⛱️Qwen2 ModelLoader

The ⛱️Qwen2 ModelLoader is the whole ballgame for this pack. The other two nodes - ⛱️Qwen2 and ⛱️Qwen2 Chat - are just text in, text out; they can't do anything until this node has loaded a Qwen Instruct model off HuggingFace and handed you the Qwen2 model and tokenizer outputs to wire into them. So it's one node, but it's also the reason this pack is a "make sure your machine can actually handle it" kind of install.

The name is doing a lot of work, too: it's a loader, but it's also a downloader. Under the hood it just calls AutoModelForCausalLM.from_pretrained(model_name, device_map="cuda", torch_dtype="auto") and the matching tokenizer call. No API, no key, no cloud - but the first time you pick a model it pulls several gigabytes into HuggingFace's cache (usually ~/.cache/huggingface, not your ComfyUI models/ folder), which is exactly why the README's first piece of advice is "keep your network open." Watch the download the first run and you'll see where the time went.

The one input that matters

model_name is a dropdown of nine HuggingFace repo IDs, and it's the only input there is. Qwen2.5's 0.5B, 1.5B and 3B Instruct are the ones a normal consumer GPU can chew through; 7B lands around 16GB in fp16, which is genuinely comfortable on a 24GB card and doable on 16GB with a tight workflow. That's where most people should stop. The 14B, 32B and 72B entries exist, but there's no quantization path in this pack - the loader is hardcoded to CUDA with fp16-ish auto dtype, so the whole model goes on the GPU. A 72B in fp16 is roughly 140GB of VRAM. The README claims ">24GB VRAM" for it, and that's optimistic to the point of being wrong; treat anything above 7B as a multi-GPU or "I have a serious rig" choice.

That missing quantization option is the pack's real limit. If you want a 7B running on a 12GB card, or Qwen quantized to GGUF at all, this isn't the pack for it - ComfyUI-GGUF and similar are the better rabbit hole.

Outputs

Two: Qwen2 (the loaded model) and tokenizer. Take both and feed them into the matching inputs on either ⛱️Qwen2 or ⛱️Qwen2 Chat. They always go together - miss one and you'll get a type error before anything runs.

Install

ComfyUI Manager is the easy path - search for ComfyUI-Qwen and let it handle the rest. Manual install is short:

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

Then restart ComfyUI. The requirements are thin - just transformers>=4.40.0 - which is both the good news and the first thing that bites you. If you're on an old transformers, expect import or generation errors; the README's upgrade is pip install git+https://github.com/huggingface/transformers.

Common issues

• CUDA out of memory - the most common one by far, and there's no slider to fix it. Pick a smaller model_name. That's the whole fix.
• First run appears frozen - it's downloading, not dead. Give it time and watch the network.
• device errors on non-NVIDIA setups - CUDA is hardcoded, so Apple Silicon and CPU-only machines are effectively out of luck with this pack.