generate_picture and output_path error FileNotFound

[Presjar]

Hello!

I have been following along the bouncing ball of getting PaperPiAI setup on a RPI 2W with the latest Raspbian OS 64bit (Trixie Lite).

It runs python 3.13.5

When running the generation venv/bin/python src/generate_picture.py --width=480 --height=800 --prompt=space /home/rpi/PaperPiAI/genImages
the output eventually results in the below. This folder exists and I can cd to it via the full path shown in the output. I also get the same error when trying to use output directory as /tmp and also when provided no prompt.

Creating image with prompt 'space'
Using seed 2196
Saving to /home/rpi/PaperPiAI/genImages/space_seed_2196_steps_5.png
Running command:
OnnxStream/src/build/sd --xl --turbo --models-path models/stable-diffusion-xl-turbo  -1.0-anyshape-onnxstream --rpi-lowmem --prompt space --seed 2196 --output /home/rpi/PaperPiAI/genImages/space_seed_2196_steps_5.png --steps 5 --res 480x800
----------------[start]-------------------
positive_prompt: space
SDXL turbo doesn't support negative_prompts
output_path: /home/rpi/PaperPiAI/genImages/space_seed_2196_steps_5.png
steps: 5
seed: 2196
using FP32 arithmetic (FP16 not detected)
----------------[prompt]------------------
Token: "space</w>"
----------------[diffusion]---------------
step:0          840688ms
Command executed successfully.
Traceback (most recent call last):
  File "/home/rpi/PaperPiAI/src/generate_picture.py", line 63, in <module>
    shutil.copyfile(fullpath, shared_fullpath)
    ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/lib/python3.13/shutil.py", line 260, in copyfile
    with open(src, 'rb') as fsrc:
         ~~~~^^^^^^^^^^^
FileNotFoundError: [Errno 2] No such file or directory: '/home/rpi/PaperPiAI/genImages/space_seed_2196_steps_5.png'

As the original install method used in your setup script for install of XNNPACK and OnnxStream results in compilation errors due to issues on the XNNPACK side of things, vitoplantamura/OnnxStream provided a fix to build XNNPACK and then ONNXStream in one go. https://github.com/vitoplantamura/OnnxStream/issues/133#issuecomment-3716914971 which I now have setup. This involved updating GCC and G++ to v15.

Do you have any ideas on this? Thanks :)

This is the install script I completed in parts

#!/bin/bash

INSTALL_DIR="$PWD"
echo Installing to "$INSTALL_DIR"

sudo apt-get update
sudo apt-get upgrade
sudo apt-get -y install tmux vim
sudo apt-get -y install cmake
sudo apt-get -y install python3-dev python3-venv python3-pip
sudo apt-get -y install imagemagick
sudo apt-get -y install git git-lfs
sudo apt-get -y install libopencv-dev  python3-opencv

cd "$INSTALL_DIR"
python3 -m venv venv
. venv/bin/activate

python -m pip install opencv_contrib_python
python -m pip install inky[rpi]==1.5.0
python -m pip install pillow

# Following updated instructions taken directly from [OnnxStream repo](https://github.com/vitoplantamura/OnnxStream).

git clone https://github.com/vitoplantamura/OnnxStream.git
cd OnnxStream/src
mkdir build
cd build
cmake ..
cmake --build . --config Release

cd "$INSTALL_DIR"
mkdir models
cd models
git clone --depth=1 https://huggingface.co/vitoplantamura/stable-diffusion-xl-turbo-1.0-anyshape-onnxstream

[dylski]

I have tried and failed to get it to work with Trixie (Debian 13). I think the OS takes up too much RAM and everything I tried got slightly closer to getting it to work (sometimes it succeeded) but it was never reliable.

The key seems to be minimising RAM usage as it appears that the process is getting killed due to Out Of Memory (OOM) error.

What I've tried

Steps 1-4 are probably the most critical for memory.

1. Disable ZRAM-based swap and Maximize Swap File Size

Trixie defaults to ZRAM (compressed RAM swap), which uses precious physical memory. We must disable it and enforce a large $\mathbf{2\text{GB}}$ swap file on the SD card using rpi-swap.

1. Stop the default ZRAM service:

   sudo systemctl stop dev-zram0.swap

2. Use rpi-swap to create a 2GB persistent swapfile:
   • Ensure the rpi-swap package is installed: sudo apt install rpi-swap
   • Create the configuration file to override the ZRAM default:

     sudo nano /etc/rpi/swap.conf.d/80-use-swapfile.conf

   • Add the following content to set the Swapfile mechanism and a fixed size of 2048MiB (2GB):

     [Main]
     Mechanism=swapfile
     [File]
     FixedSizeMiB=2048

   • Reboot the Pi immediately for this change to take effect: sudo reboot
3. Verify the change after rebooting by checking the swap size:

   free -m

---

2. Lower Kernel Memory Reservation (vm.min_free_kbytes)

This parameter reserves a block of RAM for the kernel. By default, it's too large ($\mathbf{16\text{MB}}$) for a 512MB system and effectively triggers an OOM killer.

1. Edit the system configuration file:

   sudo nano /etc/sysctl.d/98-rpi.conf

2. Change the line vm.min_free_kbytes = 16384 to the minimum safe limit of 1024 (1MB):

   vm.min_free_kbytes = 1024
   

---

3. Aggressively Use Swap (vm.swappiness)

This tells the kernel to aggressively use the new 2GB swap file to push idle data out of physical RAM, keeping it clear for the demanding Stable Diffusion job.

1. In the same configuration file (98-rpi.conf), add this setting:

   vm.swappiness = 90
   

2. Apply the changes immediately:

   sudo sysctl --load /etc/sysctl.d/98-rpi.conf

---

4. Minimise GPU RAM

To set GPU RAM to 16MB. I don't recall how I did on but I think you either do it through sudo raspi-config for older models or edit /boot/config.txt (adding gpu_mem=) for newer ones.

---

5. Disable Unnecessary System Services (Optional?)

Disable background services that consume memory but are not required for a headless e-ink frame. You can run systemctl list-units --type=servive --state=running to get a list of running services. I don't know how much these helped but I disabled modem, avah1 (makes the rpi discoverable on the intranet - I'm logging in locally or with IP address) and bluetooth.

# Stop and disable Mobile Broadband Manager
sudo systemctl disable ModemManager.service
sudo systemctl stop ModemManager.service

# Stop and disable Zeroconf/mDNS service discovery
sudo systemctl disable avahi-daemon.service
sudo systemctl stop avahi-daemon.service

# Stop and disable bluetooth Daemon
sudo systemctl disable bluetooth.service
sudo systemctl stop bluetooth.service

[Presjar]

Oh I see! "The No such file or directory" is actually because the diffusion terminated before the image was created, therefore nothing to copy?

I should expect diffusion to show 5 steps and the completion time for each in ms if all worked correctly?

I tried also with a single step, and that resulted in no ms being shown for step 1 but as I read in another issue, running a generation too often doesn't allow the memory to flush properly.

I'll have some more attempts after making the memory changes you suggested. But might just need to try Debian 12.

[Delph]

Oh I see! "The No such file or directory" is actually because the diffusion terminated before the image was created, therefore nothing to copy?

I should expect diffusion to show 5 steps and the completion time for each in ms if all worked correctly?

I tried also with a single step, and that resulted in no ms being shown for step 1 but as I read in another issue, running a generation too often doesn't allow the memory to flush properly.

I'll have some more attempts after making the memory changes you suggested. But might just need to try Debian 12.

Looks like it, we don't actually check the exit value, we just optimistically print the Command executed successfully message;

PaperPiAI/src/generate_picture.py

Line 59 in 2524df9

subprocess.run(cmd)

[dylski]

Yes, @Presjar you'd expect to see 5 steps if successful.
My bad, for blindly printing success without checking the exit value.
Sorry for the confusion!