Adds opt-in process isolation for custom nodes using pyisolate's
bwrap sandbox and JSON-RPC bridge. Each isolated node pack runs in
its own child process with zero-copy tensor transfer via shared memory.
Core infrastructure:
- CLI flag --use-process-isolation to enable isolation
- Host/child startup fencing via PYISOLATE_CHILD env var
- Manifest-driven node discovery and extension loading
- JSON-RPC bridge between host and child processes
- Shared memory forensics for leak detection
Proxy layer:
- ModelPatcher, CLIP, VAE, and ModelSampling proxies
- Host service proxies (folder_paths, model_management, progress, etc.)
- Proxy base with automatic method forwarding
Execution integration:
- Extension wrapper with V3 hidden param mapping
- Runtime helpers for isolated node execution
- Host policy for node isolation decisions
- Fenced sampler device handling and model ejection parity
Serializers for cross-process data transfer:
- File3D (GLB), PLY (structured + gaussian), NPZ (streaming frames),
VIDEO (VideoFromFile + VideoFromComponents) serializers
- data_type flag in SerializerRegistry for type-aware dispatch
- Isolated get_temp_directory() fence
New core save nodes:
- SavePLY and SaveNPZ with comfytype registrations (Ply, Npz)
DynamicVRAM compatibility:
- comfy-aimdo early init gated by isolation fence
Tests:
- Integration and policy tests for isolation lifecycle
- Manifest loader, host policy, proxy, and adapter unit tests
Depends on: pyisolate >= 0.9.2
Integrate comfy-aimdo 0.2 which takes a different approach to
installing the memory allocator hook. Instead of using the complicated
and buggy pytorch MemPool+CudaPluggableAlloctor, cuda is directly hooked
making the process much more transparent to both comfy and pytorch. As
far as pytorch knows, aimdo doesnt exist anymore, and just operates
behind the scenes.
Remove all the mempool setup stuff for dynamic_vram and bump the
comfy-aimdo version. Remove the allocator object from memory_management
and demote its use as an enablment check to a boolean flag.
Comfy-aimdo 0.2 also support the pytorch cuda async allocator, so
remove the dynamic_vram based force disablement of cuda_malloc and
just go back to the old settings of allocators based on command line
input.
