feat: multi-GPU queue routing for legacy still renders (M2)

dispatch_order_line_render now routes still renders (non-animation,
non-cinematic) to a secondary GPU queue when MULTI_GPU_LIGHT_RENDER_QUEUE
is configured and that queue has active Celery workers.

- config.py: multi_gpu_light_render_queue setting (default "" = disabled)
- render_order_line.py: implement the routing stub — load output_type via
  selectinload, check render_settings.animation + .cinematic flags, call
  _inspect_active_worker_queues (reused from workflow_graph_runtime) with
  0.5s timeout to check if the light queue is live

No behaviour change when MULTI_GPU_LIGHT_RENDER_QUEUE is not set.
Enable by setting it to "asset_pipeline_light" and adding a
render-worker-light service with concurrency=1 to docker-compose.

docs: learnings erfasst — multi-GPU queue routing M2

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

LEARNINGS.md

@@ -7,6 +7,9 @@
 
 ## Learnings
 
+### 2026-07-22 | Architecture | Multi-GPU Queue Routing für Legacy-Dispatch implementiert
+M2: `dispatch_order_line_render` leitet Still-Renders an eine zweite GPU-Queue um wenn `MULTI_GPU_LIGHT_RENDER_QUEUE` gesetzt ist UND diese Queue aktive Worker hat (Celery inspect). Animationen und Cinematics bleiben immer auf `asset_pipeline`. Neue Settings-Option `multi_gpu_light_render_queue: str = ""` in config.py — Default leer = disabled. `_inspect_active_worker_queues()` aus `workflow_graph_runtime.py` wiederverwendet (timeout=0.5s). Kein Breaking-Change: ohne die Env-Variable verhält sich die Dispatch-Funktion exakt wie vorher.
+
 ### 2026-07-22 | Architecture | BLENDER_CINEMATIC Workflow-Graph-Node implementiert
 Der cinematic Render-Pfad hatte keinen eigenen Workflow-Graph-Node — Migration 071 hatte alle cinematic Output-Types auf `legacy_only` gezwungen als Sicherheitsnetz. M1 fügt jetzt `StepName.BLENDER_CINEMATIC` hinzu, zusammen mit: (1) Node-Definition im `workflow_node_registry.py` mit denselben Szene/Camera/Material-Feldern wie BLENDER_STILL, aber ohne Animations-Params (frame_count/fps sind im cinematic_render.py-Script hartkodiert auf 250 @ 25fps), (2) `render_cinematic_task` in `tasks.py` — folgt dem Pattern von `render_order_line_still_task`, gibt mp4 aus, nutzt `_finalize_graph_turntable_output`/`_finalize_shadow_turntable_output` da cinematic = mp4, published mit `asset_type="turntable"`, (3) STEP_TASK_MAP + `_ORDER_LINE_RENDER_STEPS` + `_build_task_kwargs` + `_predict_render_output_artifact` + `_artifact_kind_override_for_step` in `workflow_graph_runtime.py` alle aktualisiert. Kein neues DB-Migration nötig: Admins können cinematic Output-Types jetzt manuell von `legacy_only` auf `graph` umstellen und ein Workflow-Definition mit BLENDER_CINEMATIC-Node zuweisen.
 

backend/app/config.py

@@ -75,7 +75,14 @@ class Settings(BaseSettings):
 
     # Redis / Celery
     redis_url: str = "redis://localhost:6379/0"
+    # Queue for shadow-mode workflow renders (second GPU worker).
     workflow_shadow_render_queue: str = "asset_pipeline_light"
+    # When non-empty AND that queue has active workers, still renders from the
+    # legacy dispatch path are routed here instead of asset_pipeline, enabling
+    # concurrent still rendering on multi-GPU setups.
+    # Set MULTI_GPU_LIGHT_RENDER_QUEUE=asset_pipeline_light in docker-compose
+    # for the render-worker-light service and restart the workers.
+    multi_gpu_light_render_queue: str = ""
 
     @model_validator(mode="after")
     def normalize_runtime_hosts(self) -> "Settings":

backend/app/domains/pipeline/tasks/render_order_line.py

@@ -42,13 +42,32 @@ def dispatch_order_line_render(order_line_id: str):
                 logger.info(f"OrderLine {order_line_id}: order {order.status.value} — not dispatching")
                 return
 
-        # All renders go to asset_pipeline (single-GPU default).
-        # For multi-GPU setups: enable render-worker-light in docker-compose
-        # and change target_queue logic below to route small stills to
-        # asset_pipeline_light for concurrent rendering.
-        pass
+        # Multi-GPU routing: route still renders to a secondary GPU queue when
+        # MULTI_GPU_LIGHT_RENDER_QUEUE is configured and that queue is active.
+        is_still_render = False
+        if line:
+            from sqlalchemy.orm import selectinload
+            line_full = session.execute(
+                select(OrderLine)
+                .options(selectinload(OrderLine.output_type))
+                .where(OrderLine.id == order_line_id)
+            ).scalar_one_or_none()
+            if line_full and line_full.output_type:
+                rs = line_full.output_type.render_settings or {}
+                is_still_render = not rs.get("animation") and not rs.get("cinematic")
 
+    light_queue = app_settings.multi_gpu_light_render_queue.strip()
     target_queue = "asset_pipeline"
+    if light_queue and is_still_render:
+        from app.domains.rendering.workflow_graph_runtime import _inspect_active_worker_queues
+        active_queues = _inspect_active_worker_queues(timeout=0.5)
+        if light_queue in active_queues:
+            target_queue = light_queue
+            logger.info(
+                "Multi-GPU routing: order_line %s (still) -> queue=%s",
+                order_line_id,
+                target_queue,
+            )
     logger.info(f"Dispatching render for order line: {order_line_id} -> queue={target_queue}")
     render_order_line_task.apply_async(args=[order_line_id], queue=target_queue)