revert flake
This commit is contained in:
@@ -4,16 +4,16 @@
|
||||
}:
|
||||
{
|
||||
# Activer Ollama comme service
|
||||
services.ollama = {
|
||||
enable = true;
|
||||
acceleration = "cuda";
|
||||
environmentVariables = {
|
||||
OLLAMA_FLASH_ATTENTION = "1"; # ← Flash Attention
|
||||
OLLAMA_NUM_PARALLEL = "2"; # ← Requêtes parallèles
|
||||
OLLAMA_MAX_LOADED_MODELS = "1"; # ← Garder 2 modèles en VRAM
|
||||
OLLAMA_KEEP_ALIVE = "5m";
|
||||
};
|
||||
};
|
||||
# services.ollama = {
|
||||
# enable = true;
|
||||
# acceleration = "cuda";
|
||||
# environmentVariables = {
|
||||
# OLLAMA_FLASH_ATTENTION = "1"; # ← Flash Attention
|
||||
# OLLAMA_NUM_PARALLEL = "2"; # ← Requêtes parallèles
|
||||
# OLLAMA_MAX_LOADED_MODELS = "1"; # ← Garder 2 modèles en VRAM
|
||||
# OLLAMA_KEEP_ALIVE = "5m";
|
||||
# };
|
||||
# };
|
||||
# services.open-webui = {
|
||||
# enable = true;
|
||||
# port = 8080; # Port par défaut
|
||||
|
||||
Reference in New Issue
Block a user