MiniMax H3 TRT VAE 本机编译 (T8 EXP)

TensorRT engines are hardware-specific. There is no download that works - the engine has to be built on the machine that will run it, against that machine's GPU, driver and runtime versions. That's what this node is for: a one-shot, serial, out-of-process compile you run when you're setting up, and then never again.

Set expectations first, because this is the expensive step of the TRT VAE route. Compiling requires at least 12000 MiB of free VRAM, 24 GiB of free RAM and 20 GiB of free disk, it runs alone (nothing else sampling on the GPU), and it's the reason the pack ships a file check and a separate decoder-only node that most people should prefer. If you were hoping to shave a couple of seconds off a clip, walk away now.

What it does

Pick a kind: decoder-flex is the recommended one, and the tooltip explains why - a flex decoder is what you need for single frames or small sizes, since the fixed-shape decoder can't handle them. The others are decoder (the plain graph), encoder-t1 and encoder for the full encode path, and decoder-w4a16, a quantised variant whose output the author reviewed as broadly acceptable but with noticeably larger deviation, and with no promise of lower VRAM or higher speed.

runtime_directory blank means ComfyUI/models/vae/h3_trt/runtime/site-packages, the isolated TensorRT install. timeout_seconds defaults to 1800 and bounds one compile; cancellation ends that subprocess only, and it does so leaving your existing engines alone. If a build fails - or the machine drops below the memory floor mid-run - the node stops the child and keeps the previous engine rather than leaving you with a half-written one.

Outputs are engine_directory and report_json. Builds are recorded under output/MiniMaxH3/TRT-VAE/builds/<number>.

Before you run it

Files in place, checked by the installation node, and the ONNX graphs prepared. The flex graph is not a download - you generate it locally from the decoder ONNX with the bundled CPU tool, which only rewrites input/output shape declarations and verifies hashes on the way:

<ComfyUI-Python> trt_vae_prepare_flex.py --source <ComfyUI>/models/vae/h3_trt/minimax_h3_vae_decoder.onnx --output <ComfyUI>/models/vae/h3_trt/minimax_h3_vae_decoder_flex.onnx --report <ComfyUI>/models/vae/h3_trt/flex-preparation.json

Only the full encode path needs the extra T1 graph, exported from your native VAE weights with trt_vae_prepare_t1.py. That step uses the GPU briefly and must be serial too. If you have no use for TRT encoding - and per the next two nodes, most people don't - skip it entirely.

Then compile one kind at a time, and refresh the model lists in ComfyUI afterwards so the new engine folder shows up in the decoder node's dropdown.

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 fully. The pack installs no Python packages by design; the TensorRT runtime goes in with an explicit --target so it can't replace ComfyUI's 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 traps

Engines are bound to the runtime directory and file identities they were built against. Move the folder, change the driver, change the TensorRT build, or rebuild on another card, and you rebuild. The node will not quietly recompile when you hit Generate - it fails instead, which is the behaviour you want, because a silently recompiling node would also silently change what your workflow is doing.

Don't scatter engine files around, either. model.engine belongs with its manifest.json under h3_trt/engines/<build>/, not in diffusion_models, and copying an engine between machines is the classic way to waste an afternoon on an error that says nothing about mismatched hardware.

Finally: the reference TRT path is a Windows, single RTX 4060 Ti 16GB, Python 3.12, PyTorch 2.10/cu130, TensorRT 10.13.3.9 setup. Other combinations may well work, but nothing in the pack claims they were tested.