Load MiniMax H3 INT8 Fast (ROCm)

This is the node the pack exists for. MiniMax H3 is the 33B omni-modal video model - text, image, video and audio in one context, native stereo sound - and it's a monster in full precision, with a community figure of roughly 42GB for the weights alone. On a consumer card, that's the whole VRAM budget gone before you generate anything. The official MiniMax_H3_FL2VA_pruned_int8_convrot.safetensors and minimax_h3_ref2va_pruned_int8_convrot.safetensors conversions shrink that to something a sensible card can hold, and this loader is the ROCm-tuned way to run them.

What makes it different from just loading the file with a stock loader is how it quantizes. H3 is mixed-precision by design: some layers are precision-sensitive (conditioning, AdaLN, the patch layers, the audio and video output heads) and shouldn't be touched. The loader validates your checkpoint - it requires the int8_tensorwise ConvRot format with 256 group size and, notably, FP32 audio/video output heads - then hands the model to ComfyUI's native H3 loading, and only routes the handful of ConvRot projection layers through a scoped ROCm Triton INT8 backend. Everything else stays on the normal HIP/native path in its original BF16/FP16/FP32. That scoping is the clever bit: it's a fast INT8 path that doesn't quietly destroy the layers H3 needs at full precision.

The inputs

Just one: unet_name, a dropdown filtered to files matching minimax_h3 + int8 + convrot in ComfyUI/models/diffusion_models. Drop the FL2VA or Ref2VA file in there and it appears. One output: a MODEL.

The surrounding workflow matters more than the node itself. The pack's recommended chain:

MiniMaxH3INT8FastLoader
  → LoraLoaderBypassModelOnly (optional, for Turbo LoRAs)
  → ModelAttentionBackend (comfy kitchen attention)
  → MiniMaxH3FastPatch / MiniMaxH3BlockCache (optional accelerators)
  → MiniMaxH3SigmaShift
  → guider / sampler

Two specifics: use LoraLoaderBypassModelOnly rather than the regular LoRA loader for a quantized base, because ComfyUI-format H3 LoRAs add their residual without rewriting the INT8 weights; and if you're running lightx2v's Turbo LoRAs, match FL2V to FL2VA / Ref2V to Ref2VA and use their 4- or 8-step schedules, not the pack's standard 20-step validation schedule.

Install

ComfyUI Manager → search ComfyUI-INT8-Fast-ROCM-ConvRot → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/DrBearJew/ComfyUI-INT8-Fast-ROCM-ConvRot

The pack declares no pip dependencies, but this node is not self-contained: it uses ComfyUI's quant-ops registry from comfy-kitchen and needs a ROCm PyTorch build - the code literally raises a runtime error if torch.version.hip is absent. Validated stack is Linux, ComfyUI 0.32.0, comfy-kitchen 0.2.31 on an RX 7900 XTX (gfx1100). Other GPUs and ComfyUI versions are, per the README, unvalidated. If you're on NVIDIA, this node isn't for you - the broader INT8-Fast pack works there, but the H3 fast loader is ROCm-gated by design.

Before you install, read the license

MiniMax H3's community license excludes the EU, UK, South Korea and the US from its applicable territory. No amount of clever loading gets you around that - if you're in one of those regions, you're not licensed to run the local weights at all, whatever the node does. The node will work; the license question is yours to answer.

Where people get burned

• A mismatched checkpoint errors out fast. The validation is strict on purpose - if it's not INT8 ConvRot with FP32 heads, you get a clear ValueError instead of silent garbage output. Don't try to feed it a bf16 file.
• Wrong folder, empty dropdown - diffusion_models, not checkpoints.
• The fast backend silently falls back. Only the ConvRot projection shapes (groupsize 256) take the Triton path; anything else routes to HIP/eager. That's by design and invisible - it's how the precision boundaries survive.