
----------------------------------------------Summary-----------------------------------------------
batch_size_per_gpu: 1
max_batch_size_per_gpu: 3
gradient_accumulation_steps: 1
global_batch_size: 2240
dp_size: 2240
tp_size: 1
pp_size: 1
sp_size: 1
ds_zero: STAGE_3
total_num_gpus: 2240
seq_len: 2048
total_num_tokens: 300.0 G
num_params_total: 529.51 G
activation_recomputation: FULL
achived_flops: 48.0
flops_efficiency: 0.15384615384615385
hbm_memory_efficiency: 1
num_flops_total_per_micro_batch: 8819.83 T
weight_memory_per_gpu: 2.53 GB
gradient_memory_per_gpu: 471.86 MB
optimizer_state_memory_per_gpu: 3.77 GB
activation_memory_bs1: 8.81 GB
activation_memory_per_gpu: 8.81 GB
latency_per_micro_batch: 3.06 minutes
latency_fwd: 45.96 s
latency_fwd_attn: 15.78 s
latency_fwd_mlp: 30.06 s
latency_fwd_layernorm: 22.66 ms
latency_fwd_tp_comm: 0.0 us
latency_fwd_input_embedding: 1.32 ms
latency_fwd_output_embedding_loss: 87.86 ms
latency_per_iter: 3.06 minutes
total_training_latency: 139.07 days
gpu_hours: 7476572
----------------------------------------------------------------------------------------------------
