Repository navigation
Intel Alder Lake UHD: Enhance Caption and audio generation fail on XPU #1367
Replies: 1 comment
|
Short answer: this GPU isn't expected to work on the XPU path as ACE-Step ships today, and the reason is specific. Why both features fail. ACE-Step always runs Intel XPU in bfloat16, and there's no setting to change that: Audio generation: the DiT/VAE handler sets self.dtype = torch.bfloat16 if resolved_device == "xpu" else torch.float32 (init_service_orchestrator.py#L101). That's why it fails at the "move to XPU and convert to the VAE dtype" step even with the LM disabled. Is the GPU supported? The warning you saw is PyTorch saying it doesn't officially support this device. PyTorch's XPU support targets Arc A/B-series and the Arc graphics in Core Ultra chips, not the Xe-LP UHD in 12th-gen Alder Lake. So ACE-Step can't promise anything here beyond what torch-xpu-ops provides. (README-XPU.md mentions "Intel integrated graphics" generally, which is broader than what PyTorch supports.) What you can try: Run in float32 on the XPU. Edit the two lines above locally: in init_service_orchestrator.py L101, use torch.float32 for xpu, and in llm_inference.py L572, change ["cuda", "xpu"] to ["cuda"]. That avoids the failing conversion completely. It does need about twice the memory, and the iGPU shares system RAM, so close other applications and keep CPU offload enabled. Even if it runs, don't expect it to be much faster than CPU on a GT1 iGPU. |
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Title
Intel Alder Lake UHD: Enhance Caption and audio generation fail on XPU
Description
I am testing ACE-Step 1.5 on a Windows system with Intel UHD Graphics
(Alder Lake-P GT1, PCI ID
8086:46A3). PyTorch detects the XPU, but warnsthat this GPU is not officially supported.
What works
including Enhance Caption and audio generation.
arithmetic, CPU-to-XPU transfer, boolean attention-mask conversion, and
small matrix multiplication.
What fails
Level Zero error 45 (
UR_RESULT_ERROR_INVALID_ARGUMENT).reaches reference-audio preparation while moving the tensor to XPU and
converting it to the VAE dtype.
float16orbfloat16fails with Level Zero error 45. This also failswith PyTorch
2.9.0+xpu; it is not limited to the original2.14.1+xpuenvironment.The detailed PyTorch reproducer and environment information are in
intel/torch-xpu-ops#5855.
Environment
8086:46A3)32.0.101.7092(newest available for this system)2.14.1+xpuin the ACE-Step environmentacestep-5Hz-lm-0.6BmodelQuestion
Is this Intel UHD generation expected to work with ACE-Step through PyTorch
XPU, or should this GPU be considered unsupported for these model paths?
Are there recommended ACE-Step settings or a known workaround for this class
of XPU/Level Zero errors?
All reactions