
----------------------------------------------Summary-----------------------------------------------
batch_size_per_gpu: 2
max_batch_size_per_gpu: 3
gradient_accumulation_steps: 1
global_batch_size: 2240
dp_size: 1120
tp_size: 1
pp_size: 1
sp_size: 1
ds_zero: STAGE_3
total_num_gpus: 1120
seq_len: 2048
total_num_tokens: 300.0 G
num_params_total: 529.51 G
activation_recomputation: FULL
achived_flops: 98.0
flops_efficiency: 0.3141025641025641
hbm_memory_efficiency: 1
num_flops_total_per_micro_batch: 17639.67 T
weight_memory_per_gpu: 3.0 GB
gradient_memory_per_gpu: 943.72 MB
optimizer_state_memory_per_gpu: 7.55 GB
activation_memory_bs1: 8.81 GB
activation_memory_per_gpu: 17.62 GB
latency_per_micro_batch: 3.0 minutes
latency_fwd: 45.05 s
latency_fwd_attn: 15.46 s
latency_fwd_mlp: 29.45 s
latency_fwd_layernorm: 45.31 ms
latency_fwd_tp_comm: 0.0 us
latency_fwd_input_embedding: 1.32 ms
latency_fwd_output_embedding_loss: 86.06 ms
latency_per_iter: 3.0 minutes
total_training_latency: 136.23 days
gpu_hours: 3661994
----------------------------------------------------------------------------------------------------
