🧠 Offload Model to DRAM

The mirror image of the CLIP offload, and arguably the more useful half of the pair. Sampling a 20B Qwen-Image-Edit UNet leaves almost nothing for the VAE decode that comes after - which is when ComfyUI either spills to CPU, crawls, or dies outright. ArchAi3D_Offload_Model moves the diffusion model out of VRAM the moment the KSampler is done, so the decoder gets the card to itself.

What it does

A side-effect node in the same DRAM cache family as the CLIP version: grab the diffusion model, move its weights from VRAM to CPU RAM, pass your data through. The difference is placement in the pipeline. The CLIP offload goes before sampling; this one goes after - its example is literally "LATENT from KSampler → VAE Decode."

Inputs and outputs

• model (required, MODEL) - the diffusion model to offload, straight from your model loader.
• trigger (optional, any type) - connect the KSampler's LATENT output here. The node runs because of the trigger, and the LATENT passes through untouched.

Outputs:

• memory_stats (STRING) - live VRAM/RAM/cache status so you can verify the VRAM was actually freed.
• dram_id (STRING) - the cache key matching the pack's triggered model loader, so the same model reloads from RAM (~1s) instead of disk (~8s) on the next run.
• passthrough (any type) - the LATENT, unchanged, ready for the VAE decode.

How the mechanism works

The pack's shared dram_cache module holds strong references to offloaded weights in a module-level dict, so nothing gets garbage-collected mid-workflow. The offload node stashes the model there with a deterministic key, and partially_unload() keeps it tracked in ComfyUI's current_loaded_models - if something asks for the model again, ComfyUI auto-reloads it from RAM rather than re-reading the file. The README's pipeline is the clearest statement of intent: encode → offload CLIP → sample → offload model → decode with "full VRAM available."

When to reach for it

Same trigger points as its sibling: OOM at decode, ComfyUI thrashing between the UNet and VAE, or a workflow where you chain many edit passes and want each stage to have the whole card. The pack is tested on a 24GB 3090 Ti with the 20B model in fp8/quantized form, and the changelog specifies the two flags that make the system behave: --normalvram --cache-classic. If you're on a 24GB card running Qwen Edit, this pair of offload nodes is probably the difference between "it works" and "it works without babysitting."

Install

Same pack as everything else here:

cd ComfyUI/custom_nodes
git clone https://github.com/amir84ferdos/ComfyUI-ArchAi3d-Qwen.git
cd ComfyUI-ArchAi3d-Qwen && pip install -r requirements.txt

Or ComfyUI Manager → search "ArchAi3d Qwen". Restart; the node lives under ArchAi3d/Memory.

Gotchas

• Don't skip the trigger. No trigger, no guaranteed ordering - the offload can fire before the model is done sampling and you'll evict a model you're still using.
• RAM is the ceiling. The model sits in system RAM while cached. A 20B quantized model in DRAM plus the OS on a 16GB machine is a swap storm waiting to happen; the pack's DRAM guide recommends Linux vm.swappiness tuning if you're on the edge.
• It's manual by design. ComfyUI already has automatic management; this exists for when you want to say exactly when each model leaves the GPU. If ComfyUI's defaults are working for you, leave them alone.