单机单卡仿真结果跟真实运行结果相差一个数量级

在包含1个英伟达V100S GPU的训练megatron-lm 126m的模型，实测的迭代时间为680ms-740ms之间，但运行SimAI模拟出来的迭代时间却是6042ms。

模型训练命令如下：
torchrun --nproc_per_node=1 pretrain_gpt.py --tensor-model-parallel-size 1 --pipeline-model-parallel-size 1 --num-layers 12 --hidden-size 768 --num-attention-heads 8 --seq-length 2048 --max-position-embeddings 2048 --micro-batch-size 8 --train-iters 5000000 --lr 6e-4 --data-path ../wiki_processed/my-126m_text_document --vocab-file gpt2-vocab.json --merge-file gpt2-merges.txt --split 950,40,10 --distributed-backend nccl --fp16 --recompute-activations --log-interval 10 --save-interval 2000 --save checkpoints

SIMAI模拟器运行过程如下：

1） 修改megatron_workload_with_aiob.sh，152行增加126m模型的配置，生成workload文件：
model_name=gpt_126M
num_layers=12
hidden_size=768
num_attention_heads=16
tensor_model_parallel_size=1
;;
2） 按照文档直接生成定制化workload
python -m workload_generator.AIOB_simAI_workload_generator --world_size=1 --global_batch=8 --micro_batch=1 --num_layers=12 --num_attention_heads=16 --hidden_size=768 --tensor_model_parallel_size=1 --seq_length=2048 --swiglu --ffn_hidden_size=3072 --aiob_enable
https://github.com/aliyun/aicb/tree/d9b4f5cd7d9d34a80cfbb0389831a16c7fe3ed7b

3）之后基于步骤1或步骤2生成的workload文件运行模拟器，结果跟预期相差一个数量级。
./bin/SimAI_analytical -w aicb/results/workload/None-None-world_size1-tp1-pp1-ep1-gbs8-mbs1-seq2048-MOE-False-GEMM-False-flash_attn-False.txt -g 1 -g_p_s 1 -r test126m_acib- -busbw example/busbw.yaml

Start with megatron_workload_with_aiob.sh around line 152, then inspect the workload produced by AIOB_simAI_workload_generator and the SimAI_analytical invocation. Compare the reported 8 versus 16 attention heads and other model options with the Megatron command, reproduce the 6042 ms result, and identify why it differs from the measured 680–740 ms iteration time.