setup_env_bk.sh

#!/bin/bash

export CPU_NAME="TurinC"
export MODEL_DIR=/home/amd/data/models
export DATA_DIR=/home/amd/data/criteo_multi_hot
export RESULTS_ROOT_DIR=/home/amd/workspace/dlrm_results
export CONDA_PREFIX=/home/amd/miniforge3/envs/DLRMv2_env
export USE_IPEX_RUNTIME=0

# for the dlrm_int8 model, the forward function does not take offset as input. NEED_OFFSET_INPUT should be set to 0
# all other models should set NEED_OFFSET_INPUT to 1
export NEED_OFFSET_INPUT=1


# export KMP_BLOCKTIME=1
# # export OMP_NUM_THREADS=$CPUS_PER_INSTANCE
# export KMP_AFFINITY="granularity=fine,compact,1,0"
# export DNNL_PRIMITIVE_CACHE_CAPACITY=20971520
export LD_PRELOAD="${CONDA_PREFIX}/lib/libtcmalloc.so:$LD_PRELOAD"
export LD_PRELOAD="${CONDA_PREFIX}/lib/libgomp.so:$LD_PRELOAD"
# export TCMALLOC_LARGE_ALLOC_REPORT_THRESHOLD=30469645312
# export ZENDNN_MATMUL_ALGO=FP32:3,BF16:0