danielhanchen's picture
Upload MiniMax-H3-Image INT8 and FP8
6290129
Raw History Blame Contribute Delete
2.29 kB
unsloth/MiniMax-H3-Image-FP8
"MiniMax-H3-Image is licensed under the MiniMax-H3-Image Community License Agreement,
Copyright © 2026 MiniMax. All Rights Reserved."
The weights here derive from MiniMax-H3-Image
(https://huggingface.co/MiniMaxAI/MiniMax-H3-Image).
LICENSE is the text MiniMax ships at
https://huggingface.co/MiniMaxAI/MiniMax-H3-Image/blob/main/LICENSE, verbatim.
Modified files and the changes made to them (MiniMax-H3-Image Community License
Agreement, Section III.2). Both files start from
minimax_h3_image_pruned_bf16.safetensors as published by Comfy-Org
(https://huggingface.co/Comfy-Org/MiniMax-H3-Image), which is the MiniMax-H3-Image
diffusion transformer with the dense adaLN modulation projection replaced by a
curve-form factorization of the time-embedding curve. That pruning is Comfy-Org's
work and not Unsloth's. It changes the modulation numerics: at the 6 sampling steps
the curve-form adaLN output differs from the dense release adaLN by a relative error
of 6.5e-4 to 7.5e-4 (cosine 0.9999998).
* MiniMax-H3-Image-INT8-ConvRot.safetensors: 313 Linear weights quantized to
signed 8-bit integers with per-output-channel absmax scales, after a ConvRot
block-Hadamard rotation at group size 256 applied to all 313, so the weights are
stored in a rotated basis. This changes the numerics. proj_in, the curve table,
norm_out.linear and the adaLN path are kept in fp32. Every other tensor is copied
through bit-identical at its source precision.
* MiniMax-H3-Image-FP8.safetensors: the same 313 Linear weights quantized to
float8 e4m3 with per-output-channel absmax scales, which changes the numerics.
No rotation is applied. proj_in, the curve table, norm_out.linear and the adaLN
path are kept in fp32. Every other tensor is copied through bit-identical at its
source precision.
Both files also carry the per-step output heads from the release's
pdd_fused_heads.safetensors, stored under the names pdd_proj_out.0 to pdd_proj_out.5
in fp32 with their values unchanged; only the tensor names and the container differ.
Both are serialized as safetensors, with metadata recording the model family, the
base model id and the base and heads revisions.
Not an official MiniMax or Comfy-Org product, and not endorsed by either.