
----------------------------------------------Summary-----------------------------------------------
batch_size_per_gpu: 10
max_batch_size_per_gpu: 18
gradient_accumulation_steps: 12500
global_batch_size: 4000000
dp_size: 32
tp_size: 8
pp_size: 8
sp_size: 8
ds_zero: NONE
total_num_gpus: 2048
seq_len: 2048
total_num_tokens: 1.4 T
num_params_total: 32.11 G
activation_recomputation: SELECTIVE
achived_flops: 312
flops_efficiency: 1
hbm_memory_efficiency: 1
num_flops_total_per_micro_batch: 4213.77 T
weight_memory_per_gpu: 983.6 MB
gradient_memory_per_gpu: 930.35 MB
optimizer_state_memory_per_gpu: 7.44 GB
activation_memory_bs1: 3.78 GB
activation_memory_per_gpu: 37.83 GB
latency_per_micro_batch: 211.03 ms
latency_fwd: 92.23 ms
latency_fwd_attn: 23.49 ms
latency_fwd_mlp: 40.71 ms
latency_fwd_layernorm: 467.98 us
latency_fwd_tp_comm: 22.26 ms
latency_fwd_input_embedding: 1.8 ms
latency_fwd_output_embedding_loss: 3.5 ms
latency_per_iter: 43.96 minutes
total_training_latency: 5.19 days
gpu_hours: 255106
----------------------------------------------------------------------------------------------------
