MQ Check FP4 Support

FP4 is the current endgame of squeezing big models onto small cards, and it's also the fastest way to discover your GPU is the wrong generation. MQ Check FP4 Support is a node with no inputs and exactly one job: tell you, as a plain boolean, whether your graphics card can run 4-bit floating point. If you've ever grabbed a shiny FP4 checkpoint for an RTX 3060, you'll appreciate why that question is worth asking before the loader half-finishes and dies.

A quick orientation, because FP4 sits at the sharp end of a ladder most people meet at the fp8 rung. fp8 (e4m3) is the community's "just use it" default for Flux-class models - half the VRAM of fp16 with near-invisible quality loss. FP4 is half of that, and it's where the decision stops being about file formats and starts being about hardware. NVIDIA only added FP4 (marketed as NVFP4) on Blackwell - the RTX 50-series and the B100/B200 datacenter parts. No 40-series card does it natively. That hardware gate is exactly what this node tests.

Mechanism, straight from the source rather than the README: when the workflow runs, it asks torch whether CUDA is available, reads the current device's compute capability, and returns True only when the major version is above 9. Compute capability major 10 and 12 are Blackwell and later; Ampere sits at 8, Ada at 8.9, Hopper at 9. Anything below, or no CUDA at all, or any exception along the way, and it quietly returns False.

There are no inputs to set, so the only thing that matters is the single output, FP4 Support (BOOLEAN). Wire it into anything that takes a boolean - an MQ Int Switch's condition, a Display Any node to just see the answer, or a conditional that routes to an FP4 loader versus an fp8 loader. The genuinely useful pattern is the last one: a shared workflow that picks the right quantization for whatever GPU it happens to be running on. Sharing an FP4 workflow with a friend on a 4070? This node is your early-exit guard.

Three honest caveats. First, it's NVIDIA-only by construction - the check is torch.cuda, so AMD and Intel cards always read False even where they run fp8 perfectly well. Second, True is a hardware verdict, not a software promise: it doesn't mean FP4 checkpoints exist for your model, or that your ComfyUI build's NVFP4 path is wired up. Third, don't be surprised when your RTX 4080 reports False - that's correct behavior, Ada genuinely can't do FP4.

Install it the usual way: ComfyUI Manager, search ComfyUI MqUtils, install, restart. Or old school:

cd ComfyUI/custom_nodes
git clone https://github.com/mikeshuangyan/ComfyUI_MqUtils.git

Restart ComfyUI. The pack has zero Python dependencies beyond torch (which ComfyUI already ships), downloads no models, and installs in about a second. For that reason alone it's worth having - it makes the FP4 question answerable in-graph instead of by squinting at a spec sheet.