configs/llemma/34B.yml

{
  "pipe_parallel_size": 0,
  "model_parallel_size": 8,
  "make_vocab_size_divisible_by": 1,

  # model settings
  "num_layers": 48,
  "hidden_size": 8192,
  "num_attention_heads": 64,
  "attention_type": "groupedquery",
  "num_kv_heads": 8,
  # NB: These rotary embedding and sequence length parameters
  # May differ from CodeLlama configs. They match what we used for
  # Llemma continued pretraining. See https://arxiv.org/abs/2310.10631
  # For detailed discussion
  "seq_length": 4096,
  "max_position_embeddings": 4096,
  "pos_emb": "rotary",
  "rotary_pct": 1,
  "rotary_emb_base": 1000000,
  "no_weight_tying": true,
  "gpt_j_residual": false,
  "output_layer_parallelism": "column",
  "norm": "rmsnorm",
  "rms_norm_epsilon": 1.0e-5,

  "attention_config": [[["flash"], 48]],

  "scaled_upper_triang_masked_softmax_fusion": true,
  "bias_gelu_fusion": false,
  "use_bias_in_norms": false,
  "use_bias_in_attn_linear": false,
  "activation": "swiglu",
  "mlp_multiple_of": 256,

   "optimizer": {
     "type": "Adam",
     "params": {
       "lr": 0.00005,
       "betas": [0.9, 0.95],
       "eps": 1.0e-8
     }
   },

   "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
  },

  # trained on 256 gpus
  "train_micro_batch_size_per_gpu": 2,
  "gradient_accumulation_steps": 16,
  "data_impl": "mmap",

  "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": 12000,
  "lr_decay_iters": 12000,
  "distributed_backend": "nccl",
  "lr_decay_style": "cosine",
  "min_lr": 1.65e-6,
  "warmup": 0.042, # warmup for ~500 iters
  "checkpoint_factor": 250,
  "eval_interval": 250,
  "eval_iters": 25,

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

  "tokenizer_type": "SPMTokenizer",
  #"vocab-file": # use 'tokenizer.model' from Meta CodeLlama download

  # "load": "" # set to same as "save" to resume from intermediate finetuning step
  #"load": MP=8 CodeLlama-34B checkpoint, converted from Meta CodeLlama download.
  # When resuming from mid-finetuning run, change "load" to the same as save location.
  "finetune": true, # set to false once resuming from intermediate finetuning step
}