MiniMax-H3 — combined (LoRA-fused · 8-bit conditioner · AoTI)

Joint video and synchronized soundtrack from a single denoising pass, packaged as a deployment build of MiniMaxAI/MiniMax-H3 (diffusers modular layout). This repo is tuned for Hugging Face ZeroGPU Spaces: the whole checkpoint is ~112 GB on the Hub (under the 150 GB per-Space storage quota), and each half of the split pattern fits the 95 GiB ZeroGPU card with room to spare — so it runs comfortably and fast, at ~40 s for a 5 s clip instead of the 5+ minutes a 30-step base run pays.

Full, unmodified bf16 partitions (including the 62 GiB bf16 conditioner) remain available in the private backup addlabsviral/minimax-h3-combined-bkup.

Repo layout

transformer/ — LoRA-fused, 4-step inference

The larryvrh/MiniMax-H3-Turbo-Lora Turbo checkpoint (minimax_h3_turbo_4step_ema_ckpt850, the final EMA round, sharp at 4 steps) is baked into the bf16 weights at conversion time — scale * (lora_B @ lora_A) is added into every targeted projection (fused-QKV thirds, SwiGLU gate/value swap, AdaLN rows), so an adapter wrap at load time is unnecessary and the checkpoint serves 4-step inference out of the box.

• 4 steps is the design point; 6 is a comfort setting, 28 is the unturbo fallback.
• Guidance-distilled: no CFG, no negative prompt — every step is one forward pass.
• 24 fps, num_frames of the form 17 * n + 5 (2–15 s).

text_encoder/ — 8-bit bitsandbytes

The 62 GiB bf16 Qwen3-VL conditioner is served from a single safetensors file with a bitsandbytes LLM.int8 quantization_config (load_in_8bit: true, threshold 6.0, no skip modules). from_pretrained applies the quantization automatically at load, so the conditioner runs in about a third of the memory and there is no manual quantize step. Requires bitsandbytes and transformers==5.8.0.

aoti/ — compiled transformer blocks

One AOT-Inductor MiniMaxH3TransformerBlock package, shared byte-identically by all 50 blocks, exported with a dynamic sequence dimension (original key bf16/torch2.11/sm120/dynamic, flattened to aoti/, with the loader layout aoti/package/submodules/transformer_blocks/ preserved). It removes ~0.5 s/step of kernel-launch and epilogue overhead and is card-specific: built for the ZeroGPU pool's RTX PRO 6000 (sm120) cards with torch 2.11. Everything else falls back to eager safely.

Usage

import torch
from diffusers import MiniMaxH3ModularPipeline

pipe = MiniMaxH3ModularPipeline.from_pretrained("addlabsviral/minimax-h3-combined")
pipe.load_components()          # text encoder quantizes to 8-bit per its config

state = pipe(
    prompt="a cinematic drone shot over a neon city at night",
    height=544, width=960,
    num_frames=124,             # 17 * n + 5
    num_inference_steps=4,      # the Turbo LoRA's design point
    generator=torch.Generator("cpu").manual_seed(42),
)

For the split ZeroGPU pattern (conditioner Space + denoiser/reconstructor Space), point H3_MODEL_REPO at this repo and H3_AOTI at the aoti/ package; modular_model_index.json already resolves transformer/ and text_encoder/, so load_components fetches exactly the subfolders each half declares.

Measured on ZeroGPU (rtx-pro-6000, bf16, everything resident)

Provenance & license

• Base checkpoint: MiniMaxAI/MiniMax-H3, MiniMax-H3 Community License (see LICENSE).
• Turbo LoRA folded into the transformer: larryvrh/MiniMax-H3-Turbo-Lora.
• AoTI package: compiled from this same bf16 partition for torch 2.11 / sm120 / dynamic sequence.