diffusers / stablediffusion still broken in v4.3.0 - v4.4.0

[K4r0qtuYNE5G4qgZ]

LocalAI version:
v4.3.0

Environment, CPU architecture, OS, and Version:
Linux 6.8.0-117-generic #117-Ubuntu SMP PREEMPT_DYNAMIC Tue May 5 19:26:24 UTC 2026 x86_64 x86_64 x86_64 GNU/Linux
Intel Xeon E5-2695 v4 / 128GB DDR4
NVIDIA GeForce RTX 3050

Describe the bug
diffusers backend / stablediffusion broken since ~May 12, Local AI v4.2.3

To Reproduce
Go to Studio Tab
Select stablediffusion in Model Pull out menu
Try to Generate

Expected behavior
That it works like it did prior Local AI v4.0.0 ~v4.2.3

Logs

07:09:01.545
stdout
Initializing libbackend for cuda13-diffusers
07:09:01.551
stdout
Using portable Python
07:09:01.776
stdout
Added /data/LocalAI/backends/cuda13-diffusers/lib to LD_LIBRARY_PATH for GPU libraries
07:09:07.211
stderr
Server started. Listening on: 127.0.0.1:35045
07:09:07.544
stderr
Loading model DreamShaper_8_pruned.safetensors...
07:09:07.544
stderr
Request Model: "DreamShaper_8_pruned.safetensors"
07:09:07.544
stderr
ContextSize: 4096
07:09:07.544
stderr
Seed: 833854170
07:09:07.544
stderr
NBatch: 512
07:09:07.544
stderr
F16Memory: true
07:09:07.544
stderr
MMap: true
07:09:07.544
stderr
NGPULayers: 9999999
07:09:07.544
stderr
Threads: 18
07:09:07.544
stderr
ModelFile: "/data/LocalAI/models/DreamShaper_8_pruned.safetensors"
07:09:07.544
stderr
PipelineType: "StableDiffusionPipeline"
07:09:07.544
stderr
SchedulerType: "k_dpmpp_2m"
07:09:07.544
stderr
CUDA: true
07:09:07.544
stderr
FlashAttention: "auto"
07:09:07.544
stderr
ModelPath: "/data/LocalAI/models"
07:09:07.544
stderr
07:09:07.544
stderr
Options: {}
07:09:07.544
stderr
LoadModel: PipelineType from request: StableDiffusionPipeline
07:09:07.767
stderr
Modular Diffusers is currently an experimental feature under active development. The API is subject to breaking changes in future releases.
07:09:08.006
stderr
[transformers] `Siglip2ImageProcessorFast` is deprecated. The `Fast` suffix for image processors has been removed; use `Siglip2ImageProcessor` instead.
07:09:08.914
stderr

Fetching 11 files:   0%|          | 0/11 [00:00<?, ?it/s]
Fetching 11 files: 100%|██████████| 11/11 [00:00<00:00, 6182.14it/s]
07:09:09.010
stderr

Loading pipeline components...:   0%|          | 0/6 [00:00<?, ?it/s]
Loading pipeline components...:  33%|███▎      | 2/6 [00:00<00:00, 21.54it/s]
07:09:09.013
stdout
Received termination signal. Shutting down...

Additional context
It 100% worked <Local AI v4.0.0 ~v4.2.3 then it said that there were updates for the backends and I guess around there it did break.

Current diffusers Backend:
Version 8830fd803042
Source: quay.io/go-skynet/local-ai-backends:latest-gpu-nvidia-cuda-13-diffusers

When trying to Reinstall Backend, it installs cuda-12-diffusers instead of the already installed cuda-13-diffusers and ends up with both installed. But it doesn't matter which one is set manually, both crash when using stablediffusion.

Also Model: sd-3.5-medium-ggml with Backend: stablediffusion-ggml still works fine without errors.

[K4r0qtuYNE5G4qgZ]

I just upgraded to Local AI 4.3.1 and it also did update all backends today.

Versions now:
diffusers: 9bd285ac3498
cuda13-diffusers: 9bd285ac3498

But the issue remains.

18:32:15.152
stdout
Initializing libbackend for cuda13-diffusers
18:32:15.158
stdout
Using portable Python
18:32:15.374
stdout
Added /data/LocalAI/backends/cuda13-diffusers/lib to LD_LIBRARY_PATH for GPU libraries
18:32:27.203
stderr
Server started. Listening on: 127.0.0.1:44371
18:32:29.159
stderr
Loading model DreamShaper_8_pruned.safetensors...
18:32:29.159
stderr
Request Model: "DreamShaper_8_pruned.safetensors"
18:32:29.159
stderr
ContextSize: 4096
18:32:29.159
stderr
Seed: 923319727
18:32:29.159
stderr
NBatch: 512
18:32:29.159
stderr
F16Memory: true
18:32:29.159
stderr
MMap: true
18:32:29.159
stderr
NGPULayers: 9999999
18:32:29.159
stderr
Threads: 18
18:32:29.159
stderr
ModelFile: "/data/LocalAI/models/DreamShaper_8_pruned.safetensors"
18:32:29.159
stderr
PipelineType: "StableDiffusionPipeline"
18:32:29.159
stderr
SchedulerType: "k_dpmpp_2m"
18:32:29.159
stderr
CUDA: true
18:32:29.159
stderr
FlashAttention: "auto"
18:32:29.159
stderr
ModelPath: "/data/LocalAI/models"
18:32:29.159
stderr
18:32:29.159
stderr
Options: {}
18:32:29.159
stderr
LoadModel: PipelineType from request: StableDiffusionPipeline
18:32:29.745
stderr
Modular Diffusers is currently an experimental feature under active development. The API is subject to breaking changes in future releases.
18:32:30.481
stderr
[transformers] `Siglip2ImageProcessorFast` is deprecated. The `Fast` suffix for image processors has been removed; use `Siglip2ImageProcessor` instead.
18:32:32.042
stderr

Fetching 11 files:   0%|          | 0/11 [00:00<?, ?it/s]
Fetching 11 files: 100%|██████████| 11/11 [00:00<00:00, 6927.53it/s]
18:32:32.134
stderr

Loading pipeline components...:   0%|          | 0/6 [00:00<?, ?it/s]
Loading pipeline components...:  33%|███▎      | 2/6 [00:00<00:00, 22.84it/s]
18:32:32.138
stdout
Received termination signal. Shutting down...

[K4r0qtuYNE5G4qgZ]

Upgraded to Local AI 4.3.4 and also updated all backends.

Versions now:
diffusers: 1a8161ee3024
cuda13-diffusers: 1a8161ee3024

The issue remains.

08:43:33.557
stdout
Initializing libbackend for cuda13-diffusers
08:43:33.562
stdout
Using portable Python
08:43:33.762
stdout
Added /data/LocalAI/backends/cuda13-diffusers/lib to LD_LIBRARY_PATH for GPU libraries
08:43:45.564
stderr
Server started. Listening on: 127.0.0.1:46757
08:43:47.562
stderr
Loading model DreamShaper_8_pruned.safetensors...
08:43:47.562
stderr
Request Model: "DreamShaper_8_pruned.safetensors"
08:43:47.562
stderr
ContextSize: 4096
08:43:47.562
stderr
Seed: 1625666905
08:43:47.562
stderr
NBatch: 512
08:43:47.562
stderr
F16Memory: true
08:43:47.562
stderr
MMap: true
08:43:47.562
stderr
NGPULayers: 9999999
08:43:47.562
stderr
Threads: 18
08:43:47.562
stderr
ModelFile: "/data/LocalAI/models/DreamShaper_8_pruned.safetensors"
08:43:47.562
stderr
PipelineType: "StableDiffusionPipeline"
08:43:47.562
stderr
SchedulerType: "k_dpmpp_2m"
08:43:47.562
stderr
CUDA: true
08:43:47.562
stderr
FlashAttention: "auto"
08:43:47.562
stderr
ModelPath: "/data/LocalAI/models"
08:43:47.562
stderr
08:43:47.562
stderr
Options: {}
08:43:47.562
stderr
LoadModel: PipelineType from request: StableDiffusionPipeline
08:43:48.131
stderr
Modular Diffusers is currently an experimental feature under active development. The API is subject to breaking changes in future releases.
08:43:48.866
stderr
[transformers] `Siglip2ImageProcessorFast` is deprecated. The `Fast` suffix for image processors has been removed; use `Siglip2ImageProcessor` instead.
08:43:50.449
stderr

Fetching 11 files:   0%|          | 0/11 [00:00<?, ?it/s]
Fetching 11 files: 100%|██████████| 11/11 [00:00<00:00, 21569.59it/s]
08:43:50.549
stderr

Loading pipeline components...:   0%|          | 0/6 [00:00<?, ?it/s]
Loading pipeline components...:  33%|███▎      | 2/6 [00:00<00:00, 21.09it/s]
08:43:50.553
stdout
Received termination signal. Shutting down...

[localai-bot]

Consolidating #9082 here. The root cause is a diffusers/transformers version incompatibility breaking single-file SD1.5 safetensors load ('CLIPTextModel' object has no attribute 'text_model'), plus an earlier uv-missing install gap in the cuda images. The fix is to pin/bump the diffusers+transformers versions in the diffusers backend requirements - we're identifying the correct combination. Tracking as the primary SD-broken issue.