fix(ollama): restrict to RX 9070 XT, restore mmproj

- Set HIP_VISIBLE_DEVICES=0 to use only the discrete GPU (gfx1201).
  llama.cpp was trying to split layers across the iGPU (gfx1036) which
  caused segfaults when loading the multimodal projector.
- Restore --mmproj for both HF models (multimodal works correctly with
  single GPU).
- Keep qwen3.5:9b disabled (Ollama-extracted GGUF uses old mrope_sections
  key format incompatible with this llama.cpp build).

ollama/docker-compose.yml

@@ -8,6 +8,8 @@ services:
     volumes:
       - "/srv/docker/ollama/data/models:/models:ro"
       - "./llama-swap.yaml:/etc/llama-swap/config.yaml:ro"
+    environment:
+      - HIP_VISIBLE_DEVICES=0
     devices:
       - "/dev/kfd:/dev/kfd"
       - "/dev/dri:/dev/dri"

ollama/llama-swap.yaml

@@ -2,25 +2,10 @@ healthCheckTimeout: 180
 logLevel: info
 
 models:
-  "qwen3.5:9b":
+  # NOTE: qwen3.5:9b and qwen3.5:9bctxSmall are disabled — the GGUF extracted
-    cmd: |
+  # from Ollama uses the old mrope_sections format (3 elements) which this
-      /app/llama-server
+  # llama.cpp build rejects. Download a fresh quantization from HuggingFace
-      --host 0.0.0.0 --port ${PORT}
+  # (e.g. bartowski/Qwen3.5-9B-GGUF) and add them back.
-      --model /models/qwen3.5-9b.gguf
-      --alias qwen3.5:9b
-      --n-gpu-layers 999
-      --ctx-size 8192
-      --temp 1 --top-k 20 --top-p 0.95 --presence-penalty 1.5
-
-  "qwen3.5:9bctxSmall":
-    cmd: |
-      /app/llama-server
-      --host 0.0.0.0 --port ${PORT}
-      --model /models/qwen3.5-9b.gguf
-      --alias qwen3.5:9bctxSmall
-      --n-gpu-layers 999
-      --ctx-size 131072
-      --temp 1 --top-k 20 --top-p 0.95 --presence-penalty 1.5
 
   "hf.co/HauhauCS/Qwen3.5-9B-Uncensored-HauhauCS-Aggressive:q4_k_m":
     cmd: |