bf16, fp16, fp8_scaled, fp8 and fp4 quants for https://huggingface.co/vafipas663/HY-OmniWeaving_repackaged
Also includes fp8_scaled and fp8 text encoder
Only contains converted files, if you want the fp32 model or bf16 text encoder, see the repackaged repo's diffusion_models/text_encoders or original repo's transformer/text_encoder
Original model page: https://huggingface.co/tencent/HY-OmniWeaving
Not a lot of testing has been done for the quants, although most should work fine. There could possibly be issues with fp4, all double blocks have been quantized (targetting attention, mlp and mod weights)
Confirmed, at the time of writing (The PR which adds HY-OmniWeaving support to ComfyUI has not been merged yet), text encoder quants are fine, but quantization config relying transformer quants don't work yet. Fp8 scaled will give noise and nvfp4 will just crash. fp8_e4m3fn will work if you want 8-bit for now.
- Downloads last month
- 436