post-training/configs/llama3-8b-dpo.yml

{
  "pipe_parallel_size": 0,
  "model_parallel_size": 4,
  "make_vocab_size_divisible_by": 1,

  # model settings
  "num_layers": 32,
  "hidden_size": 4096,
  "num_attention_heads": 32,
  "num_kv_heads": 8,
  # llama3 supports more than this but this is just for testing.
  "seq_length": 1024,
  "max_position_embeddings": 1024,
  "pos_emb": "rotary",
  "rotary_pct": 1,
  "rotary_emb_base": 500000,
  "rope_fusion": true,
  "no_weight_tying": true,
  "gpt_j_residual": false,
  "output_layer_parallelism": "column",
  "norm": "rmsnorm",
  "rms_norm_epsilon": 1.0e-5,

  "attention_config": [[["flash"], 32]],

  "scaled_upper_triang_masked_softmax_fusion": true,
  "bias_gelu_fusion": false,
  "use_bias_in_norms": false,
  "use_bias_in_attn_linear": false,
  "use_bias_in_mlp": false,
  "use_flashattn_swiglu": true,
  "activation": "swiglu",
  "intermediate_size": 14336,
  "mlp_multiple_of": 14336,



   "optimizer": {
     "type": "Adam",
     "params": {
       "lr": 5.0e-7,
       "betas": [0.9, 0.95],
       "eps": 1.0e-8
     }
   },
  "min_lr": 0.0,

   "zero_optimization": {
    "stage": 1,
    "allgather_partitions": true,
    "allgather_bucket_size": 1260000000,
    "overlap_comm": true,
    "reduce_scatter": true,
    "reduce_bucket_size": 1260000000,
    "contiguous_gradients": true,
    "cpu_offload": false
  },

  "train_impl": "dpo",
  "dataset_impl": "pairwise",
  "dpo_fp32": true,
  "dpo_beta": 0.01,
  "allow_chopped": false,
  "pos_train_data_paths": [ "data/pairwise/llama3_dpo_train_chosen_document" ],
  "pos_train_label_data_paths": [ "data/pairwise/llama3_dpo_train_chosen_label_document" ],
  "neg_train_data_paths": [ "data/pairwise/llama3_dpo_train_rejected_document" ],
  "neg_train_label_data_paths": [ "data/pairwise/llama3_dpo_train_rejected_label_document" ],
  "pos_valid_data_paths": [ "data/pairwise/llama3_dpo_val_chosen_document" ],
  "pos_valid_label_data_paths": [ "data/pairwise/llama3_dpo_val_chosen_label_document" ],
  "neg_valid_data_paths": [ "data/pairwise/llama3_dpo_val_rejected_document" ],
  "neg_valid_label_data_paths": [ "data/pairwise/llama3_dpo_val_rejected_label_document" ],
  "pos_test_data_paths": [ "data/pairwise/llama3_dpo_val_chosen_document" ],
  "pos_test_label_data_paths": [ "data/pairwise/llama3_dpo_val_chosen_label_document" ],
  "neg_test_data_paths": [ "data/pairwise/llama3_dpo_val_rejected_document" ],
  "neg_test_label_data_paths": [ "data/pairwise/llama3_dpo_val_rejected_label_document" ],

  "train_micro_batch_size_per_gpu": 32,
  "gradient_accumulation_steps": 2,
  "data_impl": "mmap",
  "pack_impl": "unpacked",
  "num_workers": 1,

  "checkpoint_activations": true,
  "checkpoint_num_layers": 1,
  "partition_activations": true,
  "synchronize_each_layer": true,

  "gradient_clipping": 1.0,
  "weight_decay": 0.1,
  "hidden_dropout": 0,
  "attention_dropout": 0,

  "precision": "bfloat16",
  "fp32_allreduce": true,
  "bf16": {
    "enabled": true
  },
  "data_types": {
    "grad_accum_dtype": "fp32"
  },

  "train_iters": 477,
  "lr_decay_iters": 477,
  "distributed_backend": "nccl",
  "lr_decay_style": "cosine",
  "warmup": 0.1,
  "checkpoint_factor": 1000,
  "eval_interval": 100,
  "eval_iters": 10,

  "log_interval": 1,
  "steps_per_print": 1,
  "wall_clock_breakdown": true,


  "save": "checkpoints/dpo/llama3/llama3-8b-instruct",
  #"load": "", # once run is started, to restart from intermediate ckpt use "load" = "save"
  "load": "checkpoints/neox_converted/llama3-8b-instruct",
  "vocab-file": "checkpoints/neox_converted/llama3-8b-instruct/tokenizer/tokenizer.json",
  "use_wandb": true,
  "wandb_group": "llama3-8b-instruct",
  "wandb_project": "ultrafeedback-dpo",
  "finetune": true, # set to false once resuming from intermediate finetuning step
  "tokenizer_type": "HFTokenizer",
}