RDNA35 Block Attention Diagnostics

You add a ComfyUI-RDNA35-Attention node, hit Queue, and get... nothing. No speedup, a benchmark that prints "optimized path did not run," a patch whose info output is a wall of fallback reasons. The first thing this pack's author would tell you is to run RDNA35 Block Attention Diagnostics before you do anything else, because the whole pack is built for a very specific machine, and this node exists to tell you whether you're on it.

It's the simplest node in the pack: zero inputs, one diagnostics STRING output, and it's an output node so you just drop it in and watch the text panel. It dumps your runtime in the exact shape the pack's dispatch code checks - PyTorch version, torch.version.hip, device name, a best-effort gfx target, whether that target counts as RDNA3.5 (gfx1150, gfx1151, or gfx1152), and whether Triton imports.

Why this particular list matters

This pack is an AMD-only research project. It targets PyTorch ROCm builds on RDNA3.5 hardware - that's the RX 9070 series - and every optimized path in it gates on the same conditions the diagnostics node reads. If torch.version.hip is None, you're running a CUDA build and none of the Triton kernels will ever dispatch. If the gfx target isn't 115x, the fast paths refuse to run. If Triton is missing or fails to import, every node falls back to a slower PyTorch reference with a reason.

So the Diagnostics output isn't just gossip about your machine. It's the exact gate the dispatch code in rdna35_block_attention/dispatch.py uses to decide between the Triton kernel and the reference implementation. When a benchmark later tells you fallback_reason: triton_unavailable_..., you'll already know why.

What you actually see

The output is a plain text report - no numbers to tune, no toggles to set, nothing to wire downstream except the string. The line you care about is RDNA3.5 target: True/False. If it says False, you have your answer: you're on the wrong GPU or the wrong build for the optimized paths, and the reference fallbacks are what you'll get. If it says True, you still want to check Triton: - a ROCm PyTorch build doesn't ship a working Triton by default, and getting one that matches your build is the most common setup failure in this pack.

Installing it

Same story as the rest of the pack: search for RDNA35 Attention in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Yasei-no-otoko/ComfyUI-RDNA35-Attention

Then restart ComfyUI. No pip dependencies - the base install is genuinely zero-dependency (dependencies = [] in its pyproject). All the heavy stuff is runtime: the PyTorch ROCm build, a compatible Triton, and for the PISA nodes a separately compiled rdna35_pisa_ck wheel. Nothing here downloads model files; the diagnostics node needs only your existing Python environment.

Gotchas

This node works everywhere, even on a CPU-only box - the report just tells you what's missing. That's arguably its whole job. Don't expect it to fix anything, though. It reports; the pack's patches and benchmarks do the falling back. And fair warning: this pack is a single-author research effort with essentially no community footprint yet, so when something looks broken, the README and the source are your support channel. Run this node first, and you'll be ahead of most people who hit these nodes cold.