MiniMax H3 Prompt Rewriter Unload / 手动卸载 (Advanced)

This node does one thing: it drops the optional 8B prompt-rewriter model this pack keeps cached, and returns the VRAM. If you're running MiniMax H3 on 16GB - and if you're here, you probably are - that's genuinely useful, because a Qwen3-VL-8B sitting in memory is a big chunk of the reason your next H3 run OOMs.

Let me back up a second, because the node name oversells slightly. There's no "unload" happening inside this node's little graph step the way you'd free a LoRA mid-pipeline. The pack's MiniMaxH3PromptRewriter8BT8Advanced node loads a Qwen3-VL-8B base with a LightX2V LoRA to rewrite your H3 prompt, and by default it already unloads after generating. This Unload node exists so you can force that release at a specific point in your workflow - after the rewrite is done but before the heavy H3 sampling starts, or when you're chaining multiple rewrites and want the memory back between them. It only touches the cache this pack owns. It won't touch other models in VRAM, and it doesn't restart anything.

Inputs and output

The whole interface is one boolean and a report string:

• unload_now - default true. false is a polite no-op that returns a {"status":"bypass"} report.
• report_json - a string with whatever actually got released (model, reference cleanup) or the bypass status.

That's it. You set unload_now to true, or you leave it and effectively get a documentation node. Wire report_json into a text preview or just let it print to the console.

Installing it

It ships in the T8mars MiniMax H3 Audio T8 pack, so install the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8

or via ComfyUI Manager by searching "MiniMax H3 Audio T8", then restart. The pack has no mandatory pip deps and won't download H3 weights. The one thing that matters for this node: if you actually use the 8B rewriter, you need the optional requirements-prompt-rewriter.txt installed (Transformers 4.x, PEFT 0.18–0.20, Pillow 10.x - pinned so they don't drag your other nodes into an upgrade war). No rewriter, no cache, nothing to unload.

Where it fits

The pack is heavily defensive by design - every experimental node defaults to report_only, false, or ABSTAIN rather than acting. This is that same philosophy: the author would rather you explicitly free the model than have it silently linger. Put this node right after your rewriter step if you're about to sample H3 on a tight card. If you never load the 8B rewriter, you can ignore it entirely - it's a memory-management convenience, not a workflow requirement.

One honest caveat: it releases this pack's rewriter cache only. If a whole separate 8B model is living in VRAM from another node, you're on your own. But for what it claims, it works, and for a node that's effectively a button it's refreshingly honest about what it touches.