LongCat-Video-Avatar-1.5-e4m3

FP8 (e4m3) quantized LongCat-Video-Avatar-1.5 — Meituan's audio-driven avatar video model, which animates a portrait from a speech track.

What was changed

Only the LongCatVideoAvatarTransformer3DModel is included here, quantized as described in quantization_config.json:

• weights: float8_e4m3fn, symmetric, per output channel (one fp32 scale per row)
• activations: float8_e4m3fn, dynamic per row
• matmul: torch._scaled_mm in RowWise mode (Hopper and newer)
• compute dtype: bf16
• final_layer.linear is skipped and stays bf16

Per-channel weight scales plus per-row activation scales cost slightly more bookkeeping than a single per-tensor scale, but they keep the output stable on a 48-layer, 4096-wide transformer where a per-tensor scale gets dragged around by outliers.

Usage

This repository contains the quantized transformer only. The text encoder, VAE and audio encoder come from the upstream repository; download both and point the upstream inference code at this transformer directory.

Running the fp8 path needs a GPU with torch._scaled_mm RowWise support (H100/H200 or newer). On older hardware, use the upstream bf16 weights.

License

MIT, inherited from LongCat-Video-Avatar-1.5.