🔍 Qwen Diagnostics and Downloader

QwenDiagnostics is the honest part of the ComfyUI-Qwen-Prompt-Expander pack: a preflight check that tells you why the main node isn't working before you spend an hour guessing. First-run experiences with the pack hinge on the main node silently failing to download a model - slow network, HuggingFace blocked in your region, a models/LLM folder you can't write to, a full disk. This node tests every one of those and hands you a report instead of a cryptic console trace.

It's an output node, so nothing wires out of it into your generation graph. You queue a run, it does its thing, and the diagnostic_report string appears on the node (wire it into a Show Text node if you want it in a viewer or saved to disk). It's a diagnostic tool, not a workflow component.

What it actually tests

Flip run_test to True and queue - with run_test at False it just prints a "set me to True" hint. Then it runs through, in order:

1. System info - PyTorch version, CUDA availability, GPU name and VRAM, transformers version.
2. Internet - it opens a socket to huggingface.co:443. If that fails, it stops right there, because nothing downstream is going to work.
3. HuggingFace Hub API access.
4. Model directory - creates ComfyUI/models/LLM/ if it doesn't exist, and checks write permissions (the node writes a test file and deletes it).
5. Disk space - warns if you have under 10GB free.
6. The real test: it downloads your chosen model and runs a tiny 5-token generation to prove inference works.

Success ends with 🎉 ALL TESTS PASSED! and a pointer to the main node. Failures print exactly which check broke and a suggested fix (check internet, try a VPN if HF is blocked, free disk, fix permissions).

The two inputs, honestly

• test_model - Qwen-0.5B (Smallest - ~1GB) (default) or SmolLM2-1.7B (Recommended - ~3.5GB). Pick the small one. Worth knowing: the test downloads at fp16, not quantized, so those sizes are the real disk/VRAM footprint - the SmolLM2 test is a proper 3.5GB pull, which is most of the download the main node would do anyway. Models land in the same models/LLM/ cache the main node reads, so a passed test means the main node's first real generation is basically instant.
• run_test - the Boolean trigger with the "Run"/"Ready" label. It's a latch: set it True, queue, read the report. Set it back to False when you're done so you don't re-run the whole battery by accident.

Using it

It installs with the pack, so there's no separate step: Manager → "Qwen Prompt Expander", or the git clone plus pip install -r requirements.txt from the README - same as the main node. And honestly, that's the case for reaching for it at all: run it once, right after install, before you queue your first real generation with the expander. If the download part hangs or you're in a region where HuggingFace is throttled, the report tells you which failure mode you're in rather than leaving you staring at the main node's spinner.

The one thing it doesn't test for is the pack's own risk profile. A diagnostic node proves your network and disk are fine - it can't prove the code you just installed is trustworthy. The ComfyUI ecosystem has a real malware history (the LLMVISION incident), and this pack is a small, freshly-released, community-contested one, so the sensible habit is the same as always: skim the source of any new node pack before you run it, and keep the 0.5B test model as your first download either way.