FLUX.2-dev训练显存配置

请问Flux.2-dev训练1024*1024分辨率图像大概需要多少显存，我使用diffsynth-2.0分支示例脚本在80G单卡A800会OOM。
补充：不使用fp8_models之前，max_pixels设为较小值104857（从1048576改为104857）能够运行，添加fp8_models参数，反而会OOM

另外parse_model_configs函数在使用model_paths指定模型时，fp8_models和offload_models使用相同方式指定会无法生效，可能需要改成以下形式：

def parse_model_configs(self, model_paths, model_id_with_origin_paths, fp8_models=None, offload_models=None, device="cpu"):
        model_configs = []
        if model_paths is not None:
            model_paths = json.loads(model_paths)
            fp8_models = [] if fp8_models is None else json.loads(fp8_models)
            offload_models = [] if offload_models is None else json.loads(offload_models)
            for path in model_paths:
                vram_config = self.parse_vram_config(
                    fp8=path in fp8_models,
                    offload=path in offload_models,
                    device=device
                )
                model_configs.append(ModelConfig(path=path, **vram_config))
        if model_id_with_origin_paths is not None:
            model_id_with_origin_paths = model_id_with_origin_paths.split(",")
            fp8_models = [] if fp8_models is None else fp8_models.split(",")
            offload_models = [] if offload_models is None else offload_models.split(",")
            for model_id_with_origin_path in model_id_with_origin_paths:
                model_id, origin_file_pattern = model_id_with_origin_path.split(":")
                vram_config = self.parse_vram_config(
                    fp8=model_id_with_origin_path in fp8_models,
                    offload=model_id_with_origin_path in offload_models,
                    device=device
                )
                model_configs.append(ModelConfig(model_id=model_id, origin_file_pattern=origin_file_pattern, **vram_config))
        return model_configs

Start at parse_model_configs and the DiffSynth-2.0 example script, comparing model_paths handling with model_id_with_origin_paths for fp8_models and offload_models. Reproduce the 1024×1024 FLUX.2-dev configuration on an 80G A800, then verify that the model options take effect without an unexpected OOM and document the required VRAM.