Eliovp's picture
Add modifications notice, config and manifest
80499d4 verified
|
Raw History Blame Contribute Delete
2.02 kB

Modified files

Every file below except tokenizer/tokenizer.json is a modified file of MiniMax-H3-Image (Section III.2 of the MiniMax-H3-Image Community License). Source: MiniMaxAI/MiniMax-H3-Image, revision 00dd081ebdd33315adf2c4d84f64aa327ddf3039. Changes made by ElioVP (Paiton), 11 Nov 2026.

File Made from Change
dit/dit.bin, dit/dit.idx transformer/, pdd_fused_heads.safetensors, pdd_config.json The attention and feed-forward projections of the 50 transformer blocks and the 2 token-refiner blocks are rounded to FP8 (e4m3) with one scale per output channel, using GPTQ rounding calibrated on 887,724 tokens from 18 prompts. Rows are reordered for the Paiton kernels. The timestep embedding and the AdaLN and output-norm modulation are evaluated in advance for the fixed 6-step schedule and stored as tables. The input, output and text projections are stored in float32. Packed into one file with an index.
dit/vae_latents.bin vae/config.json The latent mean and standard deviation, as float32
te/L00.bin to te/L49.bin text_encoder/ Decoder layers 0 to 49 of the Qwen3-VL text encoder, projections rounded to FP8 (e4m3) with one scale per output channel (round to nearest); the other layer weights are repacked unchanged. Layers 50 to 63 and the language-model head are left out because the model reads the hidden state after layer 50.
te/embed.bf16, te/vision.bin text_encoder/ Token embedding and vision tower, bf16, repacked
vae/dec8.bin vae/diffusion_pytorch_model.safetensors VAE decoder, block weights in int8 with one scale per output channel, the rest in bf16 and float32; post_quant_conv folded into the input projection
vae/enc.bin vae/diffusion_pytorch_model.safetensors VAE encoder in bf16 and float32, repacked
tokenizer/tokenizer.json tokenizer/tokenizer.json Unmodified