
----------------------------------------------Summary-----------------------------------------------
batch_size_per_gpu: 1
seq_len: 512
tp_size: 1
pp_size: 1
num_tokens_to_generate: 32
use_kv_cache: True
kv_cache_memory_per_gpu: 869.01 MB
kv_cache_latency: 947.1 us
layernorm_dtype_bytes: 2
embedding_memory_per_gpu: 425.98 MB
weight_memory_per_gpu: 64.22 GB
prefill_max_batch_size_per_gpu: 1
prefill_activation_memory_per_gpu: 9.81 GB
prefill_num_flops_fwd_total: 33.3 T
decode_max_batch_size_per_gpu: 17
decode_activation_memory_per_gpu: 12.79 MB
decode_num_flops_fwd_total: 64.22 G
prefill_latency: 153.6 ms
prefill_latency_fwd_attn: 51.77 ms
prefill_latency_fwd_mlp: 99.7 ms
prefill_latency_fwd_layernorm: 891.38 us
prefill_latency_fwd_tp_comm: 0.0 us
prefill_latency_fwd_input_embedding: 232.13 us
prefill_latency_fwd_output_embedding_loss: 998.64 us
decode_latency: 35.95 ms
decode_latency_fwd_attn: 11.59 ms
decode_latency_fwd_mlp: 23.18 ms
decode_latency_fwd_layernorm: 1.74 us
decode_latency_fwd_tp_comm: 0.0 us
decode_latency_fwd_input_embedding: 232.13 us
decode_latency_fwd_output_embedding_loss: 1.95 us
total_decode_latency: 1.15 s
total_latency: 1.3 s
total_per_token_latency: 40.75 ms
----------------------------------------------------------------------------------------------------
