66 lines
1.4 KiB
HCL
66 lines
1.4 KiB
HCL
job "ai-backend" {
|
|
datacenters = ["Homelab-PTECH-DC"]
|
|
region = "home"
|
|
type = "service"
|
|
|
|
group "ollama-group" {
|
|
count = 1
|
|
|
|
constraint {
|
|
attribute = "${meta.device}"
|
|
value = "p52-laptop"
|
|
}
|
|
|
|
network {
|
|
port "api" { static = 11434 }
|
|
}
|
|
|
|
task "ollama" {
|
|
driver = "podman"
|
|
|
|
env {
|
|
OLLAMA_HOST = "0.0.0.0"
|
|
OLLAMA_ORIGINS = "*"
|
|
OLLAMA_DEBUG = "1"
|
|
|
|
# --- THE FIX: USE VULKAN INSTEAD OF ROCm ---
|
|
OLLAMA_VULKAN = "1"
|
|
|
|
# We keep this just in case Vulkan falls back to ROCm,
|
|
# but Vulkan should take priority.
|
|
HSA_OVERRIDE_GFX_VERSION = "10.3.0"
|
|
}
|
|
|
|
config {
|
|
image = "docker.io/ollama/ollama:latest"
|
|
ports = ["api"]
|
|
privileged = true
|
|
|
|
# --- MOUNT EVERYTHING ---
|
|
# Since Vulkan scans all GPUs to pick the best one,
|
|
# we give it access to the whole folder.
|
|
volumes = [
|
|
"/mnt/local-ssd/nomad/stacks/ai/ai-backend/ollama:/root/.ollama",
|
|
"/dev/kfd:/dev/kfd",
|
|
"/dev/dri:/dev/dri"
|
|
]
|
|
}
|
|
|
|
service {
|
|
name = "ollama"
|
|
port = "api"
|
|
check {
|
|
type = "http"
|
|
path = "/"
|
|
interval = "20s"
|
|
timeout = "2s"
|
|
}
|
|
}
|
|
|
|
resources {
|
|
cpu = 2000
|
|
memory = 8192
|
|
}
|
|
}
|
|
}
|
|
} |