Downloads · 30 days
923
100% of all-time downloads
MiniMax-H3-Singularity-Ref2VA-NVFP4
wukaikevin/MiniMax-H3-Singularity-Ref2VA-NVFP4
MiniMax-H3-Singularity-Ref2VA-NVFP4 is a machine learning model from wukaikevin. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for minimax-h3. The card lists the license as apache-2.0.
WarmBloodAban/Minimax-h3Singularity 的 Ref2VA Pruned v1.3 (int8) 转 NVFP4 版。
All-time downloads
923
Public
Repo size
12.5 GB
Likes
1
Public
Repo makeup
Click a slice to open those files.
.safetensors12.5 GB · 100%
At a glance
- Library
- minimax-h3
- License
- apache-2.0
- Access
- Public
- Created
- Sep 15, 2026
- Updated
- Sep 15, 2026
- SHA
- 081a0fdc
Model card
From the Hugging Face model README
MiniMax-H3 Singularity Ref2VA — NVFP4 (pruned)
WarmBloodAban/Minimax-h3_Singularity 的 Ref2VA Pruned v1.3 (int8) 转 NVFP4 版。
格式
- 权重: NVFP4 E2M1, hi_lo nibble 打包 (
blocks.N.{attn.qkv_proj,attn.out_proj,mlp.fc1,mlp.fc2}.weightuint8 [out, in/2]) - 组缩放: 每 16 元素 fp8e4m3 (
weight_scale), cuBLAS blocked(128×4 tile) 布局 — 用comfy_kitchen.float_utils.from_blocked解码为线性布局, 或原样直喂torch._scaled_mm/ cuBLAS NVFP4 GEMM - 张量缩放: f32 标量 (
weight_scale_2) - 非量化部分 (adaln/norm/refiner/顶层) bf16 直拷,
adaln_t_tablef32 - 转换时已撤销 convrot Hadamard 旋转 (int8 源是旋转基, 本权重是原始基)
转换验证
- 权重级 cos(fp64) = 0.9955 (对 int8 反量化参考)
- 端到端: 有效视频生成验证通过
声明
- 原模型: WarmBloodAban/Minimax-h3_Singularity (微调质量/许可归原作者)
- 基础模型: MiniMaxAI/MiniMax-H3
- 本仓库仅包含量化格式转换产物
- Library
- minimax-h3
- License
- apache-2.0
- Created
- Sep 15, 2026
- Updated
- Sep 15, 2026