
----------------------------------------------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: 285.21 MB
kv_cache_latency: 310.84 us
layernorm_dtype_bytes: 2
embedding_memory_per_gpu: 262.14 MB
weight_memory_per_gpu: 13.15 GB
prefill_max_batch_size_per_gpu: 20
prefill_activation_memory_per_gpu: 3.22 GB
prefill_num_flops_fwd_total: 6.87 T
decode_max_batch_size_per_gpu: 230
decode_activation_memory_per_gpu: 4.2 MB
decode_num_flops_fwd_total: 13.15 G
prefill_latency: 31.89 ms
prefill_latency_fwd_attn: 10.7 ms
prefill_latency_fwd_mlp: 20.14 ms
prefill_latency_fwd_layernorm: 292.56 us
prefill_latency_fwd_tp_comm: 0.0 us
prefill_latency_fwd_input_embedding: 142.85 us
prefill_latency_fwd_output_embedding_loss: 614.55 us
decode_latency: 7.48 ms
decode_latency_fwd_attn: 2.34 ms
decode_latency_fwd_mlp: 4.68 ms
decode_latency_fwd_layernorm: 0.57 us
decode_latency_fwd_tp_comm: 0.0 us
decode_latency_fwd_input_embedding: 142.85 us
decode_latency_fwd_output_embedding_loss: 1.2 us
total_decode_latency: 239.32 ms
total_latency: 271.21 ms
total_per_token_latency: 8.48 ms
----------------------------------------------------------------------------------------------------
