mu训练和推理对齐问题

[PlutoQyl]

def set_timesteps_flux2(num_inference_steps=100, denoising_strength=1.0, dynamic_shift_len=None):
        sigma_min = 1 / num_inference_steps
        sigma_max = 1.0
        num_train_timesteps = 1000
        sigma_start = sigma_min + (sigma_max - sigma_min) * denoising_strength  # 1.0
        sigmas = torch.linspace(sigma_start, sigma_min, num_inference_steps)  # [1.0000, 0.9990, ..., 0.0010]
        if dynamic_shift_len is None:  # TODO：看起来这是一个调整的参数mu，用于调整动态_shift_len的范围
            # If you ask me why I set mu=0.8,
            # I can only say that it yields better training results.
            mu = 0.8
        else:
            mu = FlowMatchScheduler.compute_empirical_mu(dynamic_shift_len, num_inference_steps)
        sigmas = math.exp(mu) / (math.exp(mu) + (1 / sigmas - 1))  # [1.0000, 0.9996, ..., 0.0022]
        timesteps = sigmas * num_train_timesteps
        return sigmas, timesteps

在跑flux2klein训练脚本时发现，训练时候dynamic_shift_len是None，但推理时为dynamic_shift_len=height // 16 * width // 16。这里为啥不需要对齐？

[Artiprocher]

@PlutoQyl 因为训练时如果启用 dynamic_shift_len，那么 timestep 的分布会变得异常，训练效果非常糟糕。