TheCrazyInsanity/nixos-configv3
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

ts sucks vrah

Browse Source
This commit is contained in:
TheCrazyInsanity
2026-03-13 18:43:11 -04:00
parent b7e89f4da4
commit 4b3a369c6e
5 changed files with 205 additions and 24 deletions
Download Patch File Download Diff File Expand all files Collapse all files

37
machines/homepc/configuration.nix
View File

@@ -6,6 +6,7 @@
config,
lib,
pkgs,
inputs,
...
}:
{
@@ -42,11 +43,14 @@
# TODO: honestly, while I currently only have one pc that can run local AI, might change in the future.
# And this config is getting a bit complicated for a single pc config
# Should be moved to it's own shit
environment.systemPackages = with pkgs; [
ollama-cuda
opencode
llama-cpp
llama-swap
environment.systemPackages = [
pkgs.ollama-cuda
pkgs.opencode
pkgs.llama-cpp
pkgs.llama-swap
inputs.llm-agents.packages.${pkgs.stdenv.hostPlatform.system}.pi
inputs.llm-agents.packages.${pkgs.stdenv.hostPlatform.system}.rtk
];
services.ollama = {
@@ -87,32 +91,35 @@
environment.etc."llama-swap/config.yaml".text = ''
models:
"Qwen3.5-35B-A3B-GGUF":
cmd: llama-server --port ''${PORT} -hf unsloth/Qwen3.5-35B-A3B-GGUF:UD-IQ4_XS --ctx-size 128000 --temp 0.6 --top-p 0.95 --top-k 20 --min-p 0.00 -np 1 --cpu-moe --fit-target 256 --ubatch-size 1024 -fa on --slots --slot-save-path /home/laythe/llamapcache --jinja -kvu --no-mmproj --swa-checkpoints 32
cmd: llama-server --port ''${PORT} -hf unsloth/Qwen3.5-35B-A3B-GGUF:UD-IQ4_XS --ctx-size 128000 --temp 0.6 --top-p 0.95 --top-k 20 --min-p 0.00 -np 1 --fit-target 256 --ubatch-size 1024 -fa on --slots --slot-save-path /home/laythe/llamapcache --jinja -kvu --no-mmproj --swa-checkpoints 32 --no-kv-offload
ttl: 2400
"Qwen3-1.7B-GGUF":
cmd: llama-server --port ''${PORT} -hf unsloth/Qwen3-1.7B-GGUF:Q4_K_S --ctx-size 32768 --temp 0.6 --top-p 0.95 --top-k 20 --min-p 0.00 -fa on --jinja -kvu
ttl: 120
ttl: 300
"Qwen3-8B-GGUF":
cmd: llama-server --port ''${PORT} -hf unsloth/Qwen3-8B-GGUF:Q4_K_S --ctx-size 32768 --temp 0.6 --top-p 0.95 --top-k 20 --min-p 0.00 -fa on --jinja -kvu -np 1
ttl: 120
ttl: 300
"Qwen3-4B-Claude-Opus-Distill":
cmd: llama-server --port ''${PORT} -hf TeichAI/Qwen3-4B-Thinking-2507-Claude-4.5-Opus-High-Reasoning-Distill-GGUF:Q4_K_M --ctx-size 32768 --temp 0.6 --top-p 0.95 --top-k 20 --min-p 0.00 -fa on --jinja -kvu -np 1 --fit-target 256
ttl: 120
ttl: 300
"Qwen3.5-9B-Thinking":
cmd: "llama-server --port ''${PORT} -hf unsloth/Qwen3.5-9B-GGUF:Q4_K_M --ctx-size 32000 --temp 0.6 --top-p 0.95 --top-k 20 --min-p 0.00 -fa on --jinja -kvu -np 1 --fit-target 256 --chat-template-kwargs '{\"enable_thinking\": true}' --no-mmproj"
ttl: 120
cmd: "llama-server --port ''${PORT} -hf unsloth/Qwen3.5-9B-GGUF:Q4_K_M --ctx-size 128000 --temp 0.6 --top-p 0.95 --top-k 20 --min-p 0.00 -fa on --jinja -kvu -np 1 --fit-target 256 --chat-template-kwargs '{\"enable_thinking\": true}' --no-mmproj --no-kv-offload"
ttl: 300
"Qwen3.5-9B-Claude-Opus-Distill":
cmd: "llama-server --port ''${PORT} -hf Jackrong/Qwen3.5-9B-Claude-4.6-Opus-Reasoning-Distilled-GGUF:Q4_K_S --ctx-size 32000 --temp 0.6 --top-p 0.95 --top-k 20 --min-p 0.00 -fa on --jinja -kvu -np 1 --fit-target 512 --chat-template-kwargs '{\"enable_thinking\": true}' --no-mmproj"
ttl: 300
"Qwen3.5-4B-Thinking":
cmd: "llama-server --port ''${PORT} -hf unsloth/Qwen3.5-4B-GGUF:Q4_K_M --ctx-size 64000 --temp 0.6 --top-p 0.95 --top-k 20 --min-p 0.00 -fa on --jinja -kvu -np 1 --fit-target 256 --chat-template-kwargs '{\"enable_thinking\": true}'"
ttl: 120
ttl: 300
"Qwen3.5-9B-Non-Thinking":
cmd: "llama-server --port ''${PORT} -hf unsloth/Qwen3.5-9B-GGUF:Q4_K_M --ctx-size 32000 --temp 0.7 --top-p 0.8 --top-k 20 --min-p 0.00 -fa on --jinja -kvu -np 1 --fit-target 256 --no-mmproj"
ttl: 120
ttl: 300
"Qwen3.5-4B-Non-Thinking":
cmd: "llama-server --port ''${PORT} -hf unsloth/Qwen3.5-4B-GGUF:Q4_K_M --ctx-size 64000 --temp 0.7 --top-p 0.8 --top-k 20 --min-p 0.00 -fa on --jinja -kvu -np 1 --fit-target 256"
ttl: 120
ttl: 300
"Qwen3.5-0.8B-Non-Thinking":
cmd: "llama-server --port ''${PORT} -hf unsloth/Qwen3.5-0.8B-GGUF:Q4_K_M --ctx-size 64000 --temp 0.7 --top-p 0.8 --top-k 20 --min-p 0.00 -fa on --jinja -kvu -np 1 --fit-target 256"
ttl: 120'';
ttl: 300'';
# Set your time zone.
# time.timeZone = "Europe/Amsterdam";