def mixgrpo_lora_diffrhythm2():
    config = get_config()
    gpu_number = 1
    if torch.cuda.is_available():
        gpu_number = torch.cuda.device_count()

    # --- Dataset ---
    config.dataset = os.path.join(config.root_dir, "_dataset")
    config.prompt_fn = "genius"
    config.json_path = os.path.join(config.dataset, "Genius_local.json")

    # --- Reward (placeholder; will be replaced with ACE-Step AAS later) ---
    config.reward_fn = {"placeholder": 1.0}
    config.eval_reward_fn = {"placeholder": 1.0}

    # --- MixGRPO Sliding Window ---
    config.training_strategy = "part"  # "part" = MixGRPO, "all" = DanceGRPO
    config.grpo = ml_collections.ConfigDict()
    config.grpo.iters_per_group = 20
    config.grpo.group_size = 4
    config.grpo.sample_strategy = "progressive"
    config.grpo.prog_overlap = False
    config.grpo.prog_overlap_step = 1
    config.grpo.roll_back = False
    config.grpo.max_iters_per_group = 10
    config.grpo.min_iters_per_group = 1

    # --- GRPO Training Hyperparameters ---
    config.train.total_steps = 2000
    config.train.clip_range = 0.1
    config.train.adv_clip_max = 5
    config.train.kl_coeff = 0.01
    config.train.learning_rate = 5e-6
    config.train.max_grad_norm = 1.0
    config.train.ema = False  # OOM mitigation: EMA model copy holds ~model size in VRAM
    config.train.gradient_accumulation_steps = 3
    config.train.num_inner_epochs = 1
    config.train.cfg = False
    config.train.num_generations = 12
    config.train.init_same_noise = True
    # Reward interpolation: reward = (1-λ)*audiobox + λ*AAS
    # λ=0 → pure Audiobox Aesthetics, λ=1 → pure AAS, λ=0.5 → equal mix
    config.train.reward_lambda = 0.5

    # --- Sampling ---
    config.sample.use_sde = True
    config.sample.num_steps = 32
    config.sample.guidance_scale = 4.5
    config.sample.noise_level = 1.0  # eta: 0.0 = ODE, 1.0 = fully stochastic SDE
    config.sample.rollout_duration_sec = 30
    config.sample.train_batch_size = 1
    config.sample.num_batches_per_epoch = 1
    config.sample.num_audio_per_prompt = 12
    config.sample.mini_num_audio_per_prompt = 1

    # --- LoRA ---
    config.use_lora = True
    config.train.lora_rank = 32
    config.train.lora_alpha = 64
    config.train.lora_path = None

    # --- Score Head (SDE) ---
    config.score_head_path = "_train/checkpoints_score_head/score_head_step_4000.pt"

    # --- Logging & Save Overrides ---
    config.save_freq = 100

    # --- Run name ---
    config.run_name = (
        f"mixgrpo_lora_diffrhythm_G{gpu_number}"
        f"_r{config.train.lora_rank}"
        f"_grp{config.train.num_generations}"
        f"_win{config.grpo.group_size}"
        f"_steps{config.sample.num_steps}"
    )

    # --- Debug Mode ---
    config.debug = False
    if config.debug:
        config.run_name = "DEBUG_" + config.run_name
        config.sample.num_batches_per_epoch = 1
        config.sample.train_batch_size = 1
        config.sample.num_audio_per_prompt = 2
        config.sample.mini_num_audio_per_prompt = 2
        config.train.num_generations = 2
        config.sample.rollout_duration_sec = 6
        config.sample.num_steps = 8
        config.grpo.group_size = 2
        config.grpo.iters_per_group = 3
        config.train.total_steps = 10
        config.save_freq = 1
        config.wandb_init = False
        config.per_prompt_stat_tracking = False

    config.save_dir = os.path.join(config.output_dir, config.run_name)
    config.case_name = config.run_name
    return config