Krea2 SVDQuant Env Check

The short version

This is the smallest node in the pack and possibly the most useful one to run first. It has no inputs, needs no model, no checkpoint, no download - you drop it on the canvas, press Queue, and it tells you whether the int4 tensor-core kernel is live on your install. The pack's own docstring puts it plainly: run this before downloading an 8 GB checkpoint. If it says the CUDA backend isn't live, quantized Krea 2 checkpoints will be slower than FP8 - and knowing that before you download saves you an hour of confusion.

It exists because "the quantized checkpoint is slower than fp8" is the single most common report this pack gets, and the answer is almost always the same thing: ComfyUI disabled comfy_kitchen's CUDA backend because your torch was built against CUDA < 13. That's answerable without loading anything, so this node makes it answerable before the 8 GB download rather than after.

How it works

ComfyUI gates its quantization backend in comfy/quant_ops.py: if cuda_version < (13,), it calls ck.registry.disable("cuda") and every quantized layer resolves to a pure-Python eager implementation. That fallback unpacks int4 to bf16 in Python and runs an ordinary matmul - strictly slower than plain bf16, and the more aggressive the format, the worse it gets. The ordering inverts: fp8 fastest, w4a4/svdq slowest, the exact opposite of the benchmark table.

The node checks the backend registry directly and reports its status. Output is a single report string (backend status, for reading or pasting into a bug report). Because backend availability is process state rather than graph state, it's never served from cache - every queue press re-checks what's true right now, which is the whole point.

How to install

Same pack as everything else here:

cd ComfyUI/custom_nodes && git clone https://github.com/alperktt/Krea-2-SVDQuant-ComfyUI krea-2-svdquant

No Python dependencies. If you'd rather not open ComfyUI at all, the terminal equivalent is:

cd ComfyUI/custom_nodes/krea-2-svdquant
python diagnose.py --no-load

When to run it

• Before your first SVDQuant download - 30 seconds now, an 8 GB download + a slow generation later otherwise.
• When a quantized checkpoint is slower than your FP8 build - this is the definitive check.
• After a PyTorch upgrade - if you installed cu130 but it didn't take (a stale venv, a conda env still pointing at the old build), this will tell you.

The fix if it fails

If the report says the CUDA backend isn't live, check what torch actually is:

python -c "import torch; print(torch.__version__, torch.version.cuda)"

Below 13.0, install a cu130+ torch build, restart, and re-run the check. One honest caveat so you don't chase the wrong thing: on an RTX 20-series (Turing), the kernel can report healthy and still not be much faster than int8 - Turing simply has weaker int4 MMA hardware. That's not this node lying; it's your card predating the fast instruction.