[enhancement]: Fix for MPS producing noise/garbage on AMD RDNA2 GPUs (RX 6000 series on Intel Mac Pro)

[wildfang]

Is there an existing issue for this?

• I have searched the existing issues

Contact Details

No response

What should this feature add?

If you're running InvokeAI on a Mac Pro with an aftermarket AMD RDNA2 GPU (like the RX 6900 XT) and getting noise/colorful garbage instead of real images, here's the fix.

The Problem

PyTorch's F.linear has a bug in its Metal kernel that produces wrong results on RDNA2 GPUs when input dimensions are >= 256. This affects ALL neural network inference — every model produces garbage on MPS.

The Fix

Save this as mps_rdna2_fix.py in your InvokeAI venv's site-packages directory:

"""Fix for broken F.linear on MPS with AMD RDNA2 GPUs."""
import torch

_patched = False
_original_linear = torch.nn.functional.linear

def _patched_linear(input, weight, bias=None):
    if input.device.type == "mps" and input.shape[-1] >= 256:
        orig_shape = input.shape
        input_2d = input.reshape(-1, orig_shape[-1])
        output = torch.mm(input_2d, weight.t().contiguous())
        output = output.reshape(*orig_shape[:-1], weight.shape[0])
        if bias is not None:
            output = output + bias
        return output
    return _original_linear(input, weight, bias)

def apply():
    global _patched
    if _patched or not torch.backends.mps.is_available():
        return
    torch.nn.functional.linear = _patched_linear
    _patched = True
    print("[mps_rdna2_fix] Patched F.linear for MPS/RDNA2 compatibility")

Then create mps_rdna2_fix.pth in the same directory with:

import mps_rdna2_fix; mps_rdna2_fix.apply()

Restart InvokeAI. Images should now generate correctly on MPS.

Affected Hardware

• AMD Radeon RX 6900 XT (confirmed)
• Likely all RDNA2 GPUs: RX 6600, 6700, 6800, 6900 series
• Possibly RDNA1 (RX 5000 series) — unconfirmed
• NOT affected: Vega 56/64 (GCN), which are the stock Mac Pro GPUs

Tested On

• Mac Pro (2019), macOS 15.7.4
• InvokeAI 5.9.1
• PyTorch 2.2.2
• Juggernaut XL v9 (SDXL) — generates correct images at ~1.8 it/s on MPS with the fix

Alternatives

No response

Additional Content

No response