Add IPFS HLS streaming and GPU optimizations

- Add IPFSHLSOutput class that uploads segments to IPFS as they're created
- Update streaming task to use IPFS HLS output for distributed streaming
- Add /ipfs-stream endpoint to get IPFS playlist URL
- Update /stream endpoint to redirect to IPFS when available
- Add GPU persistence mode (STREAMING_GPU_PERSIST=1) to keep frames on GPU
- Add hardware video decoding (NVDEC) support for faster video processing
- Add GPU-accelerated primitive libraries: blending_gpu, color_ops_gpu, geometry_gpu
- Add streaming_gpu module with GPUFrame class for tracking CPU/GPU data location
- Add Dockerfile.gpu for building GPU-enabled worker image

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

docker-compose.yml

@@ -3,6 +3,10 @@ version: "3.8"
 services:
   redis:
     image: redis:7-alpine
+    ports:
+      - target: 6379
+        published: 16379
+        mode: host  # Bypass swarm routing mesh
     volumes:
       - redis_data:/data
     networks:
@@ -11,6 +15,9 @@ services:
       replicas: 1
       restart_policy:
         condition: on-failure
+      placement:
+        constraints:
+          - node.labels.gpu != true
 
   postgres:
     image: postgres:16-alpine
@@ -18,6 +25,10 @@ services:
       - POSTGRES_USER=artdag
       - POSTGRES_PASSWORD=artdag
       - POSTGRES_DB=artdag
+    ports:
+      - target: 5432
+        published: 15432
+        mode: host  # Expose for GPU worker on different VPC
     volumes:
       - postgres_data:/var/lib/postgresql/data
     networks:
@@ -26,12 +37,18 @@ services:
       replicas: 1
       restart_policy:
         condition: on-failure
+      placement:
+        constraints:
+          - node.labels.gpu != true
 
   ipfs:
     image: ipfs/kubo:latest
     ports:
       - "4001:4001"      # Swarm TCP
       - "4001:4001/udp"  # Swarm UDP
+      - target: 5001
+        published: 15001
+        mode: host  # API port for GPU worker on different VPC
     volumes:
       - ipfs_data:/data/ipfs
       - l1_cache:/data/cache:ro  # Read-only access to cache for adding files
@@ -42,6 +59,9 @@ services:
       replicas: 1
       restart_policy:
         condition: on-failure
+      placement:
+        constraints:
+          - node.labels.gpu != true
 
   l1-server:
     image: git.rose-ash.com/art-dag/l1-server:latest
@@ -75,6 +95,9 @@ services:
       replicas: 1
       restart_policy:
         condition: on-failure
+      placement:
+        constraints:
+          - node.labels.gpu != true
 
   l1-worker:
     image: git.rose-ash.com/art-dag/l1-server:latest
@@ -102,6 +125,9 @@ services:
       replicas: 2
       restart_policy:
         condition: on-failure
+      placement:
+        constraints:
+          - node.labels.gpu != true
 
   flower:
     image: mher/flower:2.0
@@ -118,12 +144,60 @@ services:
       replicas: 1
       restart_policy:
         condition: on-failure
+      placement:
+        constraints:
+          - node.labels.gpu != true
+
+  # GPU worker for streaming/rendering tasks
+  # Build: docker build -f Dockerfile.gpu -t git.rose-ash.com/art-dag/l1-gpu-server:latest .
+  # Requires: docker node update --label-add gpu=true <gpu-node-name>
+  l1-gpu-worker:
+    image: git.rose-ash.com/art-dag/l1-gpu-server:latest
+    # For local dev, uncomment to build from Dockerfile.gpu:
+    # build:
+    #   context: .
+    #   dockerfile: Dockerfile.gpu
+    command: sh -c "cd /app && celery -A celery_app worker --loglevel=info -E -Q gpu,celery"
+    environment:
+      # GPU node is on different VPC - use public IPs for cross-node communication
+      - REDIS_URL=redis://138.68.142.139:16379/5
+      - DATABASE_URL=postgresql://artdag:artdag@138.68.142.139:15432/artdag
+      # Connect to shared IPFS node on CPU (via public IP)
+      - IPFS_API=/ip4/138.68.142.139/tcp/15001
+      # Gateway fallback for resilience
+      - IPFS_GATEWAYS=https://ipfs.io,https://cloudflare-ipfs.com,https://dweb.link
+      # Local cache is ephemeral (tmpfs or local volume)
+      - CACHE_DIR=/data/cache
+      - C_FORCE_ROOT=true
+      - ARTDAG_CLUSTER_KEY=${ARTDAG_CLUSTER_KEY:-}
+      # GPU acceleration settings
+      - NVIDIA_VISIBLE_DEVICES=all
+      # Keep frames on GPU between operations for maximum performance
+      - STREAMING_GPU_PERSIST=1
+    volumes:
+      # Local cache - ephemeral, just for working files
+      - gpu_cache:/data/cache
+      # Note: No source mount - GPU worker uses code from image
+    depends_on:
+      - redis
+      - postgres
+      - ipfs
+    networks:
+      - celery
+    deploy:
+      replicas: 1
+      restart_policy:
+        condition: on-failure
+      placement:
+        constraints:
+          - node.labels.gpu == true
 
 volumes:
   redis_data:
   postgres_data:
   ipfs_data:
   l1_cache:
+  gpu_cache:  # Ephemeral cache for GPU workers
 
 networks:
   celery: