ACE-Step 1.5 on AMD RX 7900 XT (ROCm 7.2, Windows 11) - Working Setup + Fixes

[tuckeranglemyer-pixel]

ACE-Step 1.5 on AMD RX 7900 XT (ROCm 7.2, Windows 11) — Working Setup + Fixes

Got ACE-Step 1.5 running on an AMD GPU on Windows with ROCm 7.2. Documenting the full setup and the four fixes required since this doesn't work out of the box. Hopefully this saves someone else a few hours.

Hardware & Environment

• GPU: AMD Radeon RX 7900 XT (20 GB VRAM)
• CPU: AMD Ryzen 9 7900X3D
• RAM: 32 GB
• OS: Windows 11
• Python: 3.12.10
• Detected Tier: tier6a (top tier, max batch size 4 with LM, 8 without)

Installation Steps

1. Clone the repo

git clone https://github.com/ACE-Step/ACE-Step-1.5.git
cd ACE-Step-1.5

2. Install ROCm SDK components

pip install --no-cache-dir ^
  https://repo.radeon.com/rocm/windows/rocm-rel-7.2/rocm_sdk_core-7.2.0.dev0-py3-none-win_amd64.whl ^
  https://repo.radeon.com/rocm/windows/rocm-rel-7.2/rocm_sdk_devel-7.2.0.dev0-py3-none-win_amd64.whl ^
  https://repo.radeon.com/rocm/windows/rocm-rel-7.2/rocm_sdk_libraries_custom-7.2.0.dev0-py3-none-win_amd64.whl ^
  https://repo.radeon.com/rocm/windows/rocm-rel-7.2/rocm-7.2.0.dev0.tar.gz

3. Install PyTorch for ROCm

pip install --no-cache-dir ^
  https://repo.radeon.com/rocm/windows/rocm-rel-7.2/torch-2.9.1+rocmsdk20260116-cp312-cp312-win_amd64.whl ^
  https://repo.radeon.com/rocm/windows/rocm-rel-7.2/torchaudio-2.9.1+rocmsdk20260116-cp312-cp312-win_amd64.whl ^
  https://repo.radeon.com/rocm/windows/rocm-rel-7.2/torchvision-0.24.1+rocmsdk20260116-cp312-cp312-win_amd64.whl

4. Install ROCm-compatible dependencies

pip install -r requirements-rocm.txt

5. Verify GPU detection

py -c "import torch; print(torch.__version__); print(torch.cuda.is_available())"

Expected output:

2.9.1+rocmsdk20260116
True

Required Fixes (4 total)

Fix 1: vector_quantize_pytorch — torch.distributed.group import error

ROCm on Windows doesn't ship torch.distributed with the full distributed training backend. The vector_quantize_pytorch package imports it unconditionally and crashes.

File: C:\Users\<USER>\AppData\Local\Programs\Python\Python312\Lib\site-packages\vector_quantize_pytorch\lookup_free_quantization.py

Find (lines 14-15):

import torch.distributed as dist
from torch.distributed import nn as dist_nn

Replace with:

try:
    import torch.distributed as dist
    from torch.distributed import nn as dist_nn
except (ImportError, AttributeError):
    import types
    dist = types.SimpleNamespace(
        is_initialized=lambda: False,
        get_world_size=lambda: 1,
    )
    dist_nn = None

Fix 2: torchao — incompatible with ROCm Windows

torchao tries to register distributed ops (_c10d_functional.all_gather_into_tensor) that don't exist in the ROCm Windows build. The requirements-rocm.txt already excludes it, but if you installed from requirements.txt first, it's still there.

Fix:

pip uninstall torchao -y

Note: This disables quantization features (int8/int4 weight quantization). The model runs fine at full precision — you have 20 GB of VRAM.

Fix 3: Launch command — bypass the batch script

start_gradio_ui_rocm.bat looks for a venv_rocm virtual environment that doesn't exist when you install to system Python. It will refuse to launch even though everything is correctly installed.

Instead of:

.\start_gradio_ui_rocm.bat

Run directly:

py acestep\acestep_v15_pipeline.py

Fix 4: torchaudio defaults to torchcodec which is incompatible with ROCm Windows

torchaudio 2.9.1 defaults to torchcodec for all audio loading/saving, but torchcodec's DLLs aren't compatible with the ROCm PyTorch build. This breaks both audio export (MP3/FLAC saving) and dataset preprocessing for LoRA training.

For audio export: Set output format to WAV in the Gradio UI audio settings. MP3/FLAC export through torchcodec will fail.

For LoRA training preprocessing: Replace the audio loading function to use soundfile instead of torchaudio.

File: ACE-Step-1.5\acestep\training\dataset_builder_modules\preprocess_audio.py

Replace entire contents with:

import soundfile as sf
import torch
import numpy as np
def load_audio_stereo(audio_path, target_sample_rate=48000, max_duration=240.0):

audio, sr = sf.read(audio_path, dtype='float32')

audio = torch.from_numpy(audio)

if audio.dim() == 1:

audio = audio.unsqueeze(0)

else:

audio = audio.T

if audio.shape[0] == 1:

audio = audio.repeat(2, 1)

if sr != target_sample_rate:

import torchaudio.functional as F

audio = F.resample(audio, sr, target_sample_rate)

max_samples = int(max_duration * target_sample_rate)

if audio.shape[1] > max_samples:

audio = audio[:, :max_samples]

return audio, target_sample_rate

Performance Numbers

Metric	CPU (Ryzen 9 7900X3D)	GPU (RX 7900 XT, ROCm)
DiT diffusion (30s track)	~62 seconds	0.8–1.1 seconds
VAE decode (30s track)	~80 seconds	18–41 seconds
Total generation (30s, batch=1)	~2.5 minutes	~30–60 seconds

First generation is slow (~10-15 min) because MIOpen needs to compile and cache GPU kernels. Every generation after that uses the cached kernels and is dramatically faster.

Notes

• The ACESTEP_ROCM_DTYPE=float16 env var is supposed to reduce VRAM usage, but the model still loads at float32 regardless. Batch size of 1 is stable. Batch size of 4 caused an OOM crash (0xC0000005) at float32.
• MIOpen workspace warnings (IsEnoughWorkspace) spam the terminal but are harmless — the computations still complete.
• Audio output format: set to WAV in the UI settings. MP3 export fails because torchcodec can't find ffmpeg DLLs compatible with the ROCm PyTorch build.
• The 5Hz LM works on the PyTorch backend (nano-vllm isn't installed but it falls back gracefully).
• GPU is detected as AMD Radeon RX 7900 XT (20.0 GB, HIP 7.2.26024-f6f897bd3d).

TL;DR

ACE-Step 1.5 runs on AMD GPUs on Windows via ROCm 7.2. Four fixes needed: patch a distributed import in vector_quantize_pytorch, uninstall torchao, launch directly with py acestep\acestep_v15_pipeline.py, and replace the audio loader in preprocess_audio.py with soundfile for LoRA training. Generation goes from ~2.5 min on CPU to under 1 minute on GPU. RX 7000 series cards with 20 GB VRAM hit tier6a (top tier config).

[moffye]

Thank you for your post, the edit, did the trick once I located the file, I had the 'venv' variable set:

Required Fixes (4 total)
Fix 1: vector_quantize_pytorch — torch.distributed.group import error
ROCm on Windows doesn't ship torch.distributed with the full distributed training backend. The vector_quantize_pytorch package imports it unconditionally and crashes.

File: C:\Users<USER>\AppData\Local\Programs\Python\Python312\Lib\site-packages\vector_quantize_pytorch\lookup_free_quantization.py

Find (lines 14-15):

import torch.distributed as dist
from torch.distributed import nn as dist_nn
Replace with:

try:
import torch.distributed as dist
from torch.distributed import nn as dist_nn
except (ImportError, AttributeError):
import types
dist = types.SimpleNamespace(
is_initialized=lambda: False,
get_world_size=lambda: 1,
)
dist_nn = None

[Joydn90]

Thanks to your guide, I successfully trained my first lora, and I also have a 7900XT.

[aidanodr]

@tuckeranglemyer-pixel Thanks for the guide.

In my case it loads fine, finding my GPU as per image attached

Once opened in browser and press INITIALIZE etc it crashes inside CMD Window with errors like "IndentationError: unexpected indent".

I probably have FIX 1 and 4 replacement code NOT indented right?

This file: ACE-Step-1.5\acestep\training\dataset_builder_modules\preprocess_audio.py

I tried to use Google AI and asked it to correctly indent this code - no joy, what it gave me resulted in same errors as per the txt file below

Get this kind of RED error:

Error initializing model: Failed to load model with attention implementations ['sdpa', 'eager']: unexpected indent (lookup_free_quantization.py, line 15)

HERE is the errors and text from the CMD window:

Ace step errors in cmd.txt

Is indentation the issue? If so, Is there any way you could post the correct indentation of the code above? I am a musician, I am in IT but not a coder ( closest I come to that is building websites for clients :D )

thanking you - Aidan

[Myles-Croft]

Yes, and you have to edit another file audio_utils.py. You can use something like VS Code to lint the Python or stick into ChatGPT/CoPilot if you wanted. Got it working on a Strix Halo system which I don't have access to but if you use an AI coder you could probably get it to figure things out fairly quickly by feeding it the errors.

audio_utils.py - _save_mp3 now uses soundfile.write() directly instead of torchaudio.save()
preprocess_audio.py - Uses soundfile instead of torchaudio

[aidanodr]

Thanks @Myles-Croft @tuckeranglemyer-pixel ... I sorted the indentation issue via your suggestion to try CHATGPT. But then when I run:

py acestep\acestep_v15_pipeline.py

Still no go.

After consulting CHATGPT and tried all sorts of suggestions from there, it just appears ACE STEP 1.5 works best with NVIDIA cards vs Radeons ( Mine is a Radeon 780M Card in a AMD RYZEN 7 250 system with 32GB ram, 1TB NVME )

As per CHATGPT - Recommendation (honest) - For your setup:

✔ Best experience → CPU PyTorch
⚠ ROCm → experimental, fragile
❌ SDPA on ROCm iGPU → avoid entirely

So I have switched to CPU rather than GPU to run Ace Step 1.5. As slow as it is, it works and its solid.

Maybe at some point Ace Step 1.5 will become workable out of the box with ALL or most RYZEN GPUs - for now, I think, its just too much trouble to figure out ..

[tuckeranglemyer-pixel]

hey! yeah the indentation gets messed up when copying from GitHub markdown. make sure the try/except block in Fix 1 uses 4 spaces per indent level, not tabs. if you paste the code into VS Code it'll highlight the errors. for Fix 4 (preprocess_audio.py) every line inside the function needs to be indented 4 spaces. lmk if you're still stuck

[tuckeranglemyer-pixel]

rick beato is great, will check out the attorney channel too. 9 tracks in the pipeline is serious, excited to hear them. good luck with the mastering, 360 looks legit. feel free to reach out if you ever get that nvidia card and need help setting things up
best,
tucker

[aidanodr]

A quick Q & on off chance - Have you found a setup or setting where the output isnt slightly clipping or slightly distorted. I think the BASS seems to be a bit of a culprit?

"for the melody to instrument conversion - ACE-Step doesn't really do that unfortunately"

Interestingly SUNO does that. Drop in a single note riff recorded from your guitar, keyboard, voice - the go the Cover route while highlighting that riff wav in Suno and prompt it to use the melofy but play back as a sax or choir or whatever

Aidan

[tuckeranglemyer-pixel]

for the clipping, yeah the bass can overload the output. in the audio output settings theres an Enable Normalization checkbox and a Target Peak db slider. turn normalization on and set target peak to like -3.0 dB. that should tame it. you can add a small fade in/out (0.5s each) to catch any spikes at the start/end.
I actually didnt know about the cover route for melody transfer. thats a clever use of it. ACE-Step is more of a from-scratch generator right now, no style transfer pipeline built in. maybe something the community builds eventually since its open source. for now suno definitely has you on that specific workflow even with all the other baggage

[aidanodr]

thanks @tuckeranglemyer-pixel - for the advice above.

Re "the cover route for melody transfer" - thats what i found very useful in Suno. To be able to upload me playing a single note guitar line, upload then ask Suno to "cover" it but give me same melody line played back as a choir or a sax solo etc in the BPM of a track i might be working on in Reaper. My only other recourse would be to hire a sax player or go to a hall for a choir and record. All for whats an idea that may or may not work.

Have you tried to replace certain LYRIC lines in a track already done in Ace step?

I see two possible tools for this??

Cover - Restyle an existing song
Repaint - Fix one section of a song

Any ideas in this regard appreciated

Also I somehow missed your mention of UNTRACKED which you are developing.

Is this it - https://github.com/tuckeranglemyer-pixel/untracked-audio-engine

Im curious - what is it? And can i install, try, how?

• Aidan

[tuckeranglemyer-pixel]

for cover/repaint - havent tested those much, still mostly doing from-scratch generation. if you figure out a good workflow for replacing sections let me know, that would be useful to document here for others.
and yeah thats the repo but honestly untracked probably isnt useful for you. its a discovery tool for DJs to find underground house and garage tracks. very niche, very electronic music focused. your world is way more on the production/recording side. appreciate the interest though!

[aidanodr]

@tuckeranglemyer-pixel The Cover ( Remix it is now called I think ) and Repaint dont seem to do alot to the original track thats added as Sound source. Even when noodling with the Remix strength and Cover strength sliders. The lyric change sort of works if added to the Lyrics box for COVER. You need to add all lyrics again with changes.

No matter how i prompt I cannot really get the male voice to change in character - its the same voice all the time. At the moment its like the software is stuck with sameness, if i could put it that way. Of course maybe i am not using the various settings properly .

Did find this tute doc here:

https://github.com/ace-step/ACE-Step-1.5/blob/main/docs/en/Tutorial.md

And this recent REDDIT Post:

https://www.reddit.com/r/LocalLLaMA/comments/1r0904z/acestep_15_prompt_tips_how_i_get_more/

Best results are probably gained by restarting the Ace Step 1.5 each time and doing a new prompt with a lyric. The cover and repaint where decent change occurs to a track needs a bit of work i would say ..

Any way thats me so far .. Ill work away trying different things

[aidanodr]

@tuckeranglemyer-pixel

"if you figure out a good workflow for replacing sections let me know, that would be useful to document here for others."

Re the Repaint / Cover. I think the cover is now called REMIX in the Gradio UI.

Last evening I was experimenting with both, mainly the remix. I wanted to change the whole lyric on a track previously recorded using the Ace Step CUSTOM Generation Mode. I wanted to keep the music and the arrangement but change the lyric as i had revised that lyric substantially.

To a large extent the following worked. I added the original track ( done previously using CUSTOM Generation Mode ) as the SOURCE AUDIO. I then set the following settings to:

LM CFG Scale: 3

LM Temp: 0.4

Qual Score Sensitivity: 1

Target Peak: -4.5 ( I always set this down this low to deal with potential clipping, give it alot of headroom vs the default -1 )

Remix Strength: 0.9

Cover strength: 1

NO MUSIC CAPTION ( dont want any other things that might change the original music )

Add new changed LYRIC

THINK - off

Then press GENERATE MUSIC.

Another tip I got on REDDIT was to take note of the original tracks SEED number and use that in the seed field when doing the remix / repaint. I could not do this this time as i had not taken note of the seed number in the original:

"It is advisable to use the same seed. if a random seed "-1" is used in Advanced Settings, change it to the one that generated the original track"

The results were pretty much 90% in most cases. It did however stumble over a few lyrics. I did a few generations and will stem out the vocal line of each take .. make one good take inside REAPER ( the DAW I use ) - the rest of the music in the REAPER project I rerecord myself.

Aidan

[AriCuteGirl]

guessing it does not work uhhh for win 10 right ? if no thats fine il just use my cpu like im not upgrading to win 11 due to one ai app :d

[tuckeranglemyer-pixel]

moffye in the comments got it working on Windows 10 with an RX 7800XT. so yeah it should work. the fixes are the same, just make sure youre on Python 3.12 and follow the 4 fixes above. CPU works fine too if you dont want to deal with ROCm setup, just slower

[Myles-Croft]

audio_utils.py
preprocess_audio.py

Here are my two edits to get it working on Strix Halo systems. I find adding a padding of 1.5s intro and 2.5 outro prevents a lot of the clipping that can happen at the start/end, and as noted above lowering the output can help with the upper range clipping. Files are provided as is, they work for me. Still to figure out LLM there's no reason it shouldn't work with ROCm, Triton is problematic under Windows but can work under Linux still to test though been a bit busy.

[tuckeranglemyer-pixel]

this is awesome, thanks for sharing the files. the padding trick for clipping is smart, gonna try that. good to see more AMD systems getting documented here

[tuckeranglemyer-pixel]

documented LoKr training on AMD separately here: ace-step/ACE-Step-1.5#1232