I am... quite frustrated.

[NeonTropics]

UPDATE:

Using the --all-in-fp32 flag gets the generation times down from 20.00s/it to 5.85s/it... but it also forces LOW VRAM even with --always-normal-vram set

Card in use is a 1070 8GB

Tried different combos of

[--all-in-fp32 | --all-in-fp16]
[--unet-in-bf16 | --unet-in-fp16 | --unet-in-fp8-e4m3fn | --unet-in-fp8-e5m2]
[--vae-in-fp16 | --vae-in-fp32 | --vae-in-bf16]
[--vae-in-cpu]
[--clip-in-fp8-e4m3fn | --clip-in-fp8-e5m2 | --clip-in-fp16 | --clip-in-fp32]

... but nothing seems to change anything except using --all-in-fp32