Comfyui has Ace Step 1.5Currently broken on AMD, apparently, but if you have nvidia, you can gen songs locally, and with lots of good features.https://www.reddit.com/r/comfyui/comments/1quzawn/acestep_15_is_now_available_in_comfyui/
>>108051632the error I'm getting, with amd.
Yeah so the issue is that the audio vae of comfyui is borked on AMD.as usual, comfyui sucks, it sucks in a very Windows-like way.using --cpu-vaesuper slow. I thought this problem was fixed, but nooo.
>>108051680so far, 6 minutes of vae/clip running on cpu...
>>108051713cpu vae takes a huge amount of time with this. too bad the comfyui guy refuses to buy an amd card, so his code is constantly broken on amd.
fwiw, the real problem is hip is incomplete when it comes to audio.but this is a solved problem for vae anyway, so it shouldn't be happening, don't make me vibe code a comfyui fork.
>>108051632MY FIRST GENhttps://files.catbox.moe/hqkvul.mp3btw, has anyone ever gotten the gradio crap to work with amd?
I nuked the gradio stupid shit. what a pantload
God damn github, what a shit website, I searched it like 10 timed and it pukes like it's being ddos'd. massive trash.
>>108052051gradio didn't work on older nvidia card with 8GB either and I was using the lightest model 0.6B
>>108052218Yeah, the gradio / uv thing is trash.
>>108052275Their implementation of it is. Imma tr "comfyui" now, I hate how this dog shit ui became the standard
no luck solvingprobability tensor contains either `inf`, `nan` or element < 0(ie I still have to use cpu vae - that or buy an s tier nvidia gpu to just do stuff that comfyui is incapable of)try to keep the thread alive, I must do life crap
Neevr used comfy before. Trying both manual and portable installations. 16gb vram.VAEDecodeAudio keeps crashing:>RuntimeError: GET was unable to find an engine to execute this computation
>>108053243For me, I have to use --cpu-vaeMIOPEN_FIND_MODE=FAST python main.py --cpu-vaethe miopen_find_mode thing helps with my amd card, on Linux.anyway, --cpu-vae is necessary on amd, because comfyui's audio vae is immature. really, audio everything is immature on comfy.anyway, back slow genning.possible areas of solution, which I may try to investigate as time allows:>zluda comfyui>(actually I won't bother with this one) some other audio gen thingies don't have the cpu-vae issue. heartmula didn't, not sure why. I had to restart the server every few gens, again, not sure exactly how that happened.
>>108051632why the fuck is this a thread and why the fuck are you advertising malware instead of the gradio app from the lab that has all the full implementation? go discuss in >>>/g/ldg
HOW DO YOU KNOW IF YOU ARE USING CPU VAE????Your cpu usage will go way up. I find all cores utilized, but not all out constantly. 1%, 20%, 80%, 1% again, on the same core say CPU7.>>108054987comfyui isn't considered malware. audio gen is very different from /ldg/ and /sdg/ interests. 99% of those guys will never be interested in even listening to any music at all. music is kind of a Gen X thing. Once you get past Gen X, music just isn't that important to people. That's why "hit songs" for 2025 per Billboard were on average 3 minutes or so. vs 5 minutes as the traditional song length, with a few super long songs making it to the top some years.>>108054980correction, this is what is working for me (AMD XTX rx 6950xt):PYTORCH_ALLOC_CONF=expandable_segments:True TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1 MIOPEN_FIND_MODE=FAST python main.py --use-pytorch-cross-attention --novram --cpu-vaeI don't know if I need all options lol. so many combinations to try, but hopefully I'll take the time to look into it.This is btw what I used with SongBloomYes I am genning. It will take maybe 30 minutes for a 120 minute song, for me, I guess. We'll see. This is only my second acestep 1.5
>>108055108>>108053243yeah, so I'm getting a vae decode error too, with audio decode, despite being under cpu decode. it's not a ram issue.I had one successful gen, so I know it *can* work...
I can gen music just fine but not much luck with sound effects, farting noises, etc.
>>108055405share!https://catbox.moe/so far, my gens are taking forever, but I think this one will succeed (I had a node that overrides others, that is incompatible with ace step; it might be an old node that I should avoid anyway)
>>108055438https://files.catbox.moe/4gpmrd.mp3https://files.catbox.moe/2ye4xa.mp3will try more later
>>108055108it's a diffusion model you retard. also this model sucks for the most part. doesn't deserve it's own thread
https://files.catbox.moe/rho1z9.mp3
>>108055492Thanks!be sure to include what your style prompts were!First one is doing what I knew the base model could do. I would sometimes get like 1 second that punched through the low bitrate and laid down bass.secondisn't it strange rich people can't get a girlfriend?>>108055519Yeah but it's the first song model that has promise. heartmula doesn't follow the prompt. ace step 1.35 was low bitrate. songbloom's method is experimental and not ready for prime time.>>108055545style prompt: teen pop
>>108055545>>108055576also, it took me <38 minutes, and needed <50gb system memory (to cpu vae). idk why it needs that much lol
Official Gradio UI has a lot of features, including LoRA training etc... Plus I feel like the stuff is broken for Comfy, on 3090 it reloads prompt each time
https://files.catbox.moe/2imtt6.mp3
