SoulX-LiveAct NVFP4 for RTX 5090

Prequantized NVIDIA NVFP4 checkpoint for the unofficial SoulX-LiveAct RTX 5090 optimization fork.

This repository contains only the quantized LiveAct DiT checkpoint. The runtime code, setup instructions, Wav2Vec model, T5, VAE, CLIP, compiler cache, prompt cache, avatar cache, and sample media are not included.

Important compatibility note

This is a custom deployment checkpoint, not a drop-in Transformers or Diffusers model. It must be loaded with the NVFP4 checkpoint loader in the linked RTX 5090 fork. The tested kernel path targets NVIDIA Blackwell SM120.

Base model and provenance

• Base model: Soul-AILab/LiveAct
• Base-model revision: 1bcdef1d8941edf2a9a6e1d21f5657f0a69fcd37
• Quantization relationship: derived NVFP4 quantization
• Source config.json SHA-256: 63810cd77a5acb72a10aed4394565b75a76b229d7838b4996a32022bd4ae602e
• Checkpoint schema: soulx-liveact-nvfp4, version 1

The base model page identifies its model license as Apache-2.0. Review the base model card and license before redistribution or use. The GitHub source fork has separate licensing notices; see its NOTICE.md.

Checkpoint contents

• 6 safetensors shards
• 600 quantized blocks.* Linear modules
• 2,913 tensors
• Tensor payload: 11,088,486,368 bytes (approximately 10.33 GiB)
• Packed weight dtype: uint8
• Weight scale dtype: float8_e4m3fn
• Global scale dtype: float32
• Values per scale: 16

fp4_manifest.json records every shard size and SHA-256 checksum. export_result.json records the export result and runtime metadata.

Tested export environment

• GPU: NVIDIA GeForce RTX 5090 32 GB
• GPU architecture: SM120
• CUDA: 12.8
• PyTorch: 2.8.0+cu128
• Python: 3.12.3
• Export time: approximately 752 seconds

Download

Install the current Hugging Face CLI and download into the directory expected by the runtime:

python -m pip install -U huggingface_hub hf_xet

cd /path/to/SoulX-LiveAct-RTX5090
hf download 5461gpt/SoulX-LiveAct-NVFP4-RTX5090 \
  --local-dir models/LiveAct-NVFP4

The original BF16 model and Wav2Vec checkpoint must be downloaded separately as described in the RTX 5090 reproduction guide.

Verify

After downloading both the base model and this checkpoint:

cd /path/to/SoulX-LiveAct-RTX5090

python scripts/verify_fp4_checkpoint.py \
  --fp4_ckpt_dir models/LiveAct-NVFP4 \
  --source_ckpt_dir models/LiveAct \
  --device cpu

Do not start the GUI if verification reports a manifest, source-config, shard size, or checksum mismatch.

Run

On Windows, from the project directory:

run_gui_fp4_prequant_5090.cmd

Recommended baseline:

• Resolution: 256x416
• DiT steps: 3
• Output: 15 FPS
• FP8 KV cache
• VAE temporal cache
• Persistent prompt/avatar/compile caches

See the matching GitHub v0.1.0-rtx5090 release for complete setup and runtime documentation.

Measured warm performance

Measured on the tested RTX 5090 system:

• Startup warmup: approximately 60 seconds
• First chunk ready: approximately 2.11 seconds
• HLS playlist ready: approximately 2.28 seconds
• Steady aggregate RTF: approximately 1.039

Results depend on the complete software stack, cache state, storage placement, input duration, and generation settings. They are not guaranteed for other systems.

Limitations and responsible use

• Designed and tested for the custom SM120 NVFP4 runtime in the linked fork.
• Quality and timing can differ from the original BF16 model.
• TorchInductor and CUDA graph caches are intentionally not distributed.
• Users must have the rights and consent required for all portraits, voices, prompts, and generated media.
• Clearly disclose synthetic media where appropriate.

中文摘要

這是 SoulX-LiveAct 的 RTX 5090／SM120 NVFP4 預量化 DiT checkpoint，必須搭配上方 GitHub fork 的自訂 loader 使用，不能直接以一般 Transformers 或 Diffusers API 載入。 下載後請放在 models/LiveAct-NVFP4/，並先執行 checksum 驗證，再啟動 GUI。