ComfyUI-PipelineBarrier

A single-node ComfyUI extension that inserts an explicit GPU cache flush and garbage-collection pass between pipeline stages.

Why?

Long multi-stage workflows — especially WanVideo two-pass I2V with ComfyUI-MultiGPU — can OOM-kill the process at the start of the next stage even after the previous model has been offloaded via BlockSwap's force_offload. The problem is that PyTorch's CUDA allocator caches freed blocks rather than returning them to the OS immediately, so the system sees less free RAM than is actually available.

PipelineMemoryBarrier calls ComfyUI's soft_empty_cache(force=True), Python's gc.collect(), and PyTorch's torch.cuda.empty_cache() + torch.cuda.synchronize() on every visible GPU, ensuring caches are flushed and memory is returned to the OS before the next stage begins.

Installation

Via ComfyUI Manager (recommended)

Search for Pipeline Memory Barrier in the ComfyUI Manager node list.

Manual

cd ComfyUI/custom_nodes
git clone https://github.com/brosequist/ComfyUI-PipelineBarrier

No pip dependencies beyond what ComfyUI already installs (psutil is included in the base environment).

Usage

The node accepts a LATENT input and returns it unchanged.

Typical placement in a WanVideo two-pass workflow:

[Pass 1 Sampler] --latent--> [Pipeline Memory Barrier] --latent--> [Pass 2 Sampler]
[Pass 2 Sampler] --latent--> [Pipeline Memory Barrier] --latent--> [VAE Decode]

Parameters

| Parameter | Type | Default | Description | |-----------|------|---------|-------------| | latent | LATENT | — | Latent tensor. Returned unchanged. | | log_memory | bool | True | When enabled, logs GPU allocated/reserved and system RAM before and after the flush to the ComfyUI console. |

Example console output

[PipelineBarrier] before flush | GPU 0: alloc=14.32 GiB  reserved=15.12 GiB
[PipelineBarrier] before flush | GPU 1: alloc=7.18 GiB   reserved=8.00 GiB
[PipelineBarrier] before flush | RAM: 41.2/62.8 GiB  (66% used)
[PipelineBarrier] after  flush | GPU 0: alloc=0.01 GiB   reserved=0.01 GiB
[PipelineBarrier] after  flush | GPU 1: alloc=0.00 GiB   reserved=0.00 GiB
[PipelineBarrier] after  flush | RAM: 11.4/62.8 GiB  (18% used)

Notes

• The node costs essentially zero compute time — it is a synchronization point only.
• log_memory=False silences the console output for production workflows.
• The flush is safe to call at any point; it does not unload models that ComfyUI is still tracking, only releases PyTorch's internal cache of freed blocks.
• On multi-GPU setups all visible CUDA devices are flushed.

License

MIT