feat(qwen-image-bench): replace qwen3.6-35b-a3b on GPU1 with the T2I judge (NVFP4)

flukethoughts/Qwen-Image-Bench-NVFP4 — Qwen's text-to-image quality JUDGE (vision-intact,
NVFP4) on ana-ml2 GPU 1, replacing qwen3.6-35b-a3b:
- stacks/qwen-image-bench/ — vLLM multimodal serve (Qwen3_5ForConditionalGeneration, no
  text-only patch — vision wanted), GPU1 device pin, :8014. util 0.32 (0.22 KV-starved →
  crash-loop "no available memory for cache blocks"; util is a fraction of TOTAL so it
  must clear the ~20GB weight floor).
- litellm: removed qwen3.6-35b-a3b + -thinking; added qwen-image-bench + qwen-image-judge alias.

Verified live: healthy (KV 9.4GB / 133K tokens), text + IMAGE (vision path) both respond.
NOTE: arbo's hero-judge was bound to qwen3.6-35b-a3b — comfy-dev notified to repoint.

stacks/litellm/conf/config.yaml

@@ -41,40 +41,25 @@ model_list:
     model_info:
       mode: chat
 
-  # --- Qwen3.6-35B-A3B vision-language MoE (official FP8) — vision + chat. vLLM
+  # --- Qwen-Image-Bench (NVFP4) — text-to-image quality JUDGE, vision-intact. vLLM
-  #     on ana-ml2 GPU 1, :8007. Explicit entry shadows the "*" wildcard llama-swap
+  #     on ana-ml2 GPU 1, :8014. REPLACED qwen3.6-35b-a3b (operator 2026-06-19; the
-  #     route. REPLACED qwen3.5-9b-fp8 2026-06-14 (the 9B is retired; this is a
+  #     qwen36 VL judge was downed + its FP8 weights freed ~33GB). Multimodal — image
-  #     35B-A3B MoE — served under its TRUE name, never aliased under the old one).
+  #     + prompt in, a 5-dim verdict out (overall quality / prompt match / aesthetic /
-  #
+  #     LoRA activation / confidence). qwen-image-judge is a duplicate-entry alias.
-  #     THINKING SPLIT (2026-06-15, operator call — mirrors the glm-5.1 pattern
+  #     ⚠️ arbo's hero-judge was bound to qwen3.6-35b-a3b — comfy-dev notified to
-  #     above). One hybrid checkpoint; the per-request `enable_thinking` switch
+  #     repoint (different I/O than the old general VL judge). ---
-  #     picks the mode. LiteLLM forwards extra_body verbatim to vLLM, where
+  - model_name: qwen-image-bench
-  #     chat_template_kwargs lands in the chat template. vLLM runs
-  #     --reasoning-parser qwen3 so reasoning surfaces as reasoning_content. ---
-  # qwen3.6-35b-a3b: thinking DISABLED by default. The checkpoint defaults
-  #   thinking ON; enable_thinking=false forces the empty <think></think> block.
-  #   Reasoning is opt-in via qwen3.6-35b-a3b-thinking below.
-  - model_name: qwen3.6-35b-a3b
     litellm_params:
-      model: hosted_vllm/qwen3.6-35b-a3b
+      model: hosted_vllm/qwen-image-bench
-      api_base: http://10.250.50.54:8007/v1
+      api_base: http://10.250.50.54:8014/v1
       api_key: os.environ/VLLM_API_KEY
-      extra_body:
-        chat_template_kwargs:
-          enable_thinking: false
     model_info:
       mode: chat
-  # qwen3.6-35b-a3b-thinking: identical upstream checkpoint, thinking ENABLED
+  - model_name: qwen-image-judge
-  #   (opt-in reasoning). The qwen3 reasoning-parser splits <think>…</think> into
-  #   reasoning_content; content holds just the answer.
-  - model_name: qwen3.6-35b-a3b-thinking
     litellm_params:
-      model: hosted_vllm/qwen3.6-35b-a3b
+      model: hosted_vllm/qwen-image-bench
-      api_base: http://10.250.50.54:8007/v1
+      api_base: http://10.250.50.54:8014/v1
       api_key: os.environ/VLLM_API_KEY
-      extra_body:
-        chat_template_kwargs:
-          enable_thinking: true
     model_info:
       mode: chat
 

stacks/qwen-image-bench/.env.example

@@ -0,0 +1,20 @@
+# qwen-image-bench (flukethoughts/Qwen-Image-Bench-NVFP4) — ana-ml2 GPU 1 tunables.
+# Real .env lives at /opt/docker/compose/qwen-image-bench/.env.
+
+QIB_IMAGE=vllm/vllm-openai:latest
+QIB_CONTAINER_NAME=vllm-qwen-image-bench
+
+# Own port (8007=qwen36 [downed], 8010=qwen-gen, 8011=selene — 8014 free).
+QIB_PORT=8014
+QIB_GPU_ID=1
+
+# Shared GPU 1 (granite/selene/embed/rerank/reward + this). util is a fraction of
+# TOTAL (96GB), so it must clear the ~20GB weight floor + leave KV: 0.22 (~21GB) left
+# ~nothing for KV and crash-looped ("no available memory for cache blocks"). 0.32
+# (~31GB → ~9GB KV) fits the ~37GB freed by downing qwen36 (cf. qwen36's own 0.34).
+QIB_GPU_MEM_UTIL=0.32
+QIB_MAX_MODEL_LEN=32768
+QIB_MAX_NUM_SEQS=8
+
+# Optional upstream vLLM API key (empty = no auth; internal net only).
+API_KEY=

stacks/qwen-image-bench/compose.yaml

@@ -0,0 +1,84 @@
+# qwen-image-bench — flukethoughts/Qwen-Image-Bench-NVFP4 on ana-ml2 GPU 1,
+# REPLACING qwen3.6-35b-a3b (operator 2026-06-19). Qwen's text-to-image quality
+# JUDGE model (vision-language, NVFP4 weights / vision tower bf16). NOT generative —
+# it scores T2I outputs on 5 dims (overall quality, prompt match, aesthetic, LoRA
+# activation, confidence).
+#
+# Arch: Qwen3_5ForConditionalGeneration (dense Qwen3.5 hybrid SSM+attn + vision),
+# ~17B / ~20GB NVFP4. VISION-INTACT → served as multimodal; NO text-only patch
+# (unlike the qwen3.5-122b gen model, which had text-only weights). vLLM
+# production-validated per the model card.
+#
+# ⚠️ qwen3.6-35b-a3b was arbo's hero-judge (comfy-dev consumer). Downing it breaks
+# arbo's judging until comfy-dev repoints to qwen-image-bench (different I/O — a
+# 5-dim verdict vs a general VL judge). comfy-dev notified.
+#
+# DISPLACEMENT: GPU 1 is shared (granite/selene/embed/rerank/reward). qwen3.6 used
+# util 0.34 (~33GB); down it first, then this fits at util ~0.22 (~21GB). REVERT =
+# down this, `up -d` the qwen36 stack.
+#
+# Tunables in .env.
+
+name: qwen-image-bench
+
+services:
+  vllm-qwen-image-bench:
+    image: ${QIB_IMAGE:-vllm/vllm-openai:latest}
+    container_name: ${QIB_CONTAINER_NAME:-vllm-qwen-image-bench}
+    restart: unless-stopped
+    ipc: host
+    ports:
+      - "${QIB_PORT:-8014}:8000"
+    volumes:
+      - /tank/aimodels/huggingface:/hfcache
+      - /tank/aimodels/qwen-image-bench-nvfp4:/model:ro
+    environment:
+      - HF_HOME=/hfcache
+      - HF_HUB_CACHE=/hfcache/hub
+      - VLLM_API_KEY=${API_KEY:-}
+    command:
+      - /model
+      - --served-model-name
+      - qwen-image-bench
+      - --host
+      - 0.0.0.0
+      - --port
+      - "8000"
+      - --gpu-memory-utilization
+      - ${QIB_GPU_MEM_UTIL:-0.32}
+      - --max-model-len
+      - ${QIB_MAX_MODEL_LEN:-32768}
+      - --max-num-seqs
+      - ${QIB_MAX_NUM_SEQS:-8}
+      - --trust-remote-code
+      - --dtype
+      - auto
+      - --enable-prefix-caching
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              device_ids:
+                - "${QIB_GPU_ID:-1}"
+              capabilities:
+                - gpu
+    healthcheck:
+      test: ["CMD", "curl", "-f", "http://localhost:8000/health"]
+      interval: 30s
+      timeout: 10s
+      retries: 3
+      start_period: 600s
+    networks:
+      - tnet
+    labels:
+      - homepage.group=AI Systems
+      - homepage.name=Qwen-Image-Bench (T2I judge, NVFP4)
+      - homepage.icon=mdi-image-check
+      - homepage.description=Qwen text-to-image quality judge (NVFP4, vision-intact) on ana-ml2 GPU 1
+      - homepage.href=http://10.250.50.54:${QIB_PORT:-8014}/docs
+
+networks:
+  tnet:
+    name: traefik-net
+    external: true