Pt Prod

PtProd computes the product of a tensor along one dimension - torch.prod() as a node. Sums get all the attention; products are the quiet cousin you forget exists until you need to multiply an entire axis of values down to a single number. That's exactly what this node is for: collapse a dimension by multiplying everything in it.

How it works

Three inputs, one TENSOR output:

• tens - the tensor you're reducing.
• dim - which axis to multiply along, as a text field (e.g. 0 or 1). This is a string in the node's schema because the pack parses it, but you'll almost always type a single integer. A product over dim 0 of a (3, 4) tensor gives you a (4,) tensor where each element is the product down that column.
• keepdim - whether to keep the reduced axis as a size-1 dimension (so a (3, 4) tensor reduced over dim 0 becomes (1, 4) instead of (4,)). Default False. Leave it True if you're feeding the result into something that expects the original rank.

Output is a single TENSOR you can chain into any Pt* node.

When you'd actually reach for it

The honest use cases are narrower than sum's: probability-style math (a product of independent probabilities), volume-style computations over a dimension, factorial-ish reductions, or feature-aggregation experiments where you genuinely want a multiplicative aggregate instead of an average. There's also a classic numerical footgun that makes this node worth understanding: multiplying many values between 0 and 1 shrinks the result toward zero alarmingly fast, and multiplying large values can overflow to inf. If your product comes out as 0.0 or inf, that's floating-point reality doing its thing, not a broken node - you're often better off summing logs.

Installing it

Part of ComfyUI-Pt-Wrapper (HowToSD's no-code PyTorch pack, a spin-off of ComfyUI-Data-Analysis). ComfyUI Manager → search "ComfyUI-Pt-Wrapper", or:

cd ComfyUI/custom_nodes
git clone https://github.com/HowToSD/ComfyUI-Pt-Wrapper

Restart after; first boot is slow while pandas, scikit-learn, transformers, sentencepiece, peft and friends install. No model downloads needed.