Hi guys! I need some help! ACE-Step + Llama.cpp + acestep-5Hz-lm-4B_q8.gguf

[amoromania]

"Hi guys! I need some help. I'm running an old 4-core Athlon with an RTX 3060 12GB. Standard LLMs wouldn't start on my hardware, so I had Google's AI help me build a custom version of llama.cpp. It works perfectly now. I also managed to run Acestep 1.5 on my machine.
Currently, I've quantized your 4B model into q8.gguf (about 4GB), and it's running smoothly on my custom llama build. Could you please help me 'bridge' my llama setup with Acestep 1.5? I want to use the 4B model as the 'brain' and the Turbo musician for generation. Thanks in advance!"

[ChuxiJ]

I also hope that someone or an AI will integrate it.

[1larity]

This sounds pretty awesome.
It brings up a thought, I have multiple PCs with GPUs, @ChuxiJ do you think there would be any mileage in running the LLM on Ollama on another PC and getting Ace-step to use the Ollama API for LLM calls, so that some VRAM usage can be offloaded to another PC?
This might have some interesting user implications such as using bigger LLMs for planning and text inference, or using uncensored user provided models for explicit lyrics.

[amoromania]

"My CPU doesn't support the latest instruction sets, so whenever I tried running an AI chat before, it would take minutes to respond or wouldn't start at all. But after compiling Llama specifically for my hardware, I can now run models that fit into my VRAM with near-instant responses. It's awesome!"

[amoromania]

"I haven't slept in days, pulling all-nighters with Google's AI to rework the logic and refactor the code. It’s actually working—Llama is talking to the app now—but it still feels a bit off. It would be so much better if the devs just added an 'OpenAI' tab. Every time the chat freezes, I have to re-explain everything from scratch. I'm exhausted, but I'm hooked. And yeah, I’m not a programmer—though maybe I am one at heart."

[ChuxiJ]

@1larity This idea has been in the works for two months now—we just never had the time to implement it. It requires professionals to design and validate the testing methods to ensure accuracy. Our repository was set up a long time ago to run entirely on C++, like with ace-step.cpp, which should resolve device compatibility issues. If needed, I’d recommend the community build a dedicated implementation for this. Let’s work together to make ACE-Step 1.5 more easily deployable across all devices.