Vae steps runs OOM on a 16g Radeon under Comfyui-zluda
Vae steps fails while trying to get 38g of ram. I tried to reduce the picture size and number of frames. Anything I could try? Is it even possible to run on non-Nvidia hardware? I can run flux1-dev with fp16.
I am running comfyui through anaconda. I get "Invalid or unsupported data type BFloat16" when hitting the queue.
Oh, I have a 7900xtx
I run a 7800 XT on W11. I’m glad all the flux1-dev workflows run nicely, but unfortunately hunyuan video does not.
I get the same BFloat error when running Flux pro, so I just use Flux Lora's for a work around and it does nicely enough. I havent tried FLux dev. I would love to figure this out.
Okay, I have it working. I just installed HIP SDK driver for windows. via https://www.amd.com/en/developer/resources/rocm-hub/hip-sdk.html
I haven't witnessed the full render yet because I am fetching the files for the text encoder and now the tokenizer.
It is taking a while, but I have seen the YouTube channel "AI Search" do this but pause his screen and return to record his video afterward, so I believe it will be working after this is done downloading for the first time. you can watch his video here https://www.youtube.com/watch?v=vhaFcqvSOSU&t=1690s
This is what it is looking like for me,
and
okay, it loaded and restarted its queue, then presented me with same bfloat16 error. sorry.
I followed the same video to build my workflow, same result ;-)
Same error!
RuntimeError: mat1 and mat2 must have the same dtype, but got Float and BFloat16
bfloat16 computations are supported on Ampere or newer GPU architectures (i.e. compute capability >= 8.0).
one note about possible mitigations would be upcasting https://github.com/huggingface/datasets/issues/6566#issuecomment-1901425407 but it might be a lot of work I havent looked
it is not clear to me if this is the fact that it is switching to Cpu and then PERHAPS numpy is failing, or something else.
Anyway OOTB this need cuda afaik