File size: 1,800 Bytes
8887c71 d741544 8887c71 | 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 36 37 38 39 | {
"name": "FastH3 NVFP4 VAE decoder (fast-8L)",
"format": "fastvideo_h3_decoder_nvfp4_deploy_v1",
"architecture": "MiniMaxH3VideoViTDecoder3d",
"decoder_blocks": 8,
"hidden_size": 2048,
"nvfp4": {
"scheme": "W4A4 NVFP4 (E2M1 values, E4M3 per-16 block scales, FP32 per-tensor scale)",
"block_size": 16,
"activation_global_scale": "unit",
"rotation_group": null,
"quantized_linears": 48,
"quantized_layers": "transformer_blocks.*.{attn.to_q,attn.to_k,attn.to_v,attn.to_out.0,ff.net.0.proj,ff.net.2}",
"high_precision": "proj_in, proj_out, norms, register tokens, post_quant_conv"
},
"compute_dtype": "bfloat16 autocast",
"host_vae": "any MiniMax-H3 video VAE with >= 8 decoder blocks (e.g. the 26-block light VAE of the FastH3 V2/Trim stacks); the loader keeps 8",
"tensor_count_by_dtype": {
"float32": 89,
"bfloat16": 96
},
"nvfp4_linear_parameters": 537083904,
"source_step": 5000,
"files": {
"decoder.safetensors": {
"format": "fastvideo_h3_decoder_nvfp4_deploy_v1",
"contents": "NVFP4 linears as bf16 master weights, packed to NVFP4 at load; other tensors in their trained dtype",
"bytes": 1099886160,
"use": "reference weights, re-quantization, further training"
},
"decoder-nvfp4.safetensors": {
"format": "fastvideo_h3_decoder_nvfp4_packed_v1",
"contents": "pre-packed NVFP4: per linear packed_weight (uint8, two E2M1 codes per byte), weight_inv_scale (E4M3 per-16 block scales, FlashInfer SfLayout.layout_128x4 swizzle), weight_global_sf (fp32), bias (bf16), input_amax (fp32); other tensors in their trained dtype (fp32)",
"bytes": 328147232,
"use": "smallest download and lowest load memory; Blackwell only; decodes bit-identically to decoder.safetensors"
}
}
}
|