TurboQuant Info

This node doesn't compress anything. It's the readout. TurboQuant Info peers at the stats collected by its sibling, TurboQuant KV Patch, and prints them as text: how many tensors went through TQ3, how many MB they were before and after, the observed compression ratio, and the byte savings. Without it, the patch runs silently and you have no idea whether it did anything at all.

Why you'd reach for it

You only want this node if you've already added TurboQuantPatch to your graph, because the patch is an experiment and experiments need measurement. The whole point of the pack is studying KV-cache compression on real diffusion attention - the ratio and savings here are observed on your workflow, not the README's headline number. If you're trying to figure out whether 3-bit K/V quantization is worth chasing for the model you run, this is the node that gives you the answer in megabytes.

How it works

When TurboQuantPatch runs, it stashes a global wrapper object that tallies up compressed_bytes, original_bytes, and the number of stores. TurboQuantInfo's info() reads that global and formats it into a human-readable stats string. It's a live readout of the last patch that ran, not a stored report - so it reflects whatever you just generated with.

One honest detail worth knowing: if no inference has run since the patch was wired up (or the patch is disabled), the node doesn't fake numbers. It returns a clear "Status: Not yet active (no inference run)" message alongside the expected ~4.5x and the encoding details, so you always know whether the stats are real or just the spec sheet.

Inputs and outputs

The schema is minimal, which is the point:

model (MODEL) - connect the patched MODEL output from TurboQuantPatch here. It's the workflow contract that keeps the nodes visually linked, even though the stats actually come from the global.
stats (STRING) - the single output. The node is marked as an output node, so it displays itself; you can also feed this string into a text/display node if you want it in your saved output.

Installing it

Same pack, same steps as the patch. ComfyUI Manager finds it by searching ComfyUI-TurboQuant; manually:

cd ComfyUI/custom_nodes
git clone https://github.com/Scottcjn/ComfyUI-TurboQuant

then restart. No requirements.txt to worry about - it's pure PyTorch, already present in any ComfyUI install. If you want to sanity-check the quantization core before trusting any numbers, python -m tq3_core in the repo runs its self-test.

Gotchas

The stats are global and cumulative: run a workflow twice in a row and the numbers keep adding into the same wrapper, so judge a run by the incremental change, or restart your queue. Disabling the patch clears the stats entirely, which is why you'll suddenly see "Not yet active." And the one thing this node deliberately does not tell you: image quality. The reported ratio is bytes, and the README's ">0.97 cosine similarity" is a claim about attention-vector round-trips, not a guarantee your frames look identical. If you're A/B testing, pair the stats with your eyes. It's a tiny node, but it's the honest kind of tiny - it tells you exactly what the experiment measured and nothing more.