Enable Online FP8&NF4 on Minimax-H3

[Uxito-Ada]

Summary

Add online DiT quantization for MiniMax H3 with two backends:

• TorchAO FP8 dynamic activation and weight quantization
• bitsandbytes NF4 weight-only quantization with BF16 compute

They convert 258 Linear layers of MiniMax H3 across the main and token-refiner transformer blocks while preserving the reference dtype of the FP32 projections, text encoder, and VAEs.

Motivation

MiniMax H3's BF16 DiT profile requires most of an 80 GB H100. Online quantization provides a practical single-GPU
deployment option.

Implementation

• Add MiniMaxH3DiT.enable_quant() dispatch for TorchAO FP8 and BNB NF4.
• Add quantization loading for CUDA, single-GPU execution, and FSDP exclusion.
• Add dedicated H100 examples:
  • examples/minimax_h3/minimax_h3_fl2va_torchao_fp8_h100.py
  • examples/minimax_h3/minimax_h3_fl2va_bnb_nf4_h100.py
• Add --quantization to the existing FL2VA, Ref2VA, and JSON request examples.
• Documents.

Benchmark

The matched benchmark uses one NVIDIA H100 80 GB, MiniMax H3 FL2VA, 768p 16:9 output, five seconds, 50 inference
steps, seed 0, and the prompt:

tf-kernel FP8 shows 5.6% faster of denoising and 9.9% less memory than torchao FP8, while its one-time quantization is much slower than torchao, which leads to higher cost of a single e2e inference. @lzx1413

Generated Video Comparison

prompt: "Steam rises from the ramen while the family talks in the background."

Precision	Video
BF16	https://github.com/user-attachments/assets/72afe0cf-99b0-4a07-903f-94a412ef43d9
torchao FP8	https://github.com/user-attachments/assets/1361ecb6-0a62-48a3-8b7b-2b3b9c7c55d5
NF4	https://github.com/user-attachments/assets/1993e6c0-fe70-40bf-baed-8b16f6442fc3
tf-kernel FP8	https://github.com/user-attachments/assets/8e958cd5-3ad9-45fa-a3df-c9a4c103d098

It is seen that tf-kernel FP8 presents much more vibrant views than BNB NF4, but it is inconsistent to the original BF16 output, whose content diverges from the prompt. I suggest torchao as the default FP8 for Minimax-H3, and tf-kernel as an option. @lzx1413

Checklist

• Code follows the project's coding standards (ruff)
• Pre-commit hooks pass (pre-commit run --all-files)
• All tests pass (pytest tests/)
• New tests added for new functionality
• Documentation updated (README, CLAUDE.md, docstrings)
• Commit messages are clear and descriptive
• PR title follows the convention: [TYPE] Brief description

[lzx1413]

LGTM