MiniMax H3 Speech VRAM Preflight / 显存预检 (EXP/T8)

The MiniMax H3 diffusion model is a 33B joint audio-video transformer, and running it on a consumer GPU is a constant game of "will it fit this time?" MiniMaxH3SpeechVRAMPreflightT8 is the node that answers the cheapest version of that question before you commit to a ten-minute render that OOMs at step six. It sits right before the speech sampler, checks the present state of your GPU, and reports whether you're above a headroom threshold you set. Honest framing straight from the author: it reports current whole-device free VRAM and your DynamicVRAM configuration, it cannot predict the sampler's peak, and it never grants a "memory_safe" label. It's a gate, not a prophecy.

How it works

The node takes a model (so it sits naturally in the model line before the sampler - it passes it straight through untouched), plus two knobs:

• minimum_headroom_mib (default 512) - your personal "don't start if less than this much free VRAM" threshold. The pack's own docs use 512 MiB as the standing reference line in its test reports, so the default is a sensible place to start.
• block_when_currently_below_gate (default false) - and here's the design decision worth respecting. By default the node just reports (current_gate_pass boolean + report_json). Flip this on and it raises a hard error when you're under the gate, so a downstream node can't run. Start with it off, watch what it tells you across a few jobs, then turn it on once you trust the threshold.

Outputs: model (pass-through, so you can chain it without a wire detour), current_gate_pass, and report_json with the actual free-VRAM numbers and the DynamicVRAM config it detected.

Why the caveats matter

H3's memory behaviour is nasty in a specific way: peak usage depends on frame count, canvas size, reference count, and whether ComfyUI's DynamicVRAM is mid-eviction. A preflight that samples free VRAM right now can't see the sampler's future peak - that's the honest limitation the description pins down, and it's why this node never claims safety. What it is good for: catching the obvious "the last job left the card full" case, and giving you a record (report_json) when a render fails so you can correlate "I had 300 MiB free" with the OOM. It's a diagnostic instrument, not a magic "make it fit" button.

Installing and the bigger picture

Same pack, one install: ComfyUI Manager → "MiniMax H3 Audio T8", or git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8 into custom_nodes, restart. No pip deps. You still need to supply the H3 weights yourself - and before you go deep, know the H3 Community License restricts the weights in the US, EU, UK and South Korea, so "does it fit" can be followed by the thornier "am I licensed to run it here." Start small (short render window, low resolution), read the preflight report, then scale up. And pair it with the pack's general H3 Preflight node for the model/label/dimension checks this one deliberately doesn't do.