File size: 2,219 Bytes
3da88e3 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 | # Modifications and attribution
Community MLX/Metal adaptation and distribution by Devin Lai, November 2026.
This is an independent port, without MiniMax endorsement. No model pretraining,
fine-tuning, new training data or quality calibration was performed.
- **BF16:** original transformer files, retained byte for byte. Native BF16
matrices and FP32 precision islands remain at their original precision.
- **Packed12:** the fifty main DiT blocks use the lossless `h3-palette12-v1`
layout. BF16 payloads are reconstructed before the native matrix operation.
QKV storage follows the MLX Q/K/V ordering. This is storage compression,
not 12-bit quantization. Native refiners are shared.
- **Q8 / Q4:** the fifty main blocks and two token-refiner blocks use MLX
affine weight quantization, group size 64, with the required QKV reordering.
Native support tensors and FP32 precision islands remain unquantized.
Quantization changes generated images and can be slower than BF16.
- **Shared support:** native constants, normalization weights and native token
refiners are repacked. The original timestep projections are evaluated once
for the six released PDD steps, plus the fixed editing anchor. There is no
interpolation or timestep skipping. These frozen tables were evaluated on
Apple M5 Pro with MLX 0.32.4 and strict FP32 operations; the published loader
checks the schedule, MLX version and precision environment.
- **VAE and output heads:** original files, unchanged. The default runtime
follows the inspected BF16 decoder policy; its source weights are FP32.
- **Text encoder:** downloaded separately from the pinned official
`MiniMaxAI/MiniMax-H3` repository. It is not relicensed by this project.
The inference implementation follows the inspected SGLang and Transformers
contracts. Apache-2.0 applies to this project's runtime code, with source
identities in `provenance/sources.json`. Model weights and derivatives retain
the MiniMax-H3-Image Community License Agreement. The separate text encoder
retains the MiniMax H3 Community License Agreement. See `LICENSE-model`,
`LICENSE-text-encoder` and `NOTICE` in the main repository, and `LICENSE` in
each model component repository.
|