Some improvements to config.json readability

[francoishernandez]

These are minimal changes to improve dumped config readability:

• prevent dumping empty transforms_configs (change taken from Inference server, lots of related changes #42);
• dump "non-nested" fields before nested ones, not much but prevents having interspersed items that are difficult to spot;

Example with llama3 convert:

Before

{
  "n_sample": 0,
  "training": {
    "accum_count": [
      32
    ],
    "w_bit": 0,
    "valid_batch_size": 256,
    "batch_size": 896,
    "model_dtype": "fp16",
    "batch_size_multiple": 1,
    "normalization": "tokens",
    "quant_layers": [],
    "quant_type": "",
    "group_size": 0,
    "accum_steps": [
      0
    ],
    "optim": "fusedadam",
    "batch_type": "tokens"
  },
  "decoder_start_token": "<s>",
  "src_vocab": null,
  "tgt_vocab": null,
  "skip_empty_level": "silent",
  "model": {
    "add_ffnbias": false,
    "architecture": "transformer_lm",
    "add_qkvbias": false,
    "num_experts": 0,
    "heads": 32,
    "embeddings": {
      "src_word_vec_size": 4096,
      "tgt_word_vec_size": 4096
    },
    "mlp_activation_fn": "gated-silu",
    "rotary_interleave": false,
    "left_pad": true,
    "model_type": "text",
    "rotary_theta": 500000,
    "sliding_window": 0,
    "layer_norm": "rms",
    "layers": 32,
    "rotary_dim": 0,
    "num_experts_per_tok": 0,
    "parallel_residual": false,
    "decoder": {
      "decoder_type": "transformer_lm",
      "tgt_word_vec_size": 4096
    },
    "shared_layer_norm": false,
    "heads_kv": 8,
    "hidden_size": 4096,
    "transformer_ff": 14336,
    "norm_eps": 1e-05,
    "max_relative_positions": -1
  },
  "transforms_configs": {
    "filtertoolong": {
      "src_seq_length": 512,
      "tgt_seq_length": 512
    }
  },
  "vocab_size_multiple": 8,
  "share_vocab": true,
  "save_data": null,
  "data": null,
  "src_vocab_size": 128256,
  "transforms": [
    "filtertoolong"
  ],
  "data_task": "lm",
  "tgt_vocab_size": 128256
}

After

{
  "transforms": [
    "filtertoolong"
  ],
  "src_vocab_size": 128256,
  "decoder_start_token": "<s>",
  "share_vocab": true,
  "vocab_size_multiple": 8,
  "save_data": null,
  "skip_empty_level": "silent",
  "data": null,
  "tgt_vocab": null,
  "tgt_vocab_size": 128256,
  "n_sample": 0,
  "src_vocab": null,
  "training": {
    "group_size": 0,
    "w_bit": 0,
    "quant_layers": [],
    "valid_batch_size": 256,
    "accum_steps": [
      0
    ],
    "normalization": "tokens",
    "model_dtype": "fp16",
    "batch_size": 896,
    "quant_type": "",
    "batch_type": "tokens",
    "accum_count": [
      32
    ],
    "batch_size_multiple": 1,
    "optim": "fusedadam"
  },
  "transforms_configs": {
    "filtertoolong": {
      "tgt_seq_length": 512,
      "src_seq_length": 512
    }
  },
  "model": {
    "heads_kv": 8,
    "architecture": "transformer_lm",
    "rotary_theta": 500000,
    "mlp_activation_fn": "gated-silu",
    "left_pad": true,
    "parallel_residual": false,
    "hidden_size": 4096,
    "layer_norm": "rms",
    "norm_eps": 1e-05,
    "add_ffnbias": false,
    "sliding_window": 0,
    "transformer_ff": 14336,
    "layers": 32,
    "rotary_dim": 0,
    "rotary_interleave": false,
    "shared_layer_norm": false,
    "heads": 32,
    "num_experts": 0,
    "add_qkvbias": false,
    "num_experts_per_tok": 0,
    "max_relative_positions": -1,
    "embeddings": {
      "src_word_vec_size": 4096,
      "tgt_word_vec_size": 4096
    },
    "decoder": {
      "decoder_type": "transformer_lm",
      "tgt_word_vec_size": 4096
    }
  }
}