{
    "max_new_tokens": 756,
    "seed": -1.0,
    "temperature": 0.7,
    "top_p": 0.9,
    "top_k": 20,
    "typical_p": 1,
    "epsilon_cutoff": 0,
    "eta_cutoff": 0,
    "repetition_penalty": 1.15,
    "repetition_penalty_range": 0,
    "encoder_repetition_penalty": 1,
    "no_repeat_ngram_size": 0,
    "min_length": 1,
    "do_sample": true,
    "penalty_alpha": 0,
    "num_beams": 1,
    "length_penalty": 1,
    "early_stopping": false,
    "mirostat_mode": 0,
    "mirostat_tau": 5,
    "mirostat_eta": 0.1,
    "add_bos_token": true,
    "ban_eos_token": false,
    "truncation_length": 2048,
    "custom_stopping_strings": "",
    "skip_special_tokens": true,
    "stream": true,
    "tfs": 1,
    "top_a": 0,
    "textbox": "<human>My wife's going to bed soon. What do you think about that?\n\nThe AI responded: It seems like a reasonable decision for your wife to make if she is feeling tired and needs rest. Ensuring adequate sleep is important for overall health and well-being. Is there anything I can assist with or provide information on to support her in this endeavor?\n\n<human>Can you remember what she was doing?\n\nThe AI responded: Sure, can you please specify the context in which you would like me to recall your wife's previous activities? Without any specific timeframe or event being mentioned, it would not be possible for me to accurately recall her recent actions.\n\n<human>What is she preparing for?\n\nThe AI responded: To better understand what your wife may be preparing for, could you please share some additional details or provide more context so that I can assist you in a more effective manner?",
    "loader": "Transformers",
    "cpu_memory": 110000,
    "auto_devices": true,
    "disk": false,
    "cpu": false,
    "bf16": true,
    "load_in_8bit": false,
    "trust_remote_code": false,
    "load_in_4bit": true,
    "compute_dtype": "bfloat16",
    "quant_type": "nf4",
    "use_double_quant": false,
    "wbits": "None",
    "groupsize": "None",
    "model_type": "None",
    "pre_layer": 0,
    "triton": false,
    "desc_act": false,
    "no_inject_fused_attention": false,
    "no_inject_fused_mlp": false,
    "no_use_cuda_fp16": false,
    "threads": 0,
    "n_batch": 512,
    "no_mmap": false,
    "low_vram": false,
    "mlock": false,
    "n_gpu_layers": 0,
    "n_ctx": 2048,
    "llama_cpp_seed": 0.0,
    "gpu_split": "",
    "max_seq_len": 2048,
    "compress_pos_emb": 1,
    "alpha_value": 1,
    "gpu_memory_0": 20800
}