From b104dc46405633711bb35719052a0340b810a89e Mon Sep 17 00:00:00 2001
From: Preston Hunter <phunter@prestonhunter.space>
Date: Sat, 27 Dec 2025 14:26:56 -0500
Subject: [PATCH] Fixed eGPU deploy?

---
 stacks/ai/ai-backend.nomad | 27 +++++++++++++++------------
 1 file changed, 15 insertions(+), 12 deletions(-)

diff --git a/stacks/ai/ai-backend.nomad b/stacks/ai/ai-backend.nomad
index af5ebf0..0c0a45d 100644
--- a/stacks/ai/ai-backend.nomad
+++ b/stacks/ai/ai-backend.nomad
@@ -16,34 +16,37 @@ job "ai-backend" {
       port "api" { static = 11434 }
     }
 
-    task "ollama" {
+task "ollama" {
       driver = "podman"
-
+      
       env {
         OLLAMA_HOST    = "0.0.0.0"
         OLLAMA_ORIGINS = "*"
         
-        # CRITICAL FOR 6900XT:
-        # This tells ROCm to treat the card like a supported Pro workstation card
+        # Unlock the 6900XT (Navi 21) for ROCm
         HSA_OVERRIDE_GFX_VERSION = "10.3.0"
+        
+        # Debugging enabled so we can confirm it worked
+        OLLAMA_DEBUG = "1"
       }
 
       config {
-        # Standard image (contains ROCm libraries)
         image = "docker.io/ollama/ollama:latest"
         ports = ["api"]
-        
-        # Required for hardware access
         privileged = true
 
-        # Pass the graphics hardware to the container
+        # --- THE FIX: STRICT MAPPING ---
+        # Only map the eGPU (renderD128) and the Compute interface (kfd)
+        devices = [
+            "/dev/kfd",
+            "/dev/dri/renderD128"
+        ]
+
+        # Do NOT map the whole /dev/dri folder, or it might peek at the others
         volumes = [
-          "/mnt/local-ssd/nomad/stacks/ai/ai-backend/ollama:/root/.ollama",
-          "/dev/kfd:/dev/kfd",
-          "/dev/dri:/dev/dri"
+          "/mnt/local-ssd/nomad/stacks/ai/ai-backend/ollama:/root/.ollama"
         ]
       }
-
       service {
         name = "ollama"
         port = "api"