[bug]: Regression on MPS - SDXL with only Prompt outputs junk

[psychedelicious]

Is there an existing issue for this problem?

• I have searched the existing issues

Operating system

macOS

GPU vendor

Apple Silicon (MPS)

GPU model

No response

GPU VRAM

No response

Version number

f06765d

Browser

n/a

Python dependencies

No response

What happened

Generation with only a prompt and no other control layers outputs junk.

First bad commit:
f06765d is the first bad commit

commit f06765dfba736c8b5ca13d319953cc1b8ba1b5f3 (HEAD)
Author: Ryan Dick <[email protected]>
Date:   Mon Sep 30 22:36:25 2024 +0000

    Get alternative GGUF implementation working... barely.

 invokeai/backend/model_manager/load/model_loaders/flux.py |   3 +--
 invokeai/backend/quantization/gguf/ggml_tensor.py         |  78 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++---------------------
 invokeai/backend/quantization/gguf/loaders.py             | 114 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++--------------------------------------------------
 pyproject.toml                                            |   4 ++--
 4 files changed, 124 insertions(+), 75 deletions(-)

User reports downgrading torch and torchvision resolves the issue.

More details in discord thread: https://discord.com/channels/1020123559063990373/1292396911852126228

What you expected to happen

it works

How to reproduce the problem

No response

Additional context

No response

Discord username

i3oc9i

[psychedelicious]

More reports from discord: https://discord.com/channels/1020123559063990373/1149510134058471514/1292877390640320523

[Vargol]

Seems to be an issues with the default attention type and PyTorch 2.4.1.
It can be worked around by changing the attention type in invoke.yaml to torch-sdp

attention_type: torch-sdp

or by upgrading torch to a torch nightly and 2.5.0 test version (or downgrading to 2.3.1)

[RyanJDick]

Test results on my Apple M3:

SD1.5, 1024x1024

• torch 2.2.2 sliced: works
• torch 2.4.1 sliced: produces noise
• torch 2.4.1 non-sliced: maxes out memory and is extremely slow - did not run to completion
• torch 2.6.0.dev20241008 (nightly) sliced: produces noise
• torch 2.6.0.dev20241008 (nightly) non-sliced: RuntimeError: Invalid buffer size: 16.00 GB

SD1.5, 512x512

• torch 2.6.0.dev20241008 (nightly) sliced: produces noise
• torch 2.6.0.dev20241008 (nightly) non-sliced: works
• torch 2.4.1 sliced: produces noise
• torch 2.4.1 non-sliced: works

SDXL, 1024x1024

• torch 2.4.1 sliced: noise
• torch 2.4.1 non-sliced: works

The torch nightly version did not solve the issue for me.
I have put up a PR with a workaround: #7066. I see this as a temporary solution that we'll want to re-visit at some point.

The best solution would be to simply get sliced attention working properly on MPS with the latest torch. Might make sense to do this when we push #6550 across the finish line.