switch llama-swap to unified-vulkan image with FLUX.1-dev image generation

- Update deployment to unified-vulkan-2026-05-19 (includes llama-server,
  sd-server, whisper-server in one image)
- Fix binary paths: /app/llama-server -> llama-server (now on PATH)
- Migrate groups -> matrix to allow FLUX to evict the always-on 0.8B model
  when image generation is requested
- Add FLUX.1-dev Q4_K_S model via sd-server
- Configure OpenWebUI image generation to use llama-swap sd-server
- Update renovate versioning regex to treat all unified-vulkan date tags as
  patch updates for automerge

apps/llama/configs/config.yaml

@@ -19,18 +19,46 @@ hooks:
     preload:
       - "Qwen3.5-0.8B-GGUF-nothink:Q4_K_XL"
 
-groups:
+# matrix replaces groups (they are mutually exclusive).
-  always:
+# The small 0.8B model runs alongside any LLM.
-    persistent: true
+# FLUX runs alone — it needs all available VRAM and will evict the 0.8B first.
-    exclusive: false
+matrix:
-    swap: false
+  vars:
-    members:
+    q8: "Qwen3.5-0.8B-GGUF-nothink:Q4_K_XL"
-      - "Qwen3.5-0.8B-GGUF-nothink:Q4_K_XL"
+    flux: "flux1-dev:Q4_K_S"
+    coder: "Qwen3-Coder-Next-GGUF:Q4_K_M"
+    q35t: "Qwen3.5-35B-A3B-GGUF:Q4_K_M"
+    q35nt: "Qwen3.5-35B-A3B-GGUF-nothink:Q4_K_M"
+    q35ht: "Qwen3.5-35B-A3B-heretic-GGUF:Q4_K_M"
+    q35hnt: "Qwen3.5-35B-A3B-heretic-GGUF-nothink:Q4_K_M"
+    q4t: "Qwen3.5-4B-GGUF:Q4_K_M"
+    q4nt: "Qwen3.5-4B-GGUF-nothink:Q4_K_M"
+    q4ht: "Qwen3.5-4B-heretic-GGUF:Q4_K_M"
+    q4hnt: "Qwen3.5-4B-heretic-GGUF-nothink:Q4_K_M"
+    g26xl: "gemma-4-26B-A4B-it:UD-Q4_K_XL"
+    g26q2: "gemma-4-26B-A4B-it:UD-Q2_K_XL"
+    ge4xl: "unsloth/gemma-4-E4B-it-GGUF:UD-Q4_K_XL"
+    ge2xl: "unsloth/gemma-4-E2B-it-GGUF:UD-Q4_K_XL"
+    q36t: "unsloth/Qwen3.6-35B-A3B-GGUF:UD-Q4_K_XL"
+    q36nt: "unsloth/Qwen3.6-35B-A3B-GGUF-nothink:UD-Q4_K_XL"
+    haut: "HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive:Q4_K_M"
+    haunt: "HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-nothink:Q4_K_M"
+    mtpt: "unsloth/Qwen3.6-35B-A3B-MTP-GGUF:Q4_K_M"
+    mtpnt: "unsloth/Qwen3.6-35B-A3B-MTP-GGUF-nothink:Q4_K_M"
+
+  evict_costs:
+    flux: 10  # large files, slow to reload
+
+  sets:
+    # any LLM can run alongside the small always-on model
+    with_q8: "(coder | q35t | q35nt | q35ht | q35hnt | q4t | q4nt | q4ht | q4hnt | g26xl | g26q2 | ge4xl | ge2xl | q36t | q36nt | haut | haunt | mtpt | mtpnt) & q8"
+    # FLUX runs alone — evicts everything including q8
+    image_gen: "flux"
 
 models:
   "Qwen3-Coder-Next-GGUF:Q4_K_M":
     cmd: |
-      /app/llama-server
+      llama-server
         -hf unsloth/Qwen3-Coder-Next-GGUF:Q4_K_M
         --ctx-size 65536
         --predict 8192
@@ -44,7 +72,7 @@ models:
 
   "Qwen3.5-35B-A3B-GGUF:Q4_K_M":
     cmd: |
-      /app/llama-server
+      llama-server
         -hf unsloth/Qwen3.5-35B-A3B-GGUF:Q4_K_M
         ${ctx_256k}
         ${qwen35_think_args}
@@ -52,7 +80,7 @@ models:
 
   "Qwen3.5-35B-A3B-GGUF-nothink:Q4_K_M":
     cmd: |
-      /app/llama-server
+      llama-server
         -hf unsloth/Qwen3.5-35B-A3B-GGUF:Q4_K_M
         ${ctx_256k}
         ${qwen35_nothink_args}
@@ -62,7 +90,7 @@ models:
   # so providing url to the one from the non-heretic version.
   "Qwen3.5-35B-A3B-heretic-GGUF:Q4_K_M":
     cmd: |
-      /app/llama-server
+      llama-server
         -hf mradermacher/Qwen3.5-35B-A3B-heretic-GGUF:Q4_K_M
         ${qwen35_35b_heretic_mmproj}
         ${ctx_256k}
@@ -71,7 +99,7 @@ models:
 
   "Qwen3.5-35B-A3B-heretic-GGUF-nothink:Q4_K_M":
     cmd: |
-      /app/llama-server
+      llama-server
         -hf mradermacher/Qwen3.5-35B-A3B-heretic-GGUF:Q4_K_M
         ${qwen35_35b_heretic_mmproj}
         ${ctx_256k}
@@ -80,7 +108,7 @@ models:
 
   "Qwen3.5-0.8B-GGUF-nothink:Q4_K_XL":
     cmd: |
-      /app/llama-server
+      llama-server
         -hf unsloth/Qwen3.5-0.8B-GGUF:Q4_K_XL
         --ctx-size 4096
         ${qwen35_nothink_args}
@@ -88,7 +116,7 @@ models:
 
   "Qwen3.5-4B-GGUF:Q4_K_M":
     cmd: |
-      /app/llama-server
+      llama-server
         -hf unsloth/Qwen3.5-4B-GGUF:Q4_K_M
         ${ctx_128k}
         ${qwen35_think_args}
@@ -96,7 +124,7 @@ models:
 
   "Qwen3.5-4B-GGUF-nothink:Q4_K_M":
     cmd: |
-      /app/llama-server
+      llama-server
         -hf unsloth/Qwen3.5-4B-GGUF:Q4_K_M
         ${ctx_128k}
         ${qwen35_nothink_args}
@@ -104,7 +132,7 @@ models:
 
   "Qwen3.5-4B-heretic-GGUF:Q4_K_M":
     cmd: |
-      /app/llama-server
+      llama-server
         -hf mradermacher/Qwen3.5-4B-heretic-GGUF:Q4_K_M
         ${qwen35_4b_heretic_mmproj}
         ${ctx_128k}
@@ -113,7 +141,7 @@ models:
 
   "Qwen3.5-4B-heretic-GGUF-nothink:Q4_K_M":
     cmd: |
-      /app/llama-server
+      llama-server
         -hf mradermacher/Qwen3.5-4B-heretic-GGUF:Q4_K_M
         ${qwen35_4b_heretic_mmproj}
         ${ctx_128k}
@@ -122,7 +150,7 @@ models:
 
   "gemma-4-26B-A4B-it:UD-Q4_K_XL":
     cmd: |
-      /app/llama-server
+      llama-server
         -hf unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q4_K_XL \
         ${ctx_256k}
         ${gemma4_sampling}
@@ -130,7 +158,7 @@ models:
 
   "gemma-4-26B-A4B-it:UD-Q2_K_XL":
     cmd: |
-      /app/llama-server
+      llama-server
         -hf unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q2_K_XL \
         ${ctx_256k}
         ${gemma4_sampling}
@@ -138,7 +166,7 @@ models:
 
   "unsloth/gemma-4-E4B-it-GGUF:UD-Q4_K_XL":
     cmd: |
-      /app/llama-server
+      llama-server
         -hf unsloth/gemma-4-E4B-it-GGUF:UD-Q4_K_XL \
         ${ctx_128k}
         ${gemma4_sampling}
@@ -146,7 +174,7 @@ models:
 
   "unsloth/gemma-4-E2B-it-GGUF:UD-Q4_K_XL":
     cmd: |
-      /app/llama-server
+      llama-server
         -hf unsloth/gemma-4-E2B-it-GGUF:UD-Q4_K_XL \
         ${ctx_128k}
         ${gemma4_sampling}
@@ -154,7 +182,7 @@ models:
 
   "unsloth/Qwen3.6-35B-A3B-GGUF:UD-Q4_K_XL":
     cmd: |
-      /app/llama-server
+      llama-server
       -hf unsloth/Qwen3.6-35B-A3B-GGUF:UD-Q4_K_XL
       ${ctx_256k}
       ${qwen35_think_args}
@@ -162,7 +190,7 @@ models:
 
   "unsloth/Qwen3.6-35B-A3B-GGUF-nothink:UD-Q4_K_XL":
     cmd: |
-      /app/llama-server
+      llama-server
       -hf unsloth/Qwen3.6-35B-A3B-GGUF:UD-Q4_K_XL
       ${ctx_256k}
       ${qwen35_nothink_args}
@@ -170,7 +198,7 @@ models:
 
   "HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive:Q4_K_M":
     cmd: |
-      /app/llama-server
+      llama-server
       -hf HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive:Q4_K_M
       ${ctx_256k}
       ${qwen35_think_args}
@@ -178,7 +206,7 @@ models:
 
   "HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-nothink:Q4_K_M":
     cmd: |
-      /app/llama-server
+      llama-server
       -hf HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive:Q4_K_M
       ${ctx_256k}
       ${qwen35_nothink_args}
@@ -186,7 +214,7 @@ models:
 
   "unsloth/Qwen3.6-35B-A3B-MTP-GGUF:Q4_K_M":
     cmd: |
-      /app/llama-server
+      llama-server
       -hf unsloth/Qwen3.6-35B-A3B-MTP-GGUF:Q4_K_M
       ${ctx_256k}
       ${qwen35_think_args}
@@ -196,10 +224,29 @@ models:
 
   "unsloth/Qwen3.6-35B-A3B-MTP-GGUF-nothink:Q4_K_M":
     cmd: |
-      /app/llama-server
+      llama-server
       -hf unsloth/Qwen3.6-35B-A3B-MTP-GGUF:Q4_K_M
       ${ctx_256k}
       ${qwen35_nothink_args}
       --spec-type draft-mtp --spec-draft-n-max 1
       --parallel 1
       ${common_args}
+
+  # Image generation via stable-diffusion.cpp (sd-server)
+  # Models must be pre-downloaded to /root/.cache/sd/
+  # FLUX.1-dev: state-of-the-art open-weight text-to-image model by Black Forest Labs
+  # Download: huggingface-cli download lllyasviel/FLUX.1-dev-gguf flux1-dev-Q4_K_S.gguf --local-dir /root/.cache/sd
+  # Download VAE: huggingface-cli download black-forest-labs/FLUX.1-dev ae.safetensors --local-dir /root/.cache/sd
+  # Download text encoders: huggingface-cli download comfyanonymous/flux_text_encoders clip_l.safetensors t5xxl_fp16.safetensors --local-dir /root/.cache/sd
+  "flux1-dev:Q4_K_S":
+    checkEndpoint: "/"
+    cmd: |
+      sd-server
+        --listen-port ${PORT}
+        --diffusion-model /root/.cache/sd/flux1-dev-Q4_K_S.gguf
+        --vae /root/.cache/sd/ae.safetensors
+        --clip_l /root/.cache/sd/clip_l.safetensors
+        --t5xxl /root/.cache/sd/t5xxl_fp16.safetensors
+        --cfg-scale 1.0
+        --sampling-method euler
+        --steps 20

apps/llama/deployment.yaml

@@ -18,10 +18,10 @@ spec:
     spec:
       containers:
         - name: llama-swap
-          image: ghcr.io/mostlygeek/llama-swap:v216-vulkan-b9191
+          image: ghcr.io/mostlygeek/llama-swap:unified-vulkan-2026-05-19
           imagePullPolicy: IfNotPresent
           command:
-            - /app/llama-swap
+            - llama-swap
           args:
             - --config=/config/config.yaml
             - --watch-config

apps/openwebui/release.yaml

@@ -74,3 +74,16 @@ spec:
         value: "false"
       - name: OAUTH_MERGE_ACCOUNTS_BY_EMAIL
         value: "true"
+      # Image generation via llama-swap sd-server
+      - name: ENABLE_IMAGE_GENERATION
+        value: "true"
+      - name: IMAGE_GENERATION_ENGINE
+        value: "openai"
+      - name: IMAGES_OPENAI_API_BASE_URL
+        value: "http://llama.llama.svc.cluster.local:11434/v1"
+      - name: IMAGES_OPENAI_API_KEY
+        value: "ignored"
+      - name: IMAGE_GENERATION_MODEL
+        value: "flux1-dev:Q4_K_S"
+      - name: IMAGE_SIZE
+        value: "512x512"

renovate.json

@@ -15,21 +15,13 @@
     {
       "matchDatasources": ["docker"],
       "matchPackageNames": ["ghcr.io/mostlygeek/llama-swap"],
-      "versioning": "regex:^v(?<major>[0-9]+)-vulkan-b(?<patch>[0-9]+)$"
+      "versioning": "regex:^unified-vulkan-(?<patch>[0-9]{4}[0-9]{2}[0-9]{2})$"
     },
     {
       "matchUpdateTypes": ["patch"],
       "automerge": true,
       "automergeType": "pr",
       "platformAutomerge": true
-    },
-    {
-      "matchDatasources": ["docker"],
-      "matchPackageNames": ["ghcr.io/mostlygeek/llama-swap"],
-      "matchUpdateTypes": ["minor"],
-      "automerge": true,
-      "automergeType": "pr",
-      "platformAutomerge": true
     }
   ]
 }