Ray Init Actor (Advanced)

Ray Initializer (Advanced) is the same Raylight entry node as the plain Ray Initializer, plus the knobs you actually want once your setup stops being "two identical cards in the same slot" and starts being "two cards, one of which I also need for CLIP, VAE, and the rest of my graph."

If you only have one way to arrange your GPUs, use the basic version. If you've ever muttered "I wish I could keep GPU 0 out of this," this is the node you want - it's the only initializer with a GPU_SELECT input, and that alone is worth the switch on most multi-GPU boxes.

What the advanced inputs add

Everything from the regular initializer is here - GPU count, ulysses/ring/cfg/dp degrees, FSDP, FSDP CPU offload, attention backend, mmap loading - so I won't repeat it. The new bits:

• GPU_SELECT - a comma-separated list like 1,2 that restricts Ray workers to those GPUs. Leave it empty to use everything. This is the whole reason to pick Advanced: reserve GPU 0 for the CLIP encoder, VAE, or a controlnet pass while the workers chew on the diffusion model. The README literally suggests this use case.
• ray_object_store_gb - Ray's object-store size. The 2GB default is fine unless you're shuttling big tensors through Ray itself.
• ray_dashboard_address - point it at 127.0.0.1:8265 to watch worker memory and activity in Ray's dashboard while you debug.
• torch_dist_address - the master address for worker-side NCCL init (default 127.0.0.1:29500). The tooltip's warning is real: restart ComfyUI if you change it, it's read at startup.

Why you'd use this over the basic node

Honestly? Two reasons. First, GPU_SELECT is the difference between "add a second GPU" being a gift and being a headache - without it, Ray grabs every visible card and your VAE/CLIP ends up fighting the sampler for the same VRAM. Second, the dashboard hookup turns the whole "which worker is doing what and how much memory is it using" mystery into something you can watch live. For a tool whose failure mode is "silently slow because one worker is starved," that visibility is worth more than the name "Advanced" suggests.

Everything else is identical to the regular node: output is ray_actors_init, feeding the XFuser / Data Parallel / Unified Parallel samplers, and it's the same install path (ComfyUI Manager → "raylight", or clone + pip install -r requirements.txt). PyTorch 2.8.x is still the recommended version for FSDP, and the NCCL fix (pip install nvidia-nccl-cu12==2.28.9) applies here too.

Troubleshooting

• If a worker OOMs during NCCL collectives, the README notes Ray workers strip cudaMallocAsync from the allocator config - that's intentional, and RAYLIGHT_KEEP_CUDA_MALLOC_ASYNC=1 preserves it if you need it.
• New to Raylight entirely? Start with the basic initializer and the pack's example workflows (Wan T2V is the gentlest), then swap in Advanced once you know the degrees you need. The GPU_SELECT field defaults to empty, which means "all GPUs," so it behaves like the basic node until you tell it otherwise.