sdfsf

stacks/ai/ai-backend.nomad

@@ -6,7 +6,6 @@ job "ai-backend" {
   group "ollama-group" {
     count = 1
 
-    # Pin to P52 Laptop (eGPU Host)
     constraint {
       attribute = "${meta.device}"
       value     = "p52-laptop"
@@ -22,44 +21,34 @@ job "ai-backend" {
       env {
         OLLAMA_HOST    = "0.0.0.0"
         OLLAMA_ORIGINS = "*"
-        
-        # 1. Force 6900XT Support
-        HSA_OVERRIDE_GFX_VERSION = "10.3.0"
-        
-        # 2. Debugging
         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"]
-        
-        # Required to talk to hardware (This handles most security opts)
         privileged = true
 
-        # --- Explicit Device Mapping ---
+        # --- MOUNT EVERYTHING ---
-        # Map the Compute interface and the Physical Card
+        # Since Vulkan scans all GPUs to pick the best one, 
-        devices = [
+        # we give it access to the whole folder.
-            "/dev/kfd",
-            "/dev/dri/card1",
-            "/dev/dri/renderD128"
-        ]
-
-        # --- Volumes ---
         volumes = [
-          # 1. Your Custom Storage Path
           "/mnt/local-ssd/nomad/stacks/ai/ai-backend/ollama:/root/.ollama",
-          
+          "/dev/kfd:/dev/kfd",
-          # 2. Shared Memory Workaround (Replaces ipc_mode = "host")
+          "/dev/dri:/dev/dri"
-          # This helps the AMD driver communicate efficiently
-          "/dev/shm:/dev/shm"
         ]
       }
 
       service {
         name = "ollama"
         port = "api"
-        
         check {
           type     = "http"
           path     = "/"
@@ -70,7 +59,7 @@ job "ai-backend" {
       
       resources {
         cpu    = 2000
-        memory = 8192 # 8GB System RAM
+        memory = 8192
       }
     }
   }