Add Qwen GSQ-RCO Beta 1 profile

compose.yaml

@@ -235,6 +235,88 @@ services:
       - --spec-draft-p-min
       - "0.05"
 
+  # Beta 1 keeps the complete GSQ-RCO text model and KV cache on the RTX 5080.
+  # Only the multimodal projector runs on the RTX 3060. The measured hard
+  # boundary is 196608 tokens; 192K deliberately retains runtime headroom.
+  llama-beta1:
+    <<: *llama-common
+    container_name: mike-ai-llama-beta1
+    labels:
+      com.mike-ai.llama-profile: beta1
+    environment:
+      NVIDIA_VISIBLE_DEVICES: ${BETA1_GPU_DEVICES:-0,1}
+      NVIDIA_DRIVER_CAPABILITIES: compute,utility
+      MTMD_BACKEND_DEVICE: CUDA1
+    command:
+      - --model
+      - "/models/${BETA1_MODEL_FILE:?BETA1_MODEL_FILE is required}"
+      - --mmproj
+      - "/models/${VISION_PROJECTOR_FILE:?VISION_PROJECTOR_FILE is required}"
+      - --mmproj-offload
+      - --mmproj-device
+      - CUDA1
+      - --alias
+      - qwen-beta-1
+      - --ctx-size
+      - "${BETA1_CONTEXT:-192000}"
+      - --flash-attn
+      - "on"
+      - --cache-type-k
+      - q4_0
+      - --cache-type-v
+      - q4_0
+      - --cache-prompt
+      - --cache-ram
+      - "${LLAMA_CACHE_RAM_MIB:-32768}"
+      - --threads
+      - "${LLAMA_THREADS:-6}"
+      - --threads-batch
+      - "${LLAMA_THREADS_BATCH:-6}"
+      - --batch-size
+      - "${BETA1_BATCH_SIZE:-2048}"
+      - --ubatch-size
+      - "${BETA1_UBATCH_SIZE:-128}"
+      - --parallel
+      - "${BETA1_PARALLEL_SLOTS:-1}"
+      - --kv-unified
+      - --jinja
+      - --reasoning
+      - auto
+      - --reasoning-preserve
+      - --host
+      - 0.0.0.0
+      - --port
+      - "8080"
+      - --metrics
+      - --fit
+      - "off"
+      - --n-gpu-layers
+      - all
+      - --no-mmap
+      - --no-ui
+      - --temperature
+      - "1.0"
+      - --top-p
+      - "0.95"
+      - --top-k
+      - "20"
+      - --device
+      - CUDA0
+      - --main-gpu
+      - "0"
+      - --split-mode
+      - none
+      - --spec-type
+      - draft-mtp
+      - --spec-draft-n-max
+      - "3"
+      - --spec-draft-type-k
+      - f16
+      - --spec-draft-type-v
+      - f16
+      - --spec-draft-p-min
+      - "0.05"
+
   llama-large:
     <<: *llama-common
     container_name: mike-ai-llama-large
@@ -487,7 +569,7 @@ services:
       - /var/run/docker.sock:/var/run/docker.sock
     environment:
       CONTROLLER_TOKEN: "${CONTROLLER_TOKEN:?CONTROLLER_TOKEN is required}"
-      ALLOWED_PROFILES: fast,medium,large,ultra,uncensored
+      ALLOWED_PROFILES: fast,medium,beta1,large,ultra,uncensored
       IMAGE_WORKER: image
     networks: [control]
     security_opt: ["no-new-privileges:true"]