DirectML acceleration for Flux.schnell gguf

[JohnClaw]

My cpu is rather slow and 4-steps Flux.schnell generation of 512x512 image takes a lot of time (especially due to the fact that fastsdcpu doesn't use all cpu threads and maximum cpu frequency). However, my igpu/apu Vega 7 is rather fast in DirectML mode, much faster than my cpu. Could you add support for DirectML igpu acceleration to fastsdcpu, please?

[rupeshs]

@JohnClaw Currently fastsdcpu uses half number of CPU cores by default (

fastsdcpu/src/constants.py

Line 22 in 390f9f0

GGUF_THREADS = environ.get("GGUF_THREADS", cpu_count() // 2)

)
but you can change it be setting set GGUF_THREADS=10

[JohnClaw]

@JohnClaw Currently fastsdcpu uses half number of CPU cores by default (

fastsdcpu/src/constants.py

Line 22 in 390f9f0

GGUF_THREADS = environ.get("GGUF_THREADS", cpu_count() // 2)

)
but you can change it be setting set GGUF_THREADS=10

Thanks for advice.

[q5325150]

My cpu is rather slow and 4-steps Flux.schnell generation of 512x512 image takes a lot of time (especially due to the fact that fastsdcpu doesn't use all cpu threads and maximum cpu frequency). However, my igpu/apu Vega 7 is rather fast in DirectML mode, much faster than my cpu. Could you add support for DirectML igpu acceleration to fastsdcpu, please?

I want to konw, if a 4-core 16G cloud server is used to run it, how long will it take for Flux to process a 512x512 image?