tests/config/test_setup.yml

# 19M parameter model, & local setup with some additional simplifications
{
  # Settings to make the test setup as lightweight as possible
  "data_path": "data/enwik8/enwik8_text_document",
  "vocab_file": "data/gpt2-vocab.json",
  "merge_file": "data/gpt2-merges.txt",
  "lr_decay_iters": 20,
  "train_iters": 20,
  "hostfile": "None",
  "include": "localhost:1",
  "use_wandb": False,

  # Settings copied from 19M parameter config (some modifications above, meaning we can't use configs/19M.yml directly)
  "pipe_parallel_size": 1,
  "model_parallel_size": 1,

  # model settings
  "num_layers": 2,
  "hidden_size": 8,
  "num_attention_heads": 4,
  "seq_length": 1024,
  "max_position_embeddings": 1024,
  "pos_emb": "rotary",
  "no_weight_tying": true,
  "gpt_j_residual": false,
  "output_layer_parallelism": "column",

  "scaled_upper_triang_masked_softmax_fusion": false,
  "bias_gelu_fusion": false,
  "rope_fusion": false,
  "layernorm_fusion": false,

  # Optimizer
  "optimizer": {
    "type": "sm3",
    "params": {},
  },

  # precision
  "precision": "fp16",

  # init methods
  "init_method": "small_init",
  "output_layer_init_method": "wang_init",

  "train_micro_batch_size_per_gpu": 4,
  "gradient_accumulation_steps": 1,
  "data_impl": "mmap",
  "num_workers": 1,

  # activation checkpointing
  "checkpoint_activations": true,
  "checkpoint_num_layers": 1,
  "partition_activations": true,
  "synchronize_each_layer": true,

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

  "distributed_backend": "nccl",
  "lr_decay_style": "cosine",
  "warmup": 0.01,
  "checkpoint_factor": 1000,
  "eval_interval": 100000,
  "eval_iters": 10,

  "log_interval": 10,
  "steps_per_print": 10,
  "wall_clock_breakdown": true,

  # additional deepspeed args not specified above
  "deepspeed_extra_args": {
    "comms_logger": {
        "enabled": true,
        "verbose": true,
        "prof_all": true,
        "debug": false
    },
  }
}