train_llama2_7b_1nodes.sh

#!/bin/bash

# 使用blaslt和rocblas
export export LD_LIBRARY_PATH=/public/home/wangxj/Downloads/blas/hipblaslt-install0519/lib:$LD_LIBRARY_PATH 
export LD_LIBRARY_PATH=/public/home/wangxj/Downloads/blas/rocblas-install-0513-release/lib:$LD_LIBRARY_PATH

INITIALIZATION_ARGS=( --num-workers 2)

for para in $*
do
    if [[ $para == --data_path* ]];then
        data_path=${para#*=}
    elif [[ $para == --tokenizer_path* ]];then
        tokenizer_path=${para#*=}
    elif [[ $para == --checkpoint_path* ]];then
        checkpoint_path=${para#*=}
    elif [[ $para == --profiling* ]];then
        profiling=${para#*=}
    elif [[ $para == --reproduce* ]];then
        INITIALIZATION_ARGS=( --reproduce --num-workers 0)
        export MIOPEN_DEBUG_CONVOLUTION_DETERMINISTIC=1  # miopen 确定算法打开
        export ROCBLAS_ATOMICS_MOD=0                     # rocblas 关闭原子操作
        # 关闭miopen中的atomic操作算法, 只保留gemm算法
        export MIOPEN_DEBUG_CONV_FFT=0
        export MIOPEN_DEBUG_CONV_DIRECT=0
        export MIOPEN_DEBUG_CONV_GEMM=1
        export MIOPEN_DEBUG_CONV_WINOGRAD=0
        export MIOPEN_DEBUG_CONV_IMPLICIT_GEMM=0
    fi
done

# data path
DATA_PATH=${data_path}
TOKENIZER_MODEL_PATH=${tokenizer_path}
CHECKPOINT_PATH=${checkpoint_path}

# 运行环境参数
DIST_URL=${1}
DIST_PORT=${2}
RANK=$OMPI_COMM_WORLD_RANK
LOCAL_RANK=$OMPI_COMM_WORLD_LOCAL_RANK
WORLD_SIZE=$OMPI_COMM_WORLD_SIZE
CURRENT_DIR="$( cd "$( dirname "$0" )" && pwd )"
MEGATRON_PATH=$( dirname $( dirname ${CURRENT_DIR}))
export PYTHONPATH=${MEGATRON_PATH}/Megatron-LM:$PYTHONPATH

# default env
export GLOG_minloglevel=3
export CUDA_DEVICE_MAX_CONNECTIONS=1
export HSA_FORCE_FINE_GRAIN_PCIE=1
export OMP_NUM_THREADS=1
export GPU_MAX_HW_QUEUES=10 # 4 # 20

# tp-overlap控制参数
export NVTE_DISABLE_FC2_DGRAD_OVERLAP=1
export NVTE_NO_PIPELINE_OVERLAP=1

# torch控制多流转单流
export ALLREDUCE_STREAM_WITH_COMPUTE=1
export SENDRECV_STREAM_WITH_COMPUTE=1 

#增加编译缓存
export cache_size_limit=64

DISTRIBUTED_ARGS=(
    --rank ${RANK}
    --world-size ${WORLD_SIZE}
    --local-rank ${LOCAL_RANK}
    --dist-url tcp://${DIST_URL}:${DIST_PORT}
)

GPT_MODEL_ARGS=(
    --seq-length 4096
    --num-layers 32
    --hidden-size 4096
    --ffn-hidden-size 11008 
    --num-attention-heads 32
    --max-position-embeddings 4096
    --normalization RMSNorm # Lightop
    --position-embedding-type rope
    --untie-embeddings-and-output-weights
)

TRAINING_ARGS=(
    --transformer-impl transformer_engine
    --use-mcore-models 
    --micro-batch-size 1
    --global-batch-size 256
    --train-iters 50
    --weight-decay 0.1 
    --adam-beta1 0.9 
    --adam-beta2 0.95 
    --init-method-std 0.006 
    --clip-grad 1.0 
    --bf16
    --disable-bias-linear
    --attention-dropout 0
    --hidden-dropout 0
    --swiglu
    --lr 3.0e-5 
    --lr-decay-style cosine 
    --min-lr 3.0e-6
    --lr-warmup-iters 1
    --ckpt-format torch
    --ddp-average-in-collective
    --overlap-grad-reduce
    # --tp-comm-overlap
    # --tp-comm-overlap-rs-dgrad
    --use-flash-attn
)

MODEL_PARALLEL_ARGS=(
    --tensor-model-parallel-size 1
    --pipeline-model-parallel-size 2
    --context-parallel-size 1
    --use-distributed-optimizer 
    --sequence-parallel
)

DATA_ARGS=(
    --tokenizer-type Llama2Tokenizer
    --tokenizer-model ${TOKENIZER_MODEL_PATH}
    --data-path ${DATA_PATH} 
    --split 949,50,1
)

EVAL_AND_LOGGING_ARGS=(
    --log-throughput
    --eval-iters 5
    --log-interval 1
    --save-interval 1000 
    --eval-interval 1000 
    --save $CHECKPOINT_PATH
    --load $CHECKPOINT_PATH
    --tensorboard-dir "${CHECKPOINT_PATH}/tensorboard" 
)

TORCH_PROFIE_ARGS=(
    --profile
    --profile-ranks 0 1 2 3 4 5 6 7
    --profile-step-start 3
    --profile-step-end 4
    --profile-dir torch_prof_llama_1nodes_tp1-pp2-cp1
    --use-pytorch-profiler
)

HIP_PROFIE_ARGS=(
    --profile
    --profile-ranks 0 1 2 3 4 5 6 7
    --profile-step-start 4
    --profile-step-end 5
    --use-hip-profiler
)

APP="python -u ${MEGATRON_PATH}/pretrain_gpt.py \
    ${GPT_MODEL_ARGS[@]} \
    ${TRAINING_ARGS[@]} \
    ${MODEL_PARALLEL_ARGS[@]} \
    ${DATA_ARGS[@]} \
    ${EVAL_AND_LOGGING_ARGS[@]} \
    ${DISTRIBUTED_ARGS[@]} \
    ${INITIALIZATION_ARGS[@]} \
    "

if [[ $profiling == "torch" ]]; then
    APP+=" ${TORCH_PROFIE_ARGS[@]}"
elif [[ $profiling == "hip" ]]; then
    mkdir -p hip_prof_data
    APP+=" ${HIP_PROFIE_ARGS[@]}"
    APP="hipprof -d hip_prof_data --hip-trace --trace-off ${APP}"
fi

#for hygon cpu
${MEGATRON_PATH}/requirements/launch_with_binding.sh ${LOCAL_RANK} ${APP}