
----------------------------------------------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: 142.61 MB
kv_cache_latency: 139.88 us
layernorm_dtype_bytes: 2
embedding_memory_per_gpu: 262.14 MB
weight_memory_per_gpu: 3.48 GB
prefill_max_batch_size_per_gpu: 43
prefill_activation_memory_per_gpu: 1.74 GB
prefill_num_flops_fwd_total: 6.87 T
decode_max_batch_size_per_gpu: 527
decode_activation_memory_per_gpu: 2.36 MB
decode_num_flops_fwd_total: 13.15 G
prefill_latency: 11.4 ms
prefill_latency_fwd_attn: 3.74 ms
prefill_latency_fwd_mlp: 7.05 ms
prefill_latency_fwd_layernorm: 263.3 us
prefill_latency_fwd_tp_comm: 0.0 us
prefill_latency_fwd_input_embedding: 128.56 us
prefill_latency_fwd_output_embedding_loss: 215.09 us
decode_latency: 1.85 ms
decode_latency_fwd_attn: 526.92 us
decode_latency_fwd_mlp: 1.05 ms
decode_latency_fwd_layernorm: 0.51 us
decode_latency_fwd_tp_comm: 0.0 us
decode_latency_fwd_input_embedding: 128.56 us
decode_latency_fwd_output_embedding_loss: 0.42 us
total_decode_latency: 59.2 ms
total_latency: 70.6 ms
total_per_token_latency: 2.21 ms
----------------------------------------------------------------------------------------------------
