Add final Qwen3.8 runtime and uncensored profile

compose.yaml

@@ -76,6 +76,8 @@ services:
       - --mmproj
       - "/models/${VISION_PROJECTOR_FILE:?VISION_PROJECTOR_FILE is required}"
       - --mmproj-offload
+      - --mmproj-device
+      - CUDA1
       - --alias
       - qwen-fast
       - --ctx-size
@@ -146,6 +148,8 @@ services:
       - --mmproj
       - "/models/${VISION_PROJECTOR_FILE:?VISION_PROJECTOR_FILE is required}"
       - --mmproj-offload
+      - --mmproj-device
+      - CUDA1
       - --alias
       - qwen-medium
       - --ctx-size
@@ -202,6 +206,8 @@ services:
       - f16
       - --spec-draft-type-v
       - f16
+      - --spec-draft-p-min
+      - "0.05"
 
   llama-large:
     <<: *llama-common
@@ -218,6 +224,8 @@ services:
       - --mmproj
       - "/models/${VISION_PROJECTOR_FILE:?VISION_PROJECTOR_FILE is required}"
       - --mmproj-offload
+      - --mmproj-device
+      - CUDA1
       - --alias
       - qwen-large
       - --ctx-size
@@ -346,6 +354,85 @@ services:
       - --spec-draft-type-v
       - f16
 
+  # Deliberately less refusal-prone weight-level ablation. It remains behind
+  # the same authenticated router, tool permissions and confirmation guards as
+  # every other profile; "uncensored" never means unrestricted tool access.
+  llama-uncensored:
+    <<: *llama-common
+    container_name: mike-ai-llama-uncensored
+    labels:
+      com.mike-ai.llama-profile: uncensored
+    environment:
+      NVIDIA_VISIBLE_DEVICES: ${UNCENSORED_GPU_DEVICES:-0,1}
+      NVIDIA_DRIVER_CAPABILITIES: compute,utility
+      MTMD_BACKEND_DEVICE: CUDA1
+    command:
+      - --model
+      - "/models/${UNCENSORED_MODEL_FILE:?UNCENSORED_MODEL_FILE is required}"
+      - --mmproj
+      - "/models/${UNCENSORED_PROJECTOR_FILE:?UNCENSORED_PROJECTOR_FILE is required}"
+      - --mmproj-offload
+      - --mmproj-device
+      - CUDA1
+      - --alias
+      - qwen-uncensored
+      - --ctx-size
+      - "${UNCENSORED_CONTEXT:-80000}"
+      - --flash-attn
+      - "on"
+      - --cache-type-k
+      - q4_0
+      - --cache-type-v
+      - q4_0
+      - --threads
+      - "${LLAMA_THREADS:-6}"
+      - --threads-batch
+      - "${LLAMA_THREADS_BATCH:-6}"
+      - --batch-size
+      - "64"
+      - --ubatch-size
+      - "32"
+      - --parallel
+      - "1"
+      - --jinja
+      - --reasoning
+      - auto
+      - --host
+      - 0.0.0.0
+      - --port
+      - "8080"
+      - --metrics
+      - --fit
+      - "off"
+      - --n-gpu-layers
+      - all
+      - --no-mmap
+      - --no-ui
+      - --temperature
+      - "0.2"
+      - --top-p
+      - "0.8"
+      - --top-k
+      - "20"
+      - --device
+      - CUDA0,CUDA1
+      - --main-gpu
+      - "0"
+      - --split-mode
+      - layer
+      - --tensor-split
+      - "${UNCENSORED_TENSOR_SPLIT:-90,10}"
+      - --spec-type
+      - draft-mtp
+      - --spec-draft-n-max
+      - "${UNCENSORED_MTP_MAX:-2}"
+      - --spec-draft-p-min
+      - "0.10"
+      - --spec-draft-type-k
+      - f16
+      - --spec-draft-type-v
+      - f16
+
   llama-experimental:
     <<: *llama-common
     container_name: mike-ai-llama-experimental
@@ -400,7 +487,7 @@ services:
       - /var/run/docker.sock:/var/run/docker.sock
     environment:
       CONTROLLER_TOKEN: "${CONTROLLER_TOKEN:?CONTROLLER_TOKEN is required}"
-      ALLOWED_PROFILES: fast,medium,large,ultra,experimental
+      ALLOWED_PROFILES: fast,medium,large,ultra,uncensored,experimental
       IMAGE_WORKER: flux
     networks: [control]
     security_opt: ["no-new-privileges:true"]