Evan Carmen
7e4d22e41c
- Add model_management.py patch: detects GB10 unified memory (VRAM ≈ RAM > 0.95) - Set HIGH_VRAM mode: no pointless CPU offloading (same physical memory pool) - Increase maximum_vram_for_weights from 88% to 95% (8.4GB headroom on 128GB) - Skip torch.cuda.empty_cache() on unified memory (avoids page faults) - Return GPU for text_encoder/vae/intermediate offload devices on unified memory - MPS excluded from unified detection (has its own SHARED state) - Remove PYTORCH_NO_CUDA_MEMORY_CACHING env var (patch handles caching properly) - Mount patched file as read-only volume override in docker-compose.yml - DeepSeek review: safe and correct for DGX Spark target Co-authored-by: DeepSeek (code review)
132 lines
4.0 KiB
YAML
132 lines
4.0 KiB
YAML
services:
|
|
comfyui:
|
|
build:
|
|
context: .
|
|
dockerfile: Dockerfile
|
|
args:
|
|
# Pin ComfyUI to a known-good commit/tag if desired
|
|
COMFYUI_REF: "${COMFYUI_REF:-master}"
|
|
# SageAttention ref (e.g., "main", "v2.2.0", or specific commit)
|
|
SAGEATTN_REF: "${SAGEATTN_REF:-main}"
|
|
|
|
image: sparkyui:cu130
|
|
container_name: comfyui
|
|
|
|
# GPU enablement
|
|
deploy:
|
|
resources:
|
|
reservations:
|
|
devices:
|
|
- driver: nvidia
|
|
count: all
|
|
capabilities: [gpu]
|
|
|
|
# LAN exposure
|
|
ports:
|
|
- "${COMFYUI_PORT:-8188}:8188"
|
|
|
|
environment:
|
|
COMFYUI_PORT: "${COMFYUI_PORT:-8188}"
|
|
# Optimized for Grace-Blackwell unified memory architecture
|
|
# Key insight: DON'T use --gpu-only - let the unified memory fabric work naturally
|
|
COMFYUI_FLAGS: "${COMFYUI_FLAGS:---listen 0.0.0.0 --port 8188 --disable-pinned-memory --dont-upcast-attention}"
|
|
NVIDIA_VISIBLE_DEVICES: "all"
|
|
NVIDIA_DRIVER_CAPABILITIES: "compute,utility"
|
|
|
|
# Disable torch.compile/inductor - Triton doesn't support Blackwell sm_121a yet
|
|
TORCH_COMPILE_DISABLE: "1"
|
|
TORCHDYNAMO_DISABLE: "1"
|
|
|
|
# Grace-Blackwell unified memory optimizations
|
|
CUDA_CACHE_DISABLE: "1"
|
|
# PYTORCH_NO_CUDA_MEMORY_CACHING removed — our model_management patch handles
|
|
# caching properly by skipping empty_cache() on unified memory instead of disabling
|
|
# PyTorch's allocator entirely. Keeping caching ON reduces allocation overhead.
|
|
CUDA_DEVICE_MAX_CONNECTIONS: "1"
|
|
CUDA_DEVICE_MAX_COPY_CONNECTIONS: "4"
|
|
CUDA_MODULE_LOADING: "EAGER"
|
|
CUDA_MANAGED_FORCE_DEVICE_ALLOC: "1"
|
|
OMP_NUM_THREADS: "20"
|
|
CUBLAS_WORKSPACE_CONFIG: ":0:0"
|
|
|
|
volumes:
|
|
# Models from existing ComfyUI install (read-only)
|
|
- ${COMFYUI_HOST_PATH}/models:/opt/ComfyUI/models:ro
|
|
|
|
# Custom nodes - comment out to use container-only (fresh) custom_nodes
|
|
# If mounted, ComfyUI-Manager installs persist across container restarts
|
|
- ${SPARKYUI_DATA_PATH}/custom_nodes:/opt/ComfyUI/custom_nodes
|
|
|
|
# Outputs/inputs/workflows - persistent across restarts
|
|
- ${SPARKYUI_DATA_PATH}/output:/opt/ComfyUI/output
|
|
- ${SPARKYUI_DATA_PATH}/input:/opt/ComfyUI/input
|
|
- ${SPARKYUI_DATA_PATH}/workflows:/opt/ComfyUI/workflows
|
|
|
|
# Wheel cache (optional - for prebuilt wheels)
|
|
- ${SPARKYUI_DATA_PATH}/wheels:/opt/wheels
|
|
|
|
# Sparky patches - Grace-Blackwell unified memory optimizations
|
|
# This overrides ComfyUI's model_management.py with our patched version
|
|
- ./patches/model_management.py:/opt/ComfyUI/comfy/model_management.py:ro
|
|
|
|
networks:
|
|
- sparky_net
|
|
|
|
# Health check - ComfyUI takes time to load, so generous start period
|
|
healthcheck:
|
|
test: ["CMD", "curl", "-f", "http://localhost:8188/"]
|
|
interval: 30s
|
|
timeout: 10s
|
|
start_period: 120s
|
|
retries: 3
|
|
|
|
restart: unless-stopped
|
|
|
|
# ComfyUIMini - Mobile-friendly UI
|
|
# Access at http://<host>:3000
|
|
comfyuimini:
|
|
build:
|
|
context: ./comfyuimini
|
|
dockerfile: Dockerfile
|
|
args:
|
|
COMFYUIMINI_REF: "${COMFYUIMINI_REF:-main}"
|
|
|
|
image: comfyuimini:latest
|
|
container_name: comfyuimini
|
|
|
|
ports:
|
|
- "${COMFYUIMINI_PORT:-3000}:3000"
|
|
|
|
environment:
|
|
# node-config override - connects to comfyui container via docker network
|
|
NODE_CONFIG: >-
|
|
{
|
|
"app_port": 3000,
|
|
"comfyui_url": "http://comfyui:8188",
|
|
"comfyui_ws_url": "ws://comfyui:8188",
|
|
"output_dir": "/shared/output",
|
|
"reject_unauthorised_cert": false
|
|
}
|
|
|
|
volumes:
|
|
# Share output directory with ComfyUI for gallery feature (read-only)
|
|
- ${SPARKYUI_DATA_PATH}/output:/shared/output:ro
|
|
# Persist server-side workflows
|
|
- comfyuimini_workflows:/app/workflows
|
|
|
|
networks:
|
|
- sparky_net
|
|
|
|
depends_on:
|
|
comfyui:
|
|
condition: service_healthy
|
|
|
|
restart: unless-stopped
|
|
|
|
networks:
|
|
sparky_net:
|
|
driver: bridge
|
|
|
|
volumes:
|
|
comfyuimini_workflows:
|