Ollama

All configurations presented here use docker-compose. Read the Docker instructions first.

Avvertimento

This contents of this page have not been tested for all Ollama versions.

Ollama on Docker setup (CPU only)

Variable name	Description
DATA_PATH	directory containing model files for Ollama
OLLAMA_BASE_URL	the base URL where the Ollama Docker instance is listening on, usually http://${ADDR}:11434.

Basic setup

These instructions will cover the basic docker-compose setup to get Ollama running.

Vedi anche

• Ollama [1]

• GitHub - ollama/ollama: Get up and running with Llama 3, Mistral, Gemma, and other large language models. [2]

• ollama/envconfig/config.go at v0.4.2 · ollama/ollama · GitHub [3]

• Global Configuration Variables for Ollama · Issue #2941 · ollama/ollama · GitHub [4]

• Ollama crashes with Deepseek-Coder-V2-Lite-Instruct · Issue #6199 · ollama/ollama · GitHub [5]

1. follow the Docker instructions

2. create the jobs directories. See reference

   mkdir -p /home/jobs/scripts/by-user/root/docker/ollama
   cd /home/jobs/scripts/by-user/root/docker/ollama
   

3. create a Docker compose file

   /home/jobs/scripts/by-user/root/docker/ollama/docker-compose.yml

   version: '3'
   
   services:
     ollama:
       image: ollama/ollama:0.1.32
       volumes:
         - ${DATA_PATH}:/root/.ollama
       container_name: ollama
       tty: true
       restart: always
       hostname: ollama
       ports:
         - 11434:11434
       environment:
         # Keep all loaded models in RAM forever.
         # See source code for reason of negative value.
         # https://github.com/ollama/ollama/blob/v0.4.2/envconfig/config.go#L99
         # No default unloading after 5 minutes.
         - OLLAMA_KEEP_ALIVE=-1
   
         # Might improve inference.
         - OLLAMA_FLASH_ATTENTION=true
   
         # You will see everything happening.
         - OLLAMA_DEBUG=true
   
         # In RAM.
         - OLLAMA_MAX_LOADED_MODELS=4
   

   Nota

   Replace these variables with the appropriate values

   • DATA_PATH

   Nota

   These settings have been tested on CPU-only setups

4. create a Systemd unit file. See also the Docker compose services section

   /home/jobs/services/by-user/root/docker-compose.ollama.service

   [Unit]
   Requires=docker.service
   Requires=network-online.target
   After=docker.service
   After=network-online.target
   
   [Service]
   Type=simple
   WorkingDirectory=/home/jobs/scripts/by-user/root/docker/ollama
   
   ExecStart=/usr/bin/docker-compose up --remove-orphans
   ExecStop=/usr/bin/docker-compose down --remove-orphans
   
   Restart=always
   
   [Install]
   WantedBy=multi-user.target
   

5. run the deploy script

Improving performance

To possibly improve performance you can change the CPU governor. I haven’t verified the imact of these settings.

Vedi anche

• Set CPU governor to performance in 18.04 - Ask Ubuntu [6]

• CPU frequency scaling - ArchWiki [7]

Avvertimento

There might be some tools already managing the governor such as cpupower, thermald, power-profiles-daemon, etc… Before following the steps below make sure not to have any of them active as this might cause conflicts!

1. login as root

   sudo -i
   

2. check the current governor

   cat /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
   

3. to change to the performance governor simply echo it to the device

   echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
   

4. if you have an Intel CPU there is another setting you can check. It works in a similar way

   cat /sys/devices/system/cpu/cpu*/power/energy_perf_bias
   

5. set this one to maximum performance

   echo 0 | tee /sys/devices/system/cpu/cpu*/power/energy_perf_bias
   

6. to make these changes persistent add them to the root crontab

   crontab -e
   

   A text editor will open. Add this to the first line after the comments

   @reboot /usr/bin/sleep 120 && echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor && echo 0 | tee /sys/devices/system/cpu/cpu*/power/energy_perf_bias
   

Extras

Variable name	Description
AUTH_TOKEN	the authorization token to be used by HTTPS clients to connect to Ollama

HTTPS backend with Authorization

To be able to use Ollama remotely and in a safe manner, without having to deal with SSH tunnelling, you can use a reverse proxy.

At the current stage Ollama does not support any kind of authentication. A trick is to have something basic is by setting an authorization header in the reverse proxy. You’ll be limited to just one token.

In this example we’ll use Apache HTTPD.

Vedi anche

• Requesting support for basic auth or API key authentication · Issue #1053 · ollama/ollama · GitHub [8]

1. generate a random authorization token that you’ll later need to copy. Use a Python shell

   python3 -c 'import uuid; print(uuid.uuid4())'
   

2. serve Ollama via HTTPS by creating a new Apache virtual host. Include this file from the Apache configuration

   /etc/apache2/ollama.apache.conf

   ###########
   # Ollama  #
   ###########
   <IfModule mod_ssl.c>
   <VirtualHost *:80>
       ServerName ${FQDN}
   
       # Force https.
       UseCanonicalName on
       RewriteEngine on
       RewriteCond %{SERVER_NAME} =${FQDN}
   
       # Ignore rewrite rules for 127.0.0.1
       RewriteCond %{HTTP_HOST} !=127.0.0.1
       RewriteCond %{REMOTE_ADDR} !=127.0.0.1
   
       RewriteRule ^ https://%{SERVER_NAME}%{REQUEST_URI} [END,NE,R=permanent]
   </VirtualHost>
   </IfModule>
   
   <IfModule mod_ssl.c>
   <VirtualHost *:443> ${FQDN}
   UseCanonicalName on
   
       Keepalive On
       RewriteEngine on
   
       ServerName ${FQDN}
   
       # See https://github.com/ollama/ollama/issues/1053#issuecomment-2253445923
       RewriteCond %{HTTP:Authorization} !^Bearer\s+${AUTH_TOKEN}$
       RewriteRule ^ - [R=401,L]
   
       SSLCompression off
   
       ProxyRequests Off
       RequestHeader set X-Forwarded-Proto "https"
       RequestHeader set X-Forwarded-Port "443"
       RequestHeader set X-Forwarded-Host "${FQDN}"
       RequestHeader set X-Forwarded-For "%{REMOTE_ADDR}s"
       RequestHeader set Host "${FQDN}"
   
       ProxyPass / http://127.0.0.1:11434/ connectiontimeout=300 timeout=300 Keepalive=On max=50
       ProxyPassReverse / http://127.0.0.1:11434/
   
       Include /etc/letsencrypt/options-ssl-apache.conf
       SSLCertificateFile /etc/letsencrypt/live/${FQDN}/fullchain.pem
       SSLCertificateKeyFile /etc/letsencrypt/live/${FQDN}/privkey.pem
   </VirtualHost>
   </IfModule>
   

   Nota

   Replace these variables with the appropriate values

   • FQDN

   • AUTH_TOKEN

   Nota

   If you use Python’s uuid4 function, the RewriteCond line can be something like:

   RewriteCond %{HTTP:Authorization} !^Bearer\s+db0d7378-f45a-432a-a795-ad48bd6da621$
   

   or

   RewriteCond %{HTTP:Authorization} !^Bearer\s+bfe7e7be-45f0-44d2-891e-038df7b74738$
   

3. test the connection

   1. with the correct header value

      curl -H 'Authorization: Bearer ${AUTH_TOKEN}' https://${FQDN}
      

      results in

      Ollama is running
      

   2. with a wrong header value

      curl -H 'Authorization: Bearer fake-000' https://${FQDN}
      

      results in

      <!DOCTYPE HTML PUBLIC "-//IETF//DTD HTML 2.0//EN">
      <html><head>
      <title>401 Unauthorized</title>
      </head><body>
      <h1>Unauthorized</h1>
      <p>This server could not verify that you
      are authorized to access the document
      requested.  Either you supplied the wrong
      credentials (e.g., bad password), or your
      browser doesn't understand how to supply
      the credentials required.</p>
      <p>Additionally, a 401 Unauthorized
      error was encountered while trying to use an ErrorDocument to handle the request.</p>
      </body></html>
      

   3. without the header alltogether

      curl https://${FQDN}
      

      results in

      <!DOCTYPE HTML PUBLIC "-//IETF//DTD HTML 2.0//EN">
      <html><head>
      <title>401 Unauthorized</title>
      </head><body>
      <h1>Unauthorized</h1>
      <p>This server could not verify that you
      are authorized to access the document
      requested.  Either you supplied the wrong
      credentials (e.g., bad password), or your
      browser doesn't understand how to supply
      the credentials required.</p>
      <p>Additionally, a 401 Unauthorized
      error was encountered while trying to use an ErrorDocument to handle the request.</p>
      </body></html>
      

Android app

There are a couple of Android apps on F-Droid capable of connecting to Ollama. For them to work it is required to follow Basic setup and HTTPS backend with Authorization. You also need to pull some models: you can use Open WebUI to perform this operation using an admin user.

Vedi anche

• maid | F-Droid - Free and Open Source Android App Repository [9]

• GPTMobile | F-Droid - Free and Open Source Android App Repository [10]

1. install GPTMobile

2. open the settings and fill in the API URL, key (authorization token) and model name

3. start a new chat

Footnotes

[1]

https://ollama.com/ unknown license

[2]

ollama/ollama MIT License, Ollama contributors

[3]

ollama/ollama MIT License, Ollama contributors

[4]

ollama/ollama#2941 unknown license

[5]

ollama/ollama#6199 unknown license

[6]

https://askubuntu.com/a/1084727 2018-2020, CC BY-SA 4.0, askubuntu.com contributors

[7]

https://wiki.archlinux.org/title/CPU_frequency_scaling#Intel_performance_and_energy_bias_hint GNU Free Documentation License 1.3 or later license, ArchWiki contributors

[8]

ollama/ollama#1053 unknown license

[9]

https://f-droid.org/en/packages/com.danemadsen.maid/ unknown license

[10]

https://f-droid.org/en/packages/dev.chungjungsoo.gptmobile/ unknown license