MiniMax H3 VRAM Policy / VBAR显存预留策略 (Advanced)

The #1 community complaint about running MiniMax H3 locally is VRAM - it's a 33B model with a joint AV Transformer, and on a 16GB card the difference between "renders" and "crashes" is often a couple of hundred MiB of headroom. MiniMaxH3VRAMPolicyT8Advanced is the pack's answer: a guarded plan for how much VRAM to reserve, how much headroom to keep, and whether to block the run when the machine can't promise it. Crucially, its default is report-only - it tells you the plan without doing anything, and you have to explicitly opt into the modes that actually reserve memory.

The modes (mode)

• report_only (default) - zero side effects. It measures, computes headroom, and reports. This is the safe way to start: run it, read the numbers, learn your machine.
• fixed_total_reserved_exp - sets a total VRAM reserve. The tooltip gives you the honest anchor: 4.0 GiB is the validated conservative starting point for one exact RTX 4060 Ti 16GB running the 736×416, 124-frame Hybrid Stock20 workflow - and it's not a universal safe value, so re-measure on your GPU.
• external_usage_plus_margin_exp - reserves based on current external usage plus a margin, and requires clean_before_load first (that's a global side effect: it unloads every ComfyUI model before measuring). This is the invasive one; you asked for it by name.

The other knobs that matter

• fixed_total_reserved_gib (4.0) - the reserve amount in the fixed mode.
• maximum_reserved_gib (8) - a ceiling so a policy can't eat your whole card.
• minimum_current_headroom_mib (512) - the current-headroom floor; the README's project-wide 512MiB gate is exactly this.
• minimum_commit_headroom_gib (16) - system commit/swap headroom floor, because VRAM alone isn't the whole story on Windows.
• block_when_commit_below_gate (default on) - fail closed instead of running and OOMing.
• policy_epoch - bump it to explicitly invalidate ComfyUI's cached policy output after you change settings; the "cache made it not apply my change" confusion is real, and this is the reset button.
• require_dynamic_vram (default on) - the plan expects DynamicVRAM/VBAR-style headroom to be active.

Outputs:

vram_policy (the object you feed to the Advanced Hybrid Loader - that's how you guarantee the policy applies before the model loads, which is the whole point), current_gate_pass (a boolean you can gate a switch on), and report_json.

Installing:

part of MiniMax H3 Audio T8. Manager → "MiniMax H3 Audio T8", or:

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

Restart. No mandatory pip deps; the H3 weights are the thing eating your VRAM, so bring them.

The honest framing

This node measures and reserves; it doesn't make H3 fit on hardware that can't run it. On a 16GB card you still have to respect canvas/frame/reference budgets - the README is emphatic that 16GB is not automatically safe for every task. Use report-only first, learn your headroom, then pick the least invasive mode that keeps the run alive. And note the broader reality: if you're in the US/EU/UK/Korea, H3's local weights are licence-restricted, so this is a "when you can legally run it" tool.