Auto-enable PYTORCH_ENABLE_MPS_FALLBACK for ComfyUI compatibility

[Copilot]

ComfyUI custom node now sets PYTORCH_ENABLE_MPS_FALLBACK=1 to prevent crashes from unsupported MPS operations.

Changes

• __init__.py: Set environment variable before torch import using setdefault() to allow user override
• __init__.py: Display actual fallback status in console output
• README.md: Remove manual environment variable setup instructions

Implementation

# Enable MPS fallback to CPU for unsupported operations
# This must be set BEFORE importing torch
os.environ.setdefault("PYTORCH_ENABLE_MPS_FALLBACK", "1")

import torch

Console output now shows:

✓ FP8 MPS Metal patch installed successfully!
PYTORCH_ENABLE_MPS_FALLBACK: Enabled
Backend: Native (torch.mps.compile_shader)

Users can override by setting PYTORCH_ENABLE_MPS_FALLBACK=0 before ComfyUI startup.

Original prompt

---

This section details on the original issue you should resolve

<issue_title>AttributeError: module 'torch.mps' has no attribute 'compile_shader'</issue_title>
<issue_description>

ComfyUI Error Report

Error Details

• Node ID: 92:41
• Node Type: SamplerCustomAdvanced
• Exception Type: AttributeError
• Exception Message: module 'torch.mps' has no attribute 'compile_shader'

Stack Trace

^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/rocco/_ART/AIGENERATIVE/ComfyUI/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/rocco/_ART/AIGENERATIVE/ComfyUI/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1747, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Applications/ComfyUI.app/Contents/Resources/ComfyUI/comfy/ldm/lightricks/av_model.py", line 862, in forward
    return super().forward(
           ^^^^^^^^^^^^^^^^
  File "/Applications/ComfyUI.app/Contents/Resources/ComfyUI/comfy/ldm/lightricks/model.py", line 745, in forward
    return comfy.patcher_extension.WrapperExecutor.new_class_executor(
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Applications/ComfyUI.app/Contents/Resources/ComfyUI/comfy/patcher_extension.py", line 112, in execute
    return self.original(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Applications/ComfyUI.app/Contents/Resources/ComfyUI/comfy/ldm/lightricks/model.py", line 792, in _forward
    x = self._process_transformer_blocks(
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Applications/ComfyUI.app/Contents/Resources/ComfyUI/comfy/ldm/lightricks/av_model.py", line 774, in _process_transformer_blocks
    vx, ax = block(
             ^^^^^^
  File "/Users/rocco/_ART/AIGENERATIVE/ComfyUI/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/rocco/_ART/AIGENERATIVE/ComfyUI/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1747, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Applications/ComfyUI.app/Contents/Resources/ComfyUI/comfy/ldm/lightricks/av_model.py", line 236, in forward
    attn1_out = self.attn1(norm_vx, pe=v_pe, transformer_options=transformer_options)
                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/rocco/_ART/AIGENERATIVE/ComfyUI/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/rocco/_ART/AIGENERATIVE/ComfyUI/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1747, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Applications/ComfyUI.app/Contents/Resources/ComfyUI/comfy/ldm/lightricks/model.py", line 367, in forward
    q = self.to_q(x)
        ^^^^^^^^^^^^
  File "/Users/rocco/_ART/AIGENERATIVE/ComfyUI/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/rocco/_ART/AIGENERATIVE/ComfyUI/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1747, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Applications/ComfyUI.app/Contents/Resources/ComfyUI/comfy/ops.py", line 880, in forward
    output = self.forward_comfy_cast_weights(input, compute_dtype)
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Applications/ComfyUI.app/Contents/Resources/ComfyUI/comfy/ops.py", line 850, in forward_comfy_cast_weights
    x = self._forward(input, weight, bias)
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Applications/ComfyUI.app/Contents/Resources/ComfyUI/comfy/ops.py", line 846, in _forward
    return torch.nn.functional.linear(input, weight, bias)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/rocco/_ART/AIGENERATIVE/ComfyUI/.venv/lib/python3.12/site-packages/comfy_kitchen/tensor/base.py", line 348, in __torch_dispatch__
    return op_handlers[parent_cls](qt, args, kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/rocco/_ART/AIGENERATIVE/ComfyUI/.venv/lib/python3.12/site-packages/comfy_kitchen/tensor/fp8.py", line 149, in _handle_fp8_linear
    return torch.nn.functional.linear(*dequantize_args((input_tensor, weight, bias)))
                                       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^...

</details>



<!-- START COPILOT CODING AGENT SUFFIX -->

- Fixes audiohacking/fp8-mps-metal#5

<!-- START COPILOT CODING AGENT TIPS -->
---

💬 We'd love your input! Share your thoughts on Copilot coding agent in our [2 minute survey](https://gh.io/copilot-coding-agent-survey).