MPS Support?

[LG1234]

They audio engines only provide Auto, CPU and CUDA as the device, but auto defaults to CPU on Apple Silicon. Any chance we could expose MPS?

[diodiogod]

Implemented in v4.14.5.

What was added:

• Created torch_device_resolver.py - centralized device detection utility
• All 7 TTS engine nodes now expose 'mps' as a device option
• Intelligent auto-detection: MPS (Apple Silicon) > CUDA (NVIDIA) > CPU fallback

How it works:
When you select "auto" on a Mac with Apple Silicon:

1. Checks for MPS availability - uses it if available
2. Falls back to CUDA on other systems
3. Falls back to CPU if no GPU available

Backward compatible:

• Existing workflows unchanged
• MPS added as final option to prevent dropdown reordering
• Linux/Windows users unaffected
• "auto" mode now smarter on all platforms

Testing: I cannot test this on actual Apple Silicon hardware. The implementation uses PyTorch's standard torch.backends.mps.is_available() detection, which should work correctly. Testing on a Mac with Apple Silicon would be necessary to confirm functionality.

Test steps on Mac:

1. Update to v4.14.5
2. Create a TTS engine node (any engine)
3. Set device to "auto" or select "mps" directly
4. Check console output to verify which device was selected
5. Generate TTS to confirm it works

Please report results if you test this.

[LG1234]

Thank you. Running the 4.14.5 and selecting auto in the Chatterbox engine still runs on CPU (M4 Max, 128GB GPU)

From the log:
🎤 Generating Chatterbox_Official_23Lang for 'narrator' (lang: ge)
🌍 Loading ChatterBox Official 23-Lang model for German on auto
🌍 Loading ChatterBox Official 23-Lang model for German on cpu

[diodiogod]

if you choose MPS what happens? This is going to be really hard for me to implement because I can't test it on my own... so I don't really know if I can make it. I would love someone with a Mac to do a PR....

[LG1234]

MPS is not a selectable option in the engines currently. The only options are auto and cuda, and auto defaults to CPU. I implemented an MPS option myself, which vastly accelerates performance (>2-3 fold on M4 Max), but degrades for longer prompts due to VRAM/RAM overload, so this needs adjustment to, but probably exceeds my scope if it should be stable.

[diodiogod]

Fixed in v4.14.8 - Comprehensive MPS Support

What was the issue:
The 'torch_device_resolver' utility was implemented in v4.14.5 to detect MPS, but it wasn't being used consistently across all TTS engines. Many engines had hardcoded device selection that always chose CUDA or CPU, completely ignoring MPS availability.

What was fixed:
Added proper device resolution across ALL TTS engines:

• ChatterBox: Added resolve_torch_device() in from_local()
• F5-TTS: Added resolve_torch_device() in init()
• RVC: Fixed 5 locations where device was hardcoded to cuda/cpu
• IndexTTS: Fixed target_device resolution
• VibeVoice: Fixed device resolution in init() and target_device checks
• VibeVoice & Higgs Audio adapters: Fixed device resolution in 3 locations

How it works now:
When you select "auto" device on ANY engine, it properly detects:

1. MPS available → uses it (Apple Silicon)
2. CUDA available → uses it (NVIDIA)
3. Neither → falls back to CPU

Testing on Mac with M-series chip:

1. Update to v4.14.8
2. Create a TTS engine node (ChatterBox, F5-TTS, etc.)
3. Set device to "auto"
4. Check console output - should see "mps" in logs instead of "cpu"
5. Generate TTS - should use GPU acceleration

Please report results!

[LG1234]

MPS is now available and the performance is increased. After all chunks are processed (tested for chatterbox), I now get:

❌ TTS Text generation failed: Node processing failed: ValueError: too many values to unpack (expected 3)
Traceback:
Traceback (most recent call last):
File "/Users/XX/PycharmProjects/ComfyUI/custom_nodes/TTS-Audio-Suite/nodes/base/base_node.py", line 328, in process_with_error_handling
return process_func(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/Users/XX/PycharmProjects/ComfyUI/custom_nodes/TTS-Audio-Suite/nodes/chatterbox_official_23lang/chatterbox_official_23lang_processor.py", line 1210, in _process
processed_text_segments = [segment_text for _, segment_text, _ in character_segments_with_lang]
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/Users/XX/PycharmProjects/ComfyUI/custom_nodes/TTS-Audio-Suite/nodes/chatterbox_official_23lang/chatterbox_official_23lang_processor.py", line 1210, in
processed_text_segments = [segment_text for _, segment_text, _ in character_segments_with_lang]
^^^^^^^^^^^^^^^^^^
ValueError: too many values to unpack (expected 3)

Prompt executed in 03:13:58

I need to test the other engines further.