fspecii
Feature Request: Vulkan and/or ROCM support
Hi @korgano, thanks for the request!
ACE-Step UI is a frontend that relies on the ACE-Step model for all the GPU-heavy work, so hardware support is determined by the official ACE-Step repo rather than our UI layer.
The good news is that the ACE-Step team has been actively working on broader hardware support — they've already made improvements for Intel devices, and from what we've seen, AMD/ROCm support is something they're working on as well. We expect there will be updates on that front soon.
Once the official repo adds ROCm or Vulkan support, it should work with ACE-Step UI automatically since we just pass through to their API.
We'll keep this issue open to track it. In the meantime, keep an eye on the official ACE-Step repo for hardware support updates!
You can run Ace-Step 1.5 on AMD by following the official guide https://github.com/ace-step/ACE-Step-1.5/blob/main/docs/en/INSTALL.md#amd--rocm-gpus
Currently I am running both the Ace Step server and Ace Step UI on RX 7600 XT
You can run Ace-Step 1.5 on AMD by following the official guide https://github.com/ace-step/ACE-Step-1.5/blob/main/docs/en/INSTALL.md#amd--rocm-gpus
Currently I am running both the Ace Step server and Ace Step UI on RX 7600 XT
How did you get past this error then
ImportError: cannot import name 'group' from 'torch.distributed' (B:\ACE-Step-1.5_2\venv_rocm\Lib\site-packages\torch\distributed_init_.py)
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "C:\Users\Cass\AppData\Local\Programs\Python\Python312\Lib\threading.py", line 1009, in _bootstrap
self._bootstrap_inner()
│ └ <function Thread._bootstrap_inner at 0x000001F5EBE8B380>
â”” <WorkerThread(AnyIO worker thread, started daemon 6340)>
File "C:\Users\Cass\AppData\Local\Programs\Python\Python312\Lib\threading.py", line 1052, in _bootstrap_inner
self.run()
│ └ <function WorkerThread.run at 0x000001F6581AF4C0>
â”” <WorkerThread(AnyIO worker thread, started daemon 6340)>
File "B:\ACE-Step-1.5_2\venv_rocm\Lib\site-packages\anyio_backends_asyncio.py", line 986, in run
result = context.run(func, *args)
│ │ │ └ ('B:\ACE-Step-1.5_2\checkpoints', 'acestep-v15-base', 'auto', True, 'acestep-5Hz-lm-0.6B', 'vllm', False, True, False, Fals...
│ │ └ <function setup_event_handlers.
File "B:\ACE-Step-1.5_2\acestep\ui\gradio\events_init_.py", line 67, in
File "B:\ACE-Step-1.5_2\acestep\ui\gradio\events\generation\service_init.py", line 80, in init_service_wrapper status, enable = dit_handler.initialize_service( │ └ <function AceStepHandler.initialize_service at 0x000001F6531FA160> └ <acestep.handler.AceStepHandler object at 0x000001F655456B10>
File "B:\ACE-Step-1.5_2\acestep\handler.py", line 745, in initialize_service raise RuntimeError(
RuntimeError: Failed to load model with attention implementations ['sdpa', 'eager']: cannot import name 'group' from 'torch.distributed' (B:\ACE-Step-1.5_2\venv_rocm\Lib\site-packages\torch\distributed_init_.py)
You can run Ace-Step 1.5 on AMD by following the official guide https://github.com/ace-step/ACE-Step-1.5/blob/main/docs/en/INSTALL.md#amd--rocm-gpus Currently I am running both the Ace Step server and Ace Step UI on RX 7600 XT
How did you get past this error then
ImportError: cannot import name 'group' from 'torch.distributed' (B:\ACE-Step-1.5_2\venv_rocm\Lib\site-packages\torch\distributed__init__.py)
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "C:\Users\Cass\AppData\Local\Programs\Python\Python312\Lib\threading.py", line 1009, in _bootstrap self._bootstrap_inner() │ └ <function Thread._bootstrap_inner at 0x000001F5EBE8B380> └ <WorkerThread(AnyIO worker thread, started daemon 6340)> File "C:\Users\Cass\AppData\Local\Programs\Python\Python312\Lib\threading.py", line 1052, in _bootstrap_inner self.run() │ └ <function WorkerThread.run at 0x000001F6581AF4C0> └ <WorkerThread(AnyIO worker thread, started daemon 6340)> File "B:\ACE-Step-1.5_2\venv_rocm\Lib\site-packages\anyio_backends_asyncio.py", line 986, in run result = context.run(func, *args) │ │ │ └ ('B:\ACE-Step-1.5_2\checkpoints', 'acestep-v15-base', 'auto', True, 'acestep-5Hz-lm-0.6B', 'vllm', False, True, False, Fals... │ │ └ <function setup_event_handlers.. at 0x000001F65843C0E0> │ └ <method 'run' of '_contextvars.Context' objects> └ <_contextvars.Context object at 0x000001F658216480> File "B:\ACE-Step-1.5_2\venv_rocm\Lib\site-packages\gradio\utils.py", line 1034, in wrapper response = f(*args, **kwargs) │ │ └ {} │ └ ('B:\ACE-Step-1.5_2\checkpoints', 'acestep-v15-base', 'auto', True, 'acestep-5Hz-lm-0.6B', 'vllm', False, True, False, Fals... └ <function setup_event_handlers.. at 0x000001F657AEF7E0>
File "B:\ACE-Step-1.5_2\acestep\ui\gradio\events__init__.py", line 67, in fn=lambda *args: gen_h.init_service_wrapper(dit_handler, llm_handler, *args), │ │ │ │ │ └ ('B:\ACE-Step-1.5_2\checkpoints', 'acestep-v15-base', 'auto', True, 'acestep-5Hz-lm-0.6B', 'vllm', False, True, False, Fals... │ │ │ │ └ <acestep.llm_inference.LLMHandler object at 0x000001F6555524E0> │ │ │ └ <acestep.handler.AceStepHandler object at 0x000001F655456B10> │ │ └ <function init_service_wrapper at 0x000001F6555F16C0> │ └ <module 'acestep.ui.gradio.events.generation_handlers' from 'B:\ACE-Step-1.5_2\acestep\ui\gradio\events\generation_hand... └ ('B:\ACE-Step-1.5_2\checkpoints', 'acestep-v15-base', 'auto', True, 'acestep-5Hz-lm-0.6B', 'vllm', False, True, False, Fals...
File "B:\ACE-Step-1.5_2\acestep\ui\gradio\events\generation\service_init.py", line 80, in init_service_wrapper status, enable = dit_handler.initialize_service( │ └ <function AceStepHandler.initialize_service at 0x000001F6531FA160> └ <acestep.handler.AceStepHandler object at 0x000001F655456B10>
File "B:\ACE-Step-1.5_2\acestep\handler.py", line 745, in initialize_service raise RuntimeError(
RuntimeError: Failed to load model with attention implementations ['sdpa', 'eager']: cannot import name 'group' from 'torch.distributed' (B:\ACE-Step-1.5_2\venv_rocm\Lib\site-packages\torch\distributed__init__.py)
I am on Linux so I did not face this issue Is this error after trying to install the requirement-rocm.txt file?
How did you get past this error then
Try the fix in the attachment. First, rename the original lookup_free_quantization.py to lookup_free_quantization_.py
