Add tested 256K Ultra profile

compose.yaml

@@ -222,6 +222,77 @@ services:
       - --spec-draft-type-v
       - q4_0
 
+  # Text-only long-context profile. This exact IQ4-MIX / 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:
+    <<: *llama-common
+    container_name: mike-ai-llama-ultra
+    labels:
+      com.mike-ai.llama-profile: ultra
+    environment:
+      NVIDIA_VISIBLE_DEVICES: ${ULTRA_GPU_DEVICES:-0,1}
+      NVIDIA_DRIVER_CAPABILITIES: compute,utility
+    command:
+      - --model
+      - "/models/${ULTRA_MODEL_FILE:?ULTRA_MODEL_FILE is required}"
+      - --alias
+      - qwen-ultra
+      - --ctx-size
+      - "${ULTRA_CONTEXT:-262144}"
+      - --flash-attn
+      - "on"
+      - --cache-type-k
+      - q4_0
+      - --cache-type-v
+      - q4_0
+      - --threads
+      - "${LLAMA_THREADS:-6}"
+      - --threads-batch
+      - "${LLAMA_THREADS_BATCH:-6}"
+      - --batch-size
+      - "64"
+      - --ubatch-size
+      - "32"
+      - --parallel
+      - "1"
+      - --jinja
+      - --reasoning
+      - auto
+      - --host
+      - 0.0.0.0
+      - --port
+      - "8080"
+      - --metrics
+      - --fit
+      - "off"
+      - --n-gpu-layers
+      - all
+      - --no-mmap
+      - --no-ui
+      - --temperature
+      - "0.2"
+      - --top-p
+      - "0.8"
+      - --top-k
+      - "20"
+      - --device
+      - CUDA0,CUDA1
+      - --main-gpu
+      - "0"
+      - --split-mode
+      - layer
+      - --tensor-split
+      - "${ULTRA_TENSOR_SPLIT:-80,20}"
+      - --spec-type
+      - draft-mtp
+      - --spec-draft-n-max
+      - "2"
+      - --spec-draft-type-k
+      - f16
+      - --spec-draft-type-v
+      - f16
+
   llama-experimental:
     <<: *llama-common
     container_name: mike-ai-llama-experimental
@@ -276,7 +347,7 @@ services:
       - /var/run/docker.sock:/var/run/docker.sock
     environment:
       CONTROLLER_TOKEN: "${CONTROLLER_TOKEN:?CONTROLLER_TOKEN is required}"
-      ALLOWED_PROFILES: fast,medium,long,experimental
+      ALLOWED_PROFILES: fast,medium,long,ultra,experimental
     networks: [control]
     security_opt: ["no-new-privileges:true"]
     healthcheck: