
----------------------------------------------Summary-----------------------------------------------
batch_size_per_gpu: 4
max_batch_size_per_gpu: 6
gradient_accumulation_steps: 1
global_batch_size: 1536
dp_size: 384
tp_size: 1
pp_size: 1
sp_size: 1
ds_zero: STAGE_3
total_num_gpus: 384
seq_len: 2048
total_num_tokens: 300.0 G
num_params_total: 174.56 G
activation_recomputation: FULL
achived_flops: 144.0
flops_efficiency: 0.46153846153846156
hbm_memory_efficiency: 1
num_flops_total_per_micro_batch: 11756.88 T
weight_memory_per_gpu: 2.14 GB
gradient_memory_per_gpu: 905.97 MB
optimizer_state_memory_per_gpu: 7.25 GB
activation_memory_bs1: 4.83 GB
activation_memory_per_gpu: 19.33 GB
latency_per_micro_batch: 1.36 minutes
latency_fwd: 20.46 s
latency_fwd_attn: 7.15 s
latency_fwd_mlp: 13.19 s
latency_fwd_layernorm: 49.72 ms
latency_fwd_tp_comm: 0.0 us
latency_fwd_input_embedding: 794.52 us
latency_fwd_output_embedding_loss: 70.29 ms
latency_per_iter: 1.36 minutes
total_training_latency: 90.12 days
gpu_hours: 830532
----------------------------------------------------------------------------------------------------
