
----------------------------------------------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: 852.39 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: 107.74 ms
prefill_latency_fwd_attn: 36.24 ms
prefill_latency_fwd_mlp: 69.79 ms
prefill_latency_fwd_layernorm: 802.25 us
prefill_latency_fwd_tp_comm: 0.0 us
prefill_latency_fwd_input_embedding: 208.92 us
prefill_latency_fwd_output_embedding_loss: 699.05 us
decode_latency: 32.36 ms
decode_latency_fwd_attn: 10.43 ms
decode_latency_fwd_mlp: 20.86 ms
decode_latency_fwd_layernorm: 1.57 us
decode_latency_fwd_tp_comm: 0.0 us
decode_latency_fwd_input_embedding: 208.92 us
decode_latency_fwd_output_embedding_loss: 1.37 us
total_decode_latency: 1.04 s
total_latency: 1.14 s
total_per_token_latency: 35.72 ms
----------------------------------------------------------------------------------------------------
