ComfyUI caches the last model when RAM is plentiful (unified memory), so
memory doesn't drop after switching models even though models are being
swapped, not accumulated. Add a sidebar "Free GPU memory" button that
proxies ComfyUI's POST /free (unload_models + free_memory) via a new
/api/comfyui/free endpoint (COMFYUI_URL env). Verified it releases ~7GB.
README documents this plus the --disable-smart-memory auto-unload option.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
ComfyUIMini's built-in gallery is view-only. Inject a "Manage Photos"
link into its sidebar (via the shared head.ejs partial at build time, so
no fork) that points to the Model Manager's delete-capable Gallery. The
URL is built client-side from the browser host; the port is baked from
the MODEL_MANAGER_PORT build arg (default 8189).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Gallery view: grid of generated photos from ComfyUI's output/, full-size
lightbox, and permanent delete (with confirm). Paginated ("Load more").
- Backend: GET /api/gallery, GET /gallery/file (path-guarded image serve),
DELETE /api/gallery (path-guarded; clear error on permission denial).
- Mount ./output read-write into model-manager so the gallery can delete.
- Device-routing landing at /start: phones -> ComfyUIMini, desktops ->
the Gallery; ?force=mobile|desktop overrides. Ports come from the new
/api/ui-config (COMFYUI_PORT / COMFYUIMINI_PORT env).
- Responsive tweaks so the gallery is usable if opened directly on a phone.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Downloads previously landed in models/ owned by root because the
container ran as root. Add `user: "${PUID:-1000}:${PGID:-1000}"` to the
model-manager service and PUID/PGID to .env.example so downloaded models
are owned by the host user. Defaults to 1000:1000.
Note: existing root-owned files under models/ and sparkyui-data/ must be
chowned once (e.g. via a one-off root container) when upgrading.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
New FastAPI container (port 8189) to download and manage models:
- Installed Models, Add/Download (CivitAI/HuggingFace/direct URL), Settings views
- Persistent SQLite storage for API keys and download history (./sparkyui-data)
- Downloads land in ./models, auto-sorted into ComfyUI's standard subfolders
- Default COMFYUI_HOST_PATH and SPARKYUI_DATA_PATH to the project root
- Wire docker-compose service, env defaults, gitignore, README docs
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Major unified memory optimization changes:
1. model_management.py: HIGH_VRAM → NORMAL_VRAM
- GB10 unified memory: offloading to CPU doesn't save physical RAM
(same pool), but NORMAL_VRAM allows per-layer partial loading when
memory is tight instead of all-or-nothing OOM
- text_encoder_offload_device() and vae_offload_device() now return
CPU (allows ComfyUI to offload unused models)
- intermediate_device() still returns GPU (VAE outputs must stay in
CUDA allocator for honest memory tracking)
- User can force HIGH_VRAM with --highvram if models fit
2. utils.py: copy=True → copy=False for tensor.to(device)
- On GB10 unified memory, copy=True creates a full duplicate in both
CPU and CUDA allocators simultaneously (ComfyUI issue #10896)
- copy=False makes .to(device) a zero-copy device label change since
both allocators draw from the same physical LPDDR5X
- Halves model loading memory usage when --disable-mmap is set
3. Removed --disable-dynamic-vram from ComfyUI flags
- Was preventing AIMDO (comfy_aimdo) from initializing
- AIMDO now activates: VBAR-based page-level VRAM management at 32MB
granularity instead of blunt .to(cpu) copies
- Falls back to NORMAL_VRAM per-layer loading if AIMDO has issues
4. Added CUDA_CACHE_MAXSIZE=4294967296 (4GB kernel cache)
- PTX→SASS kernel caching for sm_121 (GB10 Blackwell)
- 3x speedup on subsequent runs reported by DGX Spark community
5. System: vm.swappiness reduced from 60 to 1
- Swap thrashing on unified memory causes silent system freezes
- Near-zero swappiness ensures clean OOM kills instead
Docker-based ComfyUI setup for NVIDIA DGX Spark ARM64 + sm_121:
- CUDA 13.0.2 base (required for compute_121 support)
- PyTorch 2.9.1+cu130 ARM64 wheels
- SageAttention compiled with TORCH_CUDA_ARCH_LIST="12.1"
- Triton/torch.compile disabled (no sm_121 support yet)
- ComfyUI-Manager auto-installed at runtime
- Configurable model/data paths via .env
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
TBNilles
Evan Carmen