2024-04-17 17:01:45 +01:00
|
|
|
{ pkgs, ... }:
|
2023-11-16 12:25:58 +00:00
|
|
|
|
|
|
|
let
|
2024-03-14 11:55:56 +00:00
|
|
|
backendDomain = "ollama.giugl.io";
|
2024-11-18 19:56:48 +00:00
|
|
|
ollamaHost = "127.0.0.1";
|
2024-03-14 11:55:56 +00:00
|
|
|
ollamaPort = 11434;
|
2025-01-15 17:20:07 +00:00
|
|
|
ollamaPkg = pkgs.unstablePkgs.ollama-cuda;
|
2023-11-16 12:25:58 +00:00
|
|
|
in
|
|
|
|
{
|
2024-03-14 11:55:56 +00:00
|
|
|
environment = {
|
2024-05-27 13:39:06 +01:00
|
|
|
systemPackages = [ ollamaPkg ];
|
2024-03-14 11:55:56 +00:00
|
|
|
};
|
|
|
|
|
2024-05-23 23:45:48 +01:00
|
|
|
services.ollama = {
|
|
|
|
enable = true;
|
2024-11-18 19:56:48 +00:00
|
|
|
|
2024-05-23 23:45:48 +01:00
|
|
|
package = ollamaPkg;
|
2024-11-18 19:56:48 +00:00
|
|
|
host = ollamaHost;
|
|
|
|
port = ollamaPort;
|
|
|
|
acceleration = "cuda";
|
2024-06-04 22:50:49 +01:00
|
|
|
environmentVariables = {
|
|
|
|
OLLAMA_FLASH_ATTENTION = "1";
|
|
|
|
OLLAMA_NUM_PARALLEL = "2";
|
2024-12-08 12:01:01 +00:00
|
|
|
OLLAMA_KV_CACHE_TYPE = "q8_0";
|
2024-06-04 22:50:49 +01:00
|
|
|
};
|
2024-05-23 23:45:48 +01:00
|
|
|
};
|
2024-03-14 11:55:56 +00:00
|
|
|
|
|
|
|
architect.vhost.${backendDomain} = {
|
2024-10-07 13:05:45 +01:00
|
|
|
dnsInterfaces = [ "tailscale" "lan" ];
|
2024-02-21 11:33:54 +00:00
|
|
|
|
|
|
|
locations."/" = {
|
2024-11-18 19:56:48 +00:00
|
|
|
host = ollamaHost;
|
2024-11-17 20:29:57 +00:00
|
|
|
port = ollamaPort;
|
2024-02-21 11:33:54 +00:00
|
|
|
allowLan = true;
|
|
|
|
allowWAN = true;
|
2024-11-17 20:29:57 +00:00
|
|
|
recommendedProxySettings = false;
|
2024-02-21 11:33:54 +00:00
|
|
|
extraConfig = ''
|
2024-04-02 22:56:17 +01:00
|
|
|
proxy_buffering off;
|
2024-02-21 11:33:54 +00:00
|
|
|
proxy_read_timeout 600s;
|
2024-05-14 17:09:45 +01:00
|
|
|
proxy_set_header Host localhost:${toString ollamaPort};
|
2024-02-21 11:33:54 +00:00
|
|
|
'';
|
2023-11-16 12:25:58 +00:00
|
|
|
};
|
|
|
|
};
|
|
|
|
}
|