Define four-profile production matrix with Medium default

compose.yaml

@@ -100,7 +100,7 @@ services:
     labels:
       com.mike-ai.llama-profile: medium
     environment:
-      NVIDIA_VISIBLE_DEVICES: ${MEDIUM_GPU_DEVICES:-0}
+      NVIDIA_VISIBLE_DEVICES: ${MEDIUM_GPU_DEVICES:-0,1}
       NVIDIA_DRIVER_CAPABILITIES: compute,utility
     command:
       - --model
@@ -111,7 +111,7 @@ services:
       - --alias
       - qwen-medium
       - --ctx-size
-      - "${MEDIUM_CONTEXT:-94208}"
+      - "${MEDIUM_CONTEXT:-160000}"
       - --flash-attn
       - "on"
       - --cache-type-k
@@ -149,28 +149,40 @@ services:
       - --top-k
       - "20"
       - --device
-      - CUDA0
+      - CUDA0,CUDA1
+      - --main-gpu
+      - "0"
       - --split-mode
-      - none
+      - layer
+      - --tensor-split
+      - "${MEDIUM_TENSOR_SPLIT:-90,10}"
+      - --spec-type
+      - draft-mtp
+      - --spec-draft-n-max
+      - "3"
+      - --spec-draft-type-k
+      - f16
+      - --spec-draft-type-v
+      - f16
 
-  llama-long:
+  llama-large:
     <<: *llama-common
-    container_name: mike-ai-llama-long
+    container_name: mike-ai-llama-large
     labels:
-      com.mike-ai.llama-profile: long
+      com.mike-ai.llama-profile: large
     environment:
-      NVIDIA_VISIBLE_DEVICES: ${LONG_GPU_DEVICES:-0}
+      NVIDIA_VISIBLE_DEVICES: ${LARGE_GPU_DEVICES:-0,1}
       NVIDIA_DRIVER_CAPABILITIES: compute,utility
     command:
       - --model
-      - "/models/${LONG_MODEL_FILE:?LONG_MODEL_FILE is required}"
+      - "/models/${LARGE_MODEL_FILE:?LARGE_MODEL_FILE is required}"
       - --mmproj
       - "/models/${VISION_PROJECTOR_FILE:?VISION_PROJECTOR_FILE is required}"
       - --no-mmproj-offload
       - --alias
-      - qwen-long
+      - qwen-large
       - --ctx-size
-      - "${LONG_CONTEXT:-131072}"
+      - "${LARGE_CONTEXT:-192000}"
       - --flash-attn
       - "on"
       - --cache-type-k
@@ -199,8 +211,6 @@ services:
       - "off"
       - --n-gpu-layers
       - all
-      - --override-tensor
-      - blk.([0-9]|1[0-1]).ffn_.*=CPU
       - --no-mmap
       - --no-ui
       - --temperature
@@ -210,19 +220,23 @@ services:
       - --top-k
       - "20"
       - --device
-      - CUDA0
+      - CUDA0,CUDA1
+      - --main-gpu
+      - "0"
       - --split-mode
-      - none
+      - layer
+      - --tensor-split
+      - "${LARGE_TENSOR_SPLIT:-86,14}"
       - --spec-type
       - draft-mtp
       - --spec-draft-n-max
-      - "2"
+      - "3"
       - --spec-draft-type-k
-      - q4_0
+      - f16
       - --spec-draft-type-v
-      - q4_0
+      - f16
 
-  # Text-only long-context profile. This exact IQ4_XS-pure / 256K / 80:20
+  # Text-only maximum-context profile. This exact IQ4_XS-pure / 256K / 80:20
   # combination completed the 220K fill test on RTX 5080 + RTX 3060.
   # Deliberately no vision projector: Ultra prioritizes maximum usable context.
   llama-ultra:
@@ -347,7 +361,7 @@ services:
       - /var/run/docker.sock:/var/run/docker.sock
     environment:
       CONTROLLER_TOKEN: "${CONTROLLER_TOKEN:?CONTROLLER_TOKEN is required}"
-      ALLOWED_PROFILES: fast,medium,long,ultra,experimental
+      ALLOWED_PROFILES: fast,medium,large,ultra,experimental
     networks: [control]
     security_opt: ["no-new-privileges:true"]
     healthcheck:
@@ -456,6 +470,7 @@ services:
       - ./platform/openwebui/theme/loader.js:/app/build/static/loader.js:ro
     environment:
       WEBUI_SECRET_KEY: "${WEBUI_SECRET_KEY:?WEBUI_SECRET_KEY is required}"
+      DEFAULT_MODELS: qwen-medium
       OLLAMA_BASE_URL: ""
       OPENAI_API_BASE_URLS: http://router:8081/v1
       OPENAI_API_KEYS: "${ROUTER_API_KEY:?ROUTER_API_KEY is required}"