configs/src_config.yml

# Hubble 1.1B - Copied from TinyLlama https://github.com/Lightning-AI/litgpt/blob/a5021be4bb48e27779586b56b062a1749ecb232f/litgpt/config.py#L1809 
# Modified from https://github.com/aflah02/gpt-neox/blob/olmo-support/configs/hubble/Speed_Exps/1_1B_Baseline_BS_8_GAS_8_No_Activation_Checkpointing_GQA_Llama_3_2_Fusions_All_4_FA_Swiglu.yml
{
  # parallelism settings ( you will want to change these based on your cluster setup, ideally scheduling pipeline stages
  # across the node boundaries )
  "pipe_parallel_size": 1,
  "model_parallel_size": 1,
  "make_vocab_size_divisible_by": 128, # Need to set as 64 because code makes it divisible by MP*makes_vocab_size_divisible_by

  # model settings
  "num_layers": 16,
  "hidden_size": 2048,
  "num_attention_heads": 32,
  "num_kv_heads": 8,
  "intermediate_size": 24576, # 8192*3
  "seq_length": 2048,
  "max_position_embeddings": 2048,
  "pos_emb": "rotary",
  "no_weight_tying": true,
  "gpt_j_residual": false,
  "output_layer_parallelism": "column",
  "norm": "rmsnorm",
  "rms_norm_epsilon": 1.0e-5,
  "rmsnorm_fusion": true,
    
  # these should provide some speedup but takes a while to build, set to true if desired
  "scaled_upper_triang_masked_softmax_fusion": true,
  "scaled_masked_softmax_fusion": true,
  "bias_gelu_fusion": false,
  "rope_fusion": true,
  # "layernorm_fusion": false,
  "use_bias_in_norms": false,
  "use_bias_in_attn_linear": false,
  "activation": "swiglu",
  "use_flashattn_swiglu": true,
  "mlp_multiple_of": 256,

  # init methods (Copied from OLMo 2)
  "init_method": "normal",
  "output_layer_init_method": "normal",
  "init_method_std": 0.02,

  # optimizer settings
  "optimizer": {
    "type": "Adam",
    "params": {
      "lr": 0.0004,
      "betas": [0.9, 0.95],
      "eps": 1.0e-8,
    }
  },

  # for all zero_optimization options, see https://www.deepspeed.ai/docs/config-json/#zero-optimizations-for-fp16-training
  "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
  },
  "min_lr": 0.00004,

  # batch / data settings
  # "n_gpus": 32,
  "train_micro_batch_size_per_gpu": 16,
  "gradient_accumulation_steps": 2,
  "train_batch_size": 1024,
  "data_impl": "mmap",

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

  # regularization
  "gradient_clipping": 1.0,
  "weight_decay": 0.1,
  "hidden_dropout": 0,
  "attention_dropout": 0,
  # Flash Attention
  "attention_config": [[["flash"], 16]],

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

  # misc. training settings
  "train_iters": 48000,
  "lr_decay_iters": 48000,
  "distributed_backend": "nccl",
  "lr_decay_style": "cosine",
  "warmup": 0.05,
  "checkpoint_factor": 1000,
  "eval_interval": 2000,
  "eval_iters": 10,
  "extra_save_iters": [0, 1, 4, 16, 64, 256, 512],
  # "keep_last_n_checkpoints": 2,

  # logging
  "log_interval": 10,
  "steps_per_print": 50,
  "wall_clock_breakdown": true,

  # "memory_profiling": true,
  # "memory_profiling_path": "/NS/llm-pretraining/work/afkhan/USC_Colab/gpt-neox/Artifacts/Profiles/Hubble-1.1B-Baseline_BS_8_GAS_8_No_Activation_Checkpointing_GQA_Llama_3_2_Fusions_All_4_FA_Swiglu",
  # "profile_step_start": 0,
  # "profile_step_stop": 100,

  # "launcher": "slurm",
  # "deepspeed_mpi": true,
  # "deepspeed_slurm": true,
  "no_ssh_check": true,
}