IMPORTANT The AMD optimizations need to be built/configured in your ComfyUI environment, drop this guided setup prompt into ChatGPT or other flagship LLM so it can guide you through optimizing your ComfyUI for AMD chips and use this AMD Triton optimized workflow.
This is a merge of TextFusion Refusal Reduction lora and Krea2 Optimized for AMD ROCm INT8 checkpoint
Credits go to @puppet_vision and @Capitan01R
Description
Comments (10)
I'm curious what kind of speeds would be expected with a 9070 XT with this checkpoint (and the work to set it up as stated in the description). Currently getting 8s/it with 2560 x 1920 renders using 2 loras and fp8 checkpoints. Would it be worth going through the extra setup?
Your GPU is bigger and badder than mine, testing 1024x1024 with stock comfyUI (--use-pytorch-cross-attention default startup parameter) i got 4.43s/it and with my comfyUI running with rocm/flashattention (--use-flashattention --disable-xformers startup parameters with the environment variable FLASH_ATTENTION_TRITON_AMD_ENABLE=TRUE) I get 4.19s/it
@MistyMystery Ah - okay. So significant gains but not knock-your-socks-off gains. Thanks for the response!
@CrawfatherlyOne If you compare it to 1368×768 / 8 steps: FP8 vs the int8_tensorwise quants the FP8 only scores 5.55s/it compared to this int8_tensorwise quant's score of 3.71s/it, so the change of precision and quantization is where this project knocks socks off.
@MistyMystery Damn I need to try this
Dropped you a message. Please reply 🫡
Its crazy that not more people are buying 9070XT instead of the crazy prices of RTX cards...
Agreed. It's a good card.
Because 5080 and 5090 can't be touched by AMD in comfyui. I have a 5090 from launch week and it is slow in MiniMax.
@randosal I'm not so sure that's as true today as it was a year ago (the VRAM advantage of the 5090 notwithstanding). I remember struggling with ZIT with 3 - 4 minutes per image but now can do a full 2048 x 2048 in about 32 seconds. Krea 2 Turbo, fp8, 5MP in 50 seconds. The advances AMD has made with HIP, ROCm, etc. are commendable, and Comfy has done great accommodating.