Qwen-Image-2.1-Turbo-FP8

This is an FP8 / INT8 quantized version of Qwen-Image-2.1-Turbo, the 8-step distilled checkpoint of Qwen-Image-2.1. - Each file replaces the 14.2 GB bf16 `transformer` with a 7.1 to 7.3 GB one. - Same-seed LPIPS vs the bf16 Turbo model (lower is better): 0.037 INT8-ConvRot, 0.089 INT8, 0.125 FP8. INT8-ConvRot is the recommended file. - Turbo shares its text encoder and VAE with Qwen-Image-2.1 (text encoder byte-identical, VAE identical in bf16), so the FP8 text encoder and bf16 VAE in unsloth/Qwen-Image-2.1-FP8 work with these files as is. Built from `Qwen/Qwen-Image-2.1-Turbo` at revision `d65dbc9a7e8f6b5479e33dee6030eaab2a906509` with the same recipe as unsloth/Qwen-Image-2.1-FP8: norms, modulation, timestep embeddings and (for INT8) `txt_in` stay bf16, every other linear is quantized. Scores are LPIPS (AlexNet) against the bf16 Turbo transformer at the card's settings (8 steps, CFG 1, the shipped sigma schedule, 1024x1024), 8 prompts x 2 seeds, with the same bf16 text encoder in every run. Everything here is safetensors rather than a pickle, so loading runs no arbitrary code. Needs a diffusers main build with pipeline-configured sampling sigmas (PR #14950).

译文尚未就绪，显示来源原文