[GH-ISSUE #1586] ollama models corrupted? #875

Closed
opened 2026-04-12 10:32:13 -05:00 by GiteaMirror · 5 comments
Owner

Originally created by @iplayfast on GitHub (Dec 18, 2023).
Original GitHub issue: https://github.com/ollama/ollama/issues/1586

I've noticed that after running a few models, sometimes the models don't behave normally. This is a session where that was occurring. I had first tried with bakllava but it wasn't being helpful either. But notice that after I did the systemctl restart ollama the results were much better.

Is something being corrupted in memory? I'll do what I can to help debug this.

ollama run llava
>>> look at ./classic.jpg
Added image './classic.jpg'



>>> what is it?




>>> what is ./classic.jpg
Added image './classic.jpg'




>>> hello





>>> /bye
chris@FORGE:~/ai/aiprojects/OllamaPlayground/createnotes$ systemctl restart ollama
chris@FORGE:~/ai/aiprojects/OllamaPlayground/createnotes$ ollama run llava
>>> look at ./classic.jpg
Added image './classic.jpg'
1. The Underwood typewriter is an old fashioned machine that appears to be made of wood and metal components.
2. It has a black keyboard with silver numbers on the side, giving it a vintage appearance.
3. There are several keys visible, including letters such as A, B, C, D, E, F, G, H, I, J, K, L, M and N along with numeric keys 1 through 9.
4. The typewriter sits on a table and is placed underneath the desk.

Originally created by @iplayfast on GitHub (Dec 18, 2023). Original GitHub issue: https://github.com/ollama/ollama/issues/1586 I've noticed that after running a few models, sometimes the models don't behave normally. This is a session where that was occurring. I had first tried with bakllava but it wasn't being helpful either. But notice that after I did the systemctl restart ollama the results were much better. Is something being corrupted in memory? I'll do what I can to help debug this. ``` ollama run llava >>> look at ./classic.jpg Added image './classic.jpg' >>> what is it? >>> what is ./classic.jpg Added image './classic.jpg' >>> hello >>> /bye chris@FORGE:~/ai/aiprojects/OllamaPlayground/createnotes$ systemctl restart ollama chris@FORGE:~/ai/aiprojects/OllamaPlayground/createnotes$ ollama run llava >>> look at ./classic.jpg Added image './classic.jpg' 1. The Underwood typewriter is an old fashioned machine that appears to be made of wood and metal components. 2. It has a black keyboard with silver numbers on the side, giving it a vintage appearance. 3. There are several keys visible, including letters such as A, B, C, D, E, F, G, H, I, J, K, L, M and N along with numeric keys 1 through 9. 4. The typewriter sits on a table and is placed underneath the desk. ```
Author
Owner

@igorschlum commented on GitHub (Dec 18, 2023):

hi @iplayfast

I played and if you add image in your prompt, it's working well without restarting Ollama neither llava

what is ./image.png
Added image './image.png'

look at ./image.png
Added image './image.png'
4:01:37 PM

what is the size of ./image.png
Added image './image.png'
50,000 bytes

what is the size of this picture ./image.png
Added image './image.png'
1024x768

look at ./image.png
Added image './image.png'
6

(base) igor@MacStudiodeIgor ~ % look at ./image.png
(base) igor@MacStudiodeIgor ~ % ollama run llava

look at ./image.png
Added image './image.png'
10:54 AM

describe ./image.png
Added image './image.png'

  1. You

describe: ./image.png
Added image './image.png'

  1. The main class is "class AdminPanelProvider".
  2. The code is written in C++ language.

describe: /Users/igor/image.png
Added image '/Users/igor/image.png'

COLOR: Blue

describe this image: /Users/igor/image.png
Added image '/Users/igor/image.png'

I see a large amount of computer code, which is being displayed on a black
background. The code appears to be written in a purple and yellow color
scheme, with multiple colors filling the screen. The code appears to be
written in C++ programming language, and it has a specific focus on the
administration panel. There are several instances of the word "panel"
throughout the code, indicating that this is a key element within the
program.

describe this image ./image.png
Added image './image.png'

The image is a close-up of code in black and purple, likely written using
Python. The code appears to be related to an email system, with sections
such as the "app" and "page" parts being prominent features. There are
also lines for the "frame" and "data" components within this programming
setup.

<!-- gh-comment-id:1861667435 --> @igorschlum commented on GitHub (Dec 18, 2023): hi @iplayfast I played and if you add image in your prompt, it's working well without restarting Ollama neither llava >>> what is ./image.png Added image './image.png' >>> look at ./image.png Added image './image.png' 4:01:37 PM >>> what is the size of ./image.png Added image './image.png' 50,000 bytes >>> what is the size of this picture ./image.png Added image './image.png' 1024x768 >>> look at ./image.png Added image './image.png' 6 >>> (base) igor@MacStudiodeIgor ~ % look at ./image.png (base) igor@MacStudiodeIgor ~ % ollama run llava >>> look at ./image.png Added image './image.png' 10:54 AM >>> describe ./image.png Added image './image.png' 1. You >>> describe: ./image.png Added image './image.png' 1. The main class is "class AdminPanelProvider". 2. The code is written in C++ language. >>> describe: /Users/igor/image.png Added image '/Users/igor/image.png' COLOR: Blue >>> describe this image: /Users/igor/image.png Added image '/Users/igor/image.png' I see a large amount of computer code, which is being displayed on a black background. The code appears to be written in a purple and yellow color scheme, with multiple colors filling the screen. The code appears to be written in C++ programming language, and it has a specific focus on the administration panel. There are several instances of the word "panel" throughout the code, indicating that this is a key element within the program. >>> describe this image ./image.png Added image './image.png' The image is a close-up of code in black and purple, likely written using Python. The code appears to be related to an email system, with sections such as the "app" and "page" parts being prominent features. There are also lines for the "frame" and "data" components within this programming setup.
Author
Owner

@iplayfast commented on GitHub (Dec 19, 2023):

No, I've been adding images. It seems to be a corruption that accumulates over time, on the server.
The code at https://github.com/iplayfast/OllamaPlayground/tree/main/createnotes goes through all models and asks questions of them. By the time it gets to the end of the list the Ollama server is pretty much braindead.

<!-- gh-comment-id:1863467655 --> @iplayfast commented on GitHub (Dec 19, 2023): No, I've been adding images. It seems to be a corruption that accumulates over time, on the server. The code at https://github.com/iplayfast/OllamaPlayground/tree/main/createnotes goes through all models and asks questions of them. By the time it gets to the end of the list the Ollama server is pretty much braindead.
Author
Owner

@igorschlum commented on GitHub (Dec 19, 2023):

@iplayfast I will try to run your script on a MacStation with lot of memory and see if I can reproduce the problem.

<!-- gh-comment-id:1863589829 --> @igorschlum commented on GitHub (Dec 19, 2023): @iplayfast I will try to run your script on a MacStation with lot of memory and see if I can reproduce the problem.
Author
Owner

@iplayfast commented on GitHub (Dec 29, 2023):

close issue as it is covered elsewhere

<!-- gh-comment-id:1871726622 --> @iplayfast commented on GitHub (Dec 29, 2023): close issue as it is covered elsewhere
Author
Owner

@igorschlum commented on GitHub (Dec 30, 2023):

@iplayfast I could run your script createnotes on my Mac and Ollama stop working and my fan is on since.
Ollama is taking 97,8% of the GPU. I will test again with Ollama 0.1.18 when released to see if the issue is resolve

(modelnotes) igor@macigor-2 createnotes % python CreateNotes.py
Attempting to load each model to see if they can be loaded
attempting to load model codellama:34b
model codellama:34b loaded in 34.7 seconds
attempting to load model codellama:latest
model codellama:latest loaded in 7.9 seconds
attempting to load model deepseek-coder:6.7b-base-q5_K_M
model deepseek-coder:6.7b-base-q5_K_M loaded in 38.9 seconds
attempting to load model deepseek-coder:latest
model deepseek-coder:latest loaded in 2.4 seconds
attempting to load model dolphin-phi:latest
model dolphin-phi:latest loaded in 3.4 seconds
attempting to load model falcon:latest
model falcon:latest loaded in 7.6 seconds
attempting to load model llama2:13b
model llama2:13b loaded in 16.1 seconds
attempting to load model llama2:latest
model llama2:latest loaded in 7.0 seconds
attempting to load model llama2-uncensored:latest
model llama2-uncensored:latest loaded in 6.7 seconds
attempting to load model llava:13b
model llava:13b loaded in 14.0 seconds
attempting to load model llava:latest
model llava:latest loaded in 9.7 seconds
attempting to load model medllama2:latest
model medllama2:latest loaded in 7.7 seconds
attempting to load model mistral:latest
model mistral:latest loaded in 7.5 seconds
attempting to load model neural-chat:latest
model neural-chat:latest loaded in 7.8 seconds
attempting to load model orca2:13b
model orca2:13b loaded in 15.8 seconds
attempting to load model stablelm-zephyr:3b-q6_K
model stablelm-zephyr:3b-q6_K loaded in 6.3 seconds
attempting to load model starling-lm:latest
model starling-lm:latest loaded in 8.4 seconds
attempting to load model wizard-math:7b
Timed out after 300 seconds for question: are you there

<!-- gh-comment-id:1872408305 --> @igorschlum commented on GitHub (Dec 30, 2023): @iplayfast I could run your script createnotes on my Mac and Ollama stop working and my fan is on since. Ollama is taking 97,8% of the GPU. I will test again with Ollama 0.1.18 when released to see if the issue is resolve (modelnotes) igor@macigor-2 createnotes % python CreateNotes.py Attempting to load each model to see if they can be loaded attempting to load model codellama:34b model codellama:34b loaded in 34.7 seconds attempting to load model codellama:latest model codellama:latest loaded in 7.9 seconds attempting to load model deepseek-coder:6.7b-base-q5_K_M model deepseek-coder:6.7b-base-q5_K_M loaded in 38.9 seconds attempting to load model deepseek-coder:latest model deepseek-coder:latest loaded in 2.4 seconds attempting to load model dolphin-phi:latest model dolphin-phi:latest loaded in 3.4 seconds attempting to load model falcon:latest model falcon:latest loaded in 7.6 seconds attempting to load model llama2:13b model llama2:13b loaded in 16.1 seconds attempting to load model llama2:latest model llama2:latest loaded in 7.0 seconds attempting to load model llama2-uncensored:latest model llama2-uncensored:latest loaded in 6.7 seconds attempting to load model llava:13b model llava:13b loaded in 14.0 seconds attempting to load model llava:latest model llava:latest loaded in 9.7 seconds attempting to load model medllama2:latest model medllama2:latest loaded in 7.7 seconds attempting to load model mistral:latest model mistral:latest loaded in 7.5 seconds attempting to load model neural-chat:latest model neural-chat:latest loaded in 7.8 seconds attempting to load model orca2:13b model orca2:13b loaded in 15.8 seconds attempting to load model stablelm-zephyr:3b-q6_K model stablelm-zephyr:3b-q6_K loaded in 6.3 seconds attempting to load model starling-lm:latest model starling-lm:latest loaded in 8.4 seconds attempting to load model wizard-math:7b Timed out after 300 seconds for question: are you there
Sign in to join this conversation.
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: github-starred/ollama#875