
----------------------------------------------Summary-----------------------------------------------
batch_size_per_gpu: 2
max_batch_size_per_gpu: 7
gradient_accumulation_steps: 1
global_batch_size: 1536
dp_size: 768
tp_size: 1
pp_size: 1
sp_size: 1
ds_zero: STAGE_3
total_num_gpus: 768
seq_len: 2048
total_num_tokens: 300.0 G
num_params_total: 174.56 G
activation_recomputation: FULL
achived_flops: 88.0
flops_efficiency: 0.28205128205128205
hbm_memory_efficiency: 1
num_flops_total_per_micro_batch: 5878.44 T
weight_memory_per_gpu: 1.69 GB
gradient_memory_per_gpu: 452.98 MB
optimizer_state_memory_per_gpu: 3.62 GB
activation_memory_bs1: 4.83 GB
activation_memory_per_gpu: 9.66 GB
latency_per_micro_batch: 1.11 minutes
latency_fwd: 16.73 s
latency_fwd_attn: 5.85 s
latency_fwd_mlp: 10.8 s
latency_fwd_layernorm: 24.86 ms
latency_fwd_tp_comm: 0.0 us
latency_fwd_input_embedding: 794.52 us
latency_fwd_output_embedding_loss: 57.51 ms
latency_per_iter: 1.11 minutes
total_training_latency: 73.73 days
gpu_hours: 1359052
----------------------------------------------------------------------------------------------------
