YossiIsr/joyai-video-edit-fp8

JoyAI-Video-Edit, FP8 (RTX 5090 serving layout)

An FP8 conversion by Plotin of JoyAI-Video-Edit (DiT checkpoint dit/joyai_video_edit_dit_0811.pth), packaged with its VAE and text encoder so one repo holds everything the streaming server loads. Not affiliated with or endorsed by JD or Xiaomi.

Path	What	From
dit/joyai_video_edit_dit_0811_fp8.safetensors	DiT with the img/txt attention and MLP Linears in FP8 e4m3, per-channel scales; other tensors bf16	converted from JoyAI-Video-Edit dit/joyai_video_edit_dit_0811.pth
vae/	unchanged	JoyAI-Video-Edit vae/
text_encoder/	unchanged (weights, tokenizer, configs, model card)	MiMo-VL-7B-RL-2508

What the conversion is. The upstream server converts these Linears to FP8 itself every time it loads (Fp8Linear.from_linear, JOYOMNI_FP8_IMG=1 JOYOMNI_FP8_TXT=1). This file stores that result, made with JOYOMNI_FP8_FAST_ACCUM=1 (the RTX 5090 recipe), so a server start skips reading the 32.5 GB bf16 checkpoint and quantizing it. FP8 weights are stored transposed (row-major); the file metadata records the format (joyai-fp8-presaved-v1) and the fast-accum setting. Loading it needs a small patch to the upstream server (build the model on the meta device, fill tensors from this file); the stock server cannot load it.

Licenses

• JoyAI-Video-Edit (DiT, VAE): Apache License 2.0, see LICENSE and NOTICE.
• MiMo-VL-7B-RL-2508 (text_encoder/): MIT, as declared on its model card; see text_encoder/LICENSE.