Exit code: 1. Reason: Traceback (most recent call last): File “/app/app.py”, line 6, in from logic import generate_music_logic, edit_music_logic, export_file, import_file ImportError: cannot import name ‘import_file’ from ‘logic’ (/app/logic.py)
Container logs:
===== Application Startup at 2026-05-26 16:02:28 =====
Traceback (most recent call last):
File "/app/app.py", line 6, in <module>
from logic import generate_music_logic, edit_music_logic, export_file, import_file
ImportError: cannot import name 'import_file' from 'logic' (/app/logic.py)
Here: many-happy-returns-2 - a Hugging Face Space by Gertie2013
@John6666
Hmm…??? The code for that section lacks the actual generation logic. It seems the API can no longer be found online either.
I had no choice but to replace it with a dummy or a local model.
If you really need that API, you’ll have to find it first—otherwise, you won’t be able to do anything… By the way, I’ve never even heard of that API…
Exit code: 1. Reason: Traceback (most recent call last): File “/app/app.py”, line 12, in tts = pipeline( “text-to-audio”, model=“tencent/SongGeneration”, ) File “/usr/local/lib/python3.13/site-packages/transformers/pipelines/_init_.py”, line 901, in pipeline config = AutoConfig.from_pretrained( model, _from_pipeline=task, code_revision=code_revision, **hub_kwargs, **model_kwargs ) File “/usr/local/lib/python3.13/site-packages/transformers/models/auto/configuration_auto.py”, line 419, in from_pretrained raise ValueError( …<2 lines>… ) ValueError: Unrecognized model in tencent/SongGeneration. Should have a `model_type` key in its config.json.
Container logs:
===== Application Startup at 2026-05-29 20:41:48 =====
Here: New Riffusion - a Hugging Face Space by Gertie2013
@John6666
I’ve fixed it so it works, but it won’t run without a GPU or Zero GPU. If you use a CPU, it probably won’t finish generating in an hour…
https://huggingface.co/spaces/Gertie2013/New-Riffusion/discussions/1
Can you help me complete this app: New Riffusion - a Hugging Face Space by Gertie2013
My app is supposed to generate music with tencent/SongGeneration: https://huggingface.co/tencent/SongGeneration
It’s also supposed to edit music with daydreamlive/DreamVAE: https://huggingface.co/daydreamlive/DreamVAE
@John6666
Hmm… I think a GPU is essential for that use case, but do you plan to allocate GPU Spaces actually? If you’re planning to implement this using CPU Spaces, it’s probably not possible. (Of course, it might be possible if you had an external service handle the processing and just retrieved the results via an API, but I assume you intend to run the inference on the HF side, right?)
Perhaps, you duplicate it to make sure it’s got Standard GPU. Keep the model IDs.
Sorry, I don’t really get what you’re saying…
Also, personally, I don’t use standard GPUs due to budget constraints, so I’ll only be able to help out a little with the coding.
What about CPU Upgrade? That might be Option 2. Coding is Option 1.
Coding tricks and the Upgraded CPU space just aren’t enough. I think it’ll be difficult to run the configuration you’re looking for without GPU space—and not just any GPU, but a relatively high-end one…
The reason the previous Gemini API worked, even if the model wasn’t exactly what you wanted, is because there’s a GPU on the other side of the network—one that’s even more powerful than what I mentioned above. (Since it’s Google, it might be a TPU, but it’s pretty much the same thing. It’s not a CPU.)
Even if you train a single ant, you can’t make it strong enough to beat an elephant… Maybe “ant” is an exaggeration—let’s say it’s strong enough to handle a mouse or a squirrel…
In terms of HF, Zero GPU Spaces costs $9 a month and lets you use it for 25 minutes a day; it’s part of the high-performance GPU Spaces category, but since there’s a daily GPU usage limit (quota), it’ll be inconvenient if you plan to use it for extended periods at once. Also, coding on Zero GPU Spaces is trickier and more finicky than on regular GPU Spaces. Still, since it guarantees performance and is relatively inexpensive, it’s ideal for prototyping or creating demos to show others, but it has too many restrictions if you just want to “use it” for general purposes.
GPUs are expensive in reality. While there are occasionally free usage quotas provided for research support or corporate demos, these are extremely rare exceptions. Most people who want to use OSS AI end up buying expensive consumer-grade GPUs for their personal PCs. There are even a few who purchase enterprise-grade GPUs—which are several times more expensive—for personal use. There are several well-known services, including HF, that offer paid rentals by the hour.
In any case, “free” and “unrestricted” access to GPUs—barring scams—is virtually nonexistent. The only exception might be the 16GB VRAM Tesla T4 GPU on Google Colab Free, which you can use for a few dozen minutes a day…
In short, the bottom line is that “what stands in the way of your (and others’) implementation goals is the financial cost.”
Well, I can’t really help you unless I know your target environment—specifically, what level of GPU you’ll actually be using…
Also, I don’t do custom development work for others. I’m not completely swamped, but I don’t have that much free time. If I’m interested, I might help out on my own, but sometimes I won’t do anything at all.
If you’re looking for a reliable workforce, you’re better off paying someone to do the job…
I used Zero GPU to actually measure the memory usage of the GPU version I created earlier, and it looks like it should run as long as there is at least about 7.5 GB of VRAM and at least about 27 GB of RAM. I can’t say for certain, but this should serve as a rough guide.
Applying this to actual HF Spaces, a T4 small instance would have enough VRAM but not enough RAM to run it. A T4 medium instance might be able to run it. An L40S instance should run slightly faster based on these calculations.
Edit:
By the way, aside from the discussion about T4, daydreamlive/DreamVAE doesn’t seem to be the kind of model used for music editing.