VAE decode stops decoding if a instance is idle for a little while with 2.7.10 windows (BMG)

[SHoogstad]

Describe the issue

Steps to reproduce:

1. generate tings
2. wait for a while
3. try to generate again

what I get in the generation process (its an 2 pass workflow) at the last decode it just stops working/decoding (not errors or warning) while if you keep it active it didn't do it this was also not an issue in 2.6.10

if more information is needed let me know!

[xiguiw]

Describe the issue

Steps to reproduce:

1. generate tings
2. wait for a while
3. try to generate again

what I get in the generation process (its an 2 pass workflow) at the last decode it just stops working/decoding (not errors or warning) while if you keep it active it didn't do it this was also not an issue in 2.6.10

if more information is needed let me know!

@SHoogstad

Thanks for interesting in IPEX.

I cannot a whole picture what happened from this description.
Let me confirm it:

1. The issue cannot be reproduced on 2.6.10+xpu, but happened on 2.7.10+xpu.
2. it's the VAE decode.
   Would you show more information about this?
   the detail steps to reproduce it.
3. After running the workflow, keep it idle for a while, then rerun the same workflow, it stops at decoding?
   What's the workflow/VAE decode is?
4. What windows version (win10/Win11) and version of GPU driver?
5. What's your BMG device type?

Please provide me the information, so that I can reproduce it.
Thanks!

[SHoogstad]

Describe the issue

Steps to reproduce:

1. generate tings
2. wait for a while
3. try to generate again

what I get in the generation process (its an 2 pass workflow) at the last decode it just stops working/decoding (not errors or warning) while if you keep it active it didn't do it this was also not an issue in 2.6.10
if more information is needed let me know!

@SHoogstad

Thanks for interesting in IPEX.

I cannot a whole picture what happened from this description. Let me confirm it:

1. The issue cannot be reproduced on 2.6.10+xpu, but happened on 2.7.10+xpu.
2. it's the VAE decode.
   Would you show more information about this?
   the detail steps to reproduce it.
3. After running the workflow, keep it idle for a while, then rerun the same workflow, it stops at decoding?
   What's the workflow/VAE decode is?
4. What windows version (win10/Win11) and version of GPU driver?
5. What's your BMG device type?

Please provide me the information, so that I can reproduce it. Thanks!

its in an issue in diffusion models like SDXL(stable diffusion xl)

let me give you a little more info:

1. yes
2. its encode not decode mis wrote it sorry about that here is the code its bassicly converting latent to an actual image
3. just as i said just leave it idle no diffusion going on or something for like 10m then encode just stop working
4. win11, driver 32.0.101.6647
5. XPU b580 12gb

[yinghu5]

Hi @SHoogstad,

If possible, could you please provide us your workflow or share the process for the issue?

We try to reproduce the issue on one XPU B580 12GB with window 11 and driver 32.0.101.6647

1. set https_proxy=http://xxx.intel.com:xxx

2. conda create -n ipex27 python=3.12

3. conda activate ipex27
   install intel extension for pytorch and do smoke test and make sure the XPU work normally.

4. git clone comfyUI and cd comfyui-master folder

5. pip install -r requirements

6. python main.py
   go http://127.0.0.1:8188/
   use the template : image =>SDXL simple

and download the SDXL 1.0 base and refine model into models/checkpoints

and try run 1
the GPU run fine and produce the image

after 10 minutues (not click anything), run 2. (the image still keep as above, after 221.27s then produce the second images)
it seems ok to run them one by one.

thank you!

[SHoogstad]

Test.json this is the workflow, the model is this one https://civitai.com/models/946205?modelVersionId=1200155 its an noobai/illustrious model

[yinghu5]

Hi @SHoogstad

thanks a lot for the workflow and models. I downloaded them and put them into related folder, then click the run .
first run:

26 minutes later, then click the "Run" again. the workflow seems run fine.

is there any special when run the workflow or in the environment?

I installed the machine from OS and miniforge. and download compfyUI-master today.
moreover, as pyTorch XPU work with SD model well. so i just install pytorch at this time
python -m pip install torch==2.7.0 torchvision==0.22.0 torchaudio==2.7.0 --index-url https://download.pytorch.org/whl/xpu
(no install IPEX, if you have that, please feel free to pip uninstall it).

[SHoogstad]

is there no use for ipex? for sd, but yeah strange that you can;t replicate it maybe the dual gpu i am running (amd and b580) is interfering somehow

[yinghu5]

Hi @SHoogstad,

Then first how about try pytorch-xpu directly
(As IPEX dev team keep upstreamming the related optimization to pytorch-xpu officially, the SD optimization should be same from IPEX and pytorch-xpu in latest version, so you can try uninstall IPEX 2.7 and see if the vae issue and there any performance change)

second, if possible, how about to try disable the AMD card (from windows device manager) and see the issue still insist?