jdopensource

/

JoyAI-Video-Edit

like 100

Follow

JD.com Open Source 401

FP8 is already the default execution path — it just happens at load time rather than as a separate checkpoint:

• the released weights are bf16 (dit/joyai_video_edit_dit_0811.pth, ~32.5 GB);
• at startup every DiT Linear is quantized to FP8 e4m3 (per-output-channel weight scales + per-token activation quantization) and runs through the cutlass FP8 GEMM in deploy/joyomni_ops; JOYOMNI_FP8_IMG=1 / JOYOMNI_FP8_TXT=1 are the defaults;
• with JOYOMNI_LOW_VRAM=1 the weights are staged to the GPU block-by-block and quantized as they land, so the full bf16 model is never GPU-resident — that's how it runs on a single RTX 5090 (32 GB): ~21.5 GiB steady at 840×480 @ 24 FPS (see DEPLOYMENT.md);
• optional INT8 attention on GeForce via SageAttention (JOYOMNI_SAGE_ATTN=1).

A pre-quantized FP8 checkpoint would mainly cut the download/disk footprint (~32.5 GB → ~16 GB), not VRAM or inference speed — the runtime already gives you those. It's not on the short-term roadmap, but if the download size is the main pain point for you, that's useful signal — let us know.