MiniMax H3 TRT VAE 安装检查 (T8 EXP)

This one does exactly one thing and tells you so in its own description: it lists what's missing from your TensorRT VAE setup. No compiling, no GPU load, no pip install. If you're about to spend twenty minutes on an engine build that needs 12GB of free VRAM and 24GB of RAM, run this first - it's five seconds and it will tell you that you forgot the .onnx.data sidecar.

What it checks

You give it one input, runtime_directory. Leave it blank and it looks at ComfyUI/models/vae/h3_trt/runtime/site-packages, the isolated TensorRT install the pack expects. Point it at an absolute path if you already have your own.

Three outputs. files_present is a boolean - false the moment anything's missing. status is the human-readable version, and it names the missing items rather than making you diff the report by hand. report_json is the full inventory if you want to script against it.

The VAE side of this pack expects a specific pile of files under models/vae/h3_trt/: the decoder ONNX with its .onnx.data, a _flex variant prepared from those two, and for the full encode path an encoder ONNX plus a separately-prepared single-frame _t1 graph. The standalone TensorRT runtime in runtime/site-packages is installed separately, deliberately, so it can't overwrite ComfyUI's own Torch and CUDA:

<ComfyUI-Python> -m pip install --target <ComfyUI>/models/vae/h3_trt/runtime/site-packages tensorrt-cu13==10.13.3.9.post1

The pack also uses psutil and nvidia-ml-py (import name pynvml), and onnx if you still need to prepare the flex graph.

What "pass" does and doesn't mean

Read the node description again, it's unusually honest: files being present is not the same as inference having been validated. This node never loads a GPU context, so it cannot tell you whether the TensorRT version in that folder actually matches your driver, or whether the engines you eventually compile will run. It's the cheap gate before the expensive gates - nothing more.

Practical order of operations, straight from the pack's own docs:

1. Install the isolated runtime, put the ONNX files in place, run this check until files_present is true.
2. Prepare the flex graph with the bundled trt_vae_prepare_flex.py (it writes a new file, never overwrites the source, and verifies hashes).
3. Run the compile node - serially, nothing else on the GPU.
4. Only then swap the decoder node into your workflow where the video VAE used to be.

Install

Manager → search MiniMax H3 Audio T8, or:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8.git minimax-h3-audio-T8

Restart completely afterwards. You can skip all of this TensorRT business entirely and the rest of the pack still imports and works - that's a design decision, not an accident, and it's the right one. If your files are in place but the check still complains, the usual culprit is a runtime folder that exists but is empty, or a .onnx.data file that got separated from its graph during the copy.