Skip to content

GitLab

Commit 1e8185f4 authored by wangxj's avatar wangxj
Browse files

添加llama和qwen的脚本

parent 3e2c4b1e
Hide whitespace changes
Inline Side-by-side
Showing with 1174 additions and 510 deletions
examples/llama/Llama2_70b.sh deleted 100755 → 0
View file @ 3e2c4b1e
#!/bin/bash
# set -eux
#export FLASH_ATTENTION_PRINT_PARAM=1
# Runs the "7B" parameter model
export HSA_FORCE_FINE_GRAIN_PCIE=1
export OMP_NUM_THREADS=1
export NCCL_P2P_LEVEL=PXB # SYS
#export HIP_ALLOC_INITIALIZE=0
# export GPU_MAX_HW_QUEUES=10
export NCCL_ALGO=Ring
export NCCL_NCHANNELS_PER_PEER=16
export NCCL_MIN_NCHANNELS=32 # 20
export NCCL_MAX_NCHANNELS=32 # 20
export NCCL_IB_TIMEOUT=22
export CUDA_DEVICE_MAX_CONNECTIONS=1
export NCCL_IB_HCA=mlx5_2:1,mlx5_3:1,mlx5_4:1,mlx5_5:1,mlx5_6:1,mlx5_7:1,mlx5_8:1,mlx5_9:1
export NCCL_NET_GDR_LEVEL=7
export NCCL_NET_GDR_READ=1
export RCCL_SDMA_COPY_ENABLE=0
export NCCL_TOPO_FILE="/public/home/wangxj/Projects/rccl-test/rccl-tests-0204/topo-input.xml"
# export NCCL_TOPO_FILE="/workspace/rccl-test/rccl-tests-0204/topo-input.xml"
export GLOG_minloglevel=3 # 打印error级别的nccl日志
source /opt/dtk/env.sh
# 导入hipblaslt库
# export LD_LIBRARY_PATH=/data/hipblaslt-install-0904/lib:$LD_LIBRARY_PATH
# export LD_LIBRARY_PATH=/data/hipblaslt-install-dtk-25.04-0212/lib:$LD_LIBRARY_PATH
export LD_LIBRARY_PATH=/public/home/wangxj/Downloads/blas/hipblaslt-install-dtk-25.04-0212/lib:$LD_LIBRARY_PATH
# 更新rocblas
# export LD_LIBRARY_PATH=/data/rocblas-install_qwen1211/lib:$LD_LIBRARY_PATH
# export LD_LIBRARY_PATH=/data/rocblas-install_qwen1228/lib:$LD_LIBRARY_PATH
# export LD_LIBRARY_PATH=/data/rocblas-install-0118-bf16/lib:$LD_LIBRARY_PATH
# export LD_LIBRARY_PATH=/data/rocblas-install-0203-release/lib:$LD_LIBRARY_PATH
export LD_LIBRARY_PATH=/public/home/wangxj/Downloads/blas/rocblas-install-0331-release/lib:$LD_LIBRARY_PATH
# torch控制多流转单流
export ALLREDUCE_STREAM_WITH_COMPUTE=1
export SENDRECV_STREAM_WITH_COMPUTE=1
# prof采集添加同步, 避免卡顿
# export GPU_FLUSH_ON_EXECUTION=1
# export HIP_DIRECT_DISPATCH=0
# 采集rocblas size
# export ROCBLAS_LAYER=3
# export HIPBLASLT_LOG_LEVEL=3
# 采集 fa size
# export FLASH_ATTENTION_PRINT_PARAM=1
#增加编译缓存
export cache_size_limit=64
# lightop算子库
export PYTORCH_ROCM_ARCH='gfx906;gfx926;gfx936'
# CHECKPOINT_PATH=./Llama-2-7b-hf-to-meg-tp1-pp2 #CHECKPOINT_PATH=./tmp_7b #
SAVE_PATH=./tmp_7b
TENSORBOARD_LOGS_PATH=./tmp_7b #$2 #<Specify path>
DATA_PATH="/public/home/gmhtest_tmp/RedPajama-Data-1T-Sample/redpajama_text_document" #<Specify path and file prefix>_text_document
# DATA_PATH="/data/datasets/oscar-1GB-head/oscar-1GB_head-llama3.2_text_document" #<Specify path and file prefix>_text_document
GPT_MODEL_ARGS=(
--num-layers 80 #80 #80 #40 # 20 #
--hidden-size 8192
--ffn-hidden-size 22016 # 28672
--num-attention-heads 64
--max-position-embeddings 8192
--group-query-attention
--num-query-groups 8
--normalization RMSNorm
--position-embedding-type rope
--untie-embeddings-and-output-weights # 分开处理embed和输出权重, 增加灵活性
)
export NVTE_FLASH_ATTN=1 # 走cutlass
# export NVTE_FLASH_ATTN_TRITON=1 # 走triton_fa
# --transformer-impl transformer_engine # 走core用这两组参数
# --use-mcore-models
# --transformer-impl local # 走legacy用这两组参数
# --use-legacy-models
TRAINING_ARGS=(
--transformer-impl local # 走legacy用这两组参数
--use-legacy-models
--micro-batch-size 1
--global-batch-size 512 #32 #512 #256 # 64 #240 #60 #512 #64
--train-iters 300
--weight-decay 0.1
--adam-beta1 0.9
--adam-beta2 0.95
--init-method-std 0.006
--clip-grad 1.0
--bf16
# --fp16 # 开启fp16需要指定loss-scale
# --loss-scale 1024
--use-distributed-optimizer
--disable-bias-linear
--attention-dropout 0
--hidden-dropout 0
# --no-gradient-accumulation-fusion
# --no-check-for-nan-in-loss-and-grad
--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 # 在dp阶段通信中, 梯度或参数将被直接平均, 而不是先求和(到一个设备)再平均
# --recompute-activations
# --recompute-granularity full # 开启重计算降低显存增加耗时
# --recompute-num-layers 1 #0 #
# --recompute-method block
--overlap-grad-reduce # 重叠ddp grad reduce
# --tp-comm-overlap # tensor parallel comm和gemm重叠, 启动core
# --tp-comm-overlap-rs-dgrad # reduce-scatter和dgrad gemm重叠, 启动core
--use-flash-attn
)
# export TORCHINDUCTOR_COORDINATE_DESCENT_TUNING=1
# export TORCHINDUCTOR_BENCHMARK_FUSION=1
# export TORCHINDUCTOR_BENCHMARK_MULTI_TEMPLATES=1
# export TORCHINDUCTOR_MAX_AUTOTUNE=1
# export TORCHINDUCTOR_CACHE_DIR=./cache
# --use-flash-attn-cutlass # cutlass fa
# --use-flash-attn-triton # triton fa
# --use-flash-attn-torch # torch fa
MODEL_PARALLEL_ARGS=(
--sequence-parallel
--tensor-model-parallel-size 4
--pipeline-model-parallel-size 8
--context-parallel-size 1
# --num-layers-per-virtual-pipeline-stage 1
# --microbatch-group-size-per-virtual-pipeline-stage 5
# --no-overlap-p2p-communication # 开启后
)
DATA_ARGS=(
--data-path $DATA_PATH
--seq-length 4096 #8192 #4096
--split 949,50,1
--tokenizer-type Llama2Tokenizer
--tokenizer-model /public/home/gmhtest_tmp/RedPajama-Data-1T-Sample/tokenizer.model
# --tokenizer-model /data/model_weights/llama3.2/tokenizer.model
)
EVAL_AND_LOGGING_ARGS=(
--log-interval 1
--log-throughput
--save-interval 500
--eval-interval 50
--eval-iters 3
--save $SAVE_PATH
--load $SAVE_PATH
--tensorboard-dir $TENSORBOARD_LOGS_PATH
)
# FINETUNE_ARGS=(
# # --finetune
# # --pretrained-checkpoint $CHECKPOINT_PATH
# --load $CHECKPOINT_PATH
# --no-load-optim
# --no-load-rng
# )
PROFILE_ARGS=(
--profile
--profile-step-start 4
--profile-step-end 5
--use-pytorch-profiler
--profile-ranks 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
--profile-dir prof_data
)
RANK=$OMPI_COMM_WORLD_RANK
LOCAL_RANK=$OMPI_COMM_WORLD_LOCAL_RANK
WORLD_SIZE=$OMPI_COMM_WORLD_SIZE
DIST_URL=${1}
DIST_PORT=34577
DISTRIBUTED_ARGS=(
--rank ${RANK}
--world-size ${WORLD_SIZE}
--local-rank ${LOCAL_RANK}
--dist-url tcp://${DIST_URL}:${DIST_PORT}
)
APP="python -u ../../pretrain_gpt.py \
${GPT_MODEL_ARGS[@]} \
${TRAINING_ARGS[@]} \
${MODEL_PARALLEL_ARGS[@]} \
${DATA_ARGS[@]} \
${EVAL_AND_LOGGING_ARGS[@]} \
${DISTRIBUTED_ARGS[@]} \
"
# 开启profile
# ${PROFILE_ARGS[@]} \
# export HIP_VISIBLE_DEVICES=0,7 # # 4,5,6,7 #,
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 # # 4,5,6,7 #,
# export CUDA_VISIBLE_DEVICES=4,5,6,7 # 0,1,2,3,
# ${APP}
case ${LOCAL_RANK} in
[0])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
# hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
numactl --cpunodebind=0 --membind=0 ${APP}
;;
[1])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
# hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
numactl --cpunodebind=1 --membind=1 ${APP}
;;
[2])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
# hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
numactl --cpunodebind=2 --membind=2 ${APP}
;;
[3])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
numactl --cpunodebind=3 --membind=3 ${APP}
# hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
;;
[4])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
numactl --cpunodebind=4 --membind=4 ${APP}
# hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
;;
[5])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
numactl --cpunodebind=5 --membind=5 ${APP}
# hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
;;
[6])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
numactl --cpunodebind=6 --membind=6 ${APP}
# hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
;;
[7])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
numactl --cpunodebind=7 --membind=7 ${APP}
# hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
;;
esac
\ No newline at end of file
examples/llama/Llama3_70b.sh deleted 100644 → 0
View file @ 3e2c4b1e
#!/bin/bash
# set -eux
#export FLASH_ATTENTION_PRINT_PARAM=1
# Runs the "7B" parameter model
export HSA_FORCE_FINE_GRAIN_PCIE=1
export OMP_NUM_THREADS=1
export NCCL_P2P_LEVEL=PXB # SYS
#export HIP_ALLOC_INITIALIZE=0
# export GPU_MAX_HW_QUEUES=10
export NCCL_ALGO=Ring
export NCCL_NCHANNELS_PER_PEER=16
export NCCL_MIN_NCHANNELS=32 # 20
export NCCL_MAX_NCHANNELS=32 # 20
export NCCL_IB_TIMEOUT=22
export CUDA_DEVICE_MAX_CONNECTIONS=1
export NCCL_IB_HCA=mlx5_2:1,mlx5_3:1,mlx5_4:1,mlx5_5:1,mlx5_6:1,mlx5_7:1,mlx5_8:1,mlx5_9:1
export NCCL_NET_GDR_LEVEL=7
export NCCL_NET_GDR_READ=1
export RCCL_SDMA_COPY_ENABLE=0
export NCCL_TOPO_FILE="/public/home/wangxj/Projects/rccl-test/rccl-tests-0204/topo-input.xml"
# export NCCL_TOPO_FILE="/workspace/rccl-test/rccl-tests-0204/topo-input.xml"
export GLOG_minloglevel=3 # 打印error级别的nccl日志
source /opt/dtk/env.sh
# 导入hipblaslt库
# export LD_LIBRARY_PATH=/data/hipblaslt-install-0904/lib:$LD_LIBRARY_PATH
# export LD_LIBRARY_PATH=/data/hipblaslt-install-dtk-25.04-0212/lib:$LD_LIBRARY_PATH
export LD_LIBRARY_PATH=/public/home/wangxj/Downloads/hipblaslt-install-dtk-25.04-0212/lib:$LD_LIBRARY_PATH
# 更新rocblas
# export LD_LIBRARY_PATH=/data/rocblas-install_qwen1211/lib:$LD_LIBRARY_PATH
# export LD_LIBRARY_PATH=/data/rocblas-install_qwen1228/lib:$LD_LIBRARY_PATH
# export LD_LIBRARY_PATH=/data/rocblas-install-0118-bf16/lib:$LD_LIBRARY_PATH
# export LD_LIBRARY_PATH=/data/rocblas-install-0203-release/lib:$LD_LIBRARY_PATH
export LD_LIBRARY_PATH=/public/home/wangxj/Downloads/rocblas-install-0203-release/lib:$LD_LIBRARY_PATH
# torch控制多流转单流
export ALLREDUCE_STREAM_WITH_COMPUTE=1
export SENDRECV_STREAM_WITH_COMPUTE=1
# prof采集添加同步, 避免卡顿
# export GPU_FLUSH_ON_EXECUTION=1
# export HIP_DIRECT_DISPATCH=0
# 采集rocblas size
# export ROCBLAS_LAYER=3
# export HIPBLASLT_LOG_LEVEL=3
# 采集 fa size
# export FLASH_ATTENTION_PRINT_PARAM=1
#增加编译缓存
export cache_size_limit=64
# CHECKPOINT_PATH=./Llama-2-7b-hf-to-meg-tp1-pp2 #CHECKPOINT_PATH=./tmp_7b #
SAVE_PATH=./tmp_7b
TENSORBOARD_LOGS_PATH=./tmp_7b #$2 #<Specify path>
DATA_PATH="/public/home/wangxj/Downloads/datasets/oscar-1GB-head/oscar-1GB_head-llama3.2_text_document" #<Specify path and file prefix>_text_document
# DATA_PATH="/data/datasets/oscar-1GB-head/oscar-1GB_head-llama3.2_text_document" #<Specify path and file prefix>_text_document
GPT_MODEL_ARGS=(
--num-layers 80 #80 #80 #40 # 20 #
--hidden-size 8192
--ffn-hidden-size 28672
--num-attention-heads 64
--max-position-embeddings 8192
--group-query-attention
--num-query-groups 8
--normalization RMSNorm
--position-embedding-type rope
--untie-embeddings-and-output-weights # 分开处理embed和输出权重, 增加灵活性
)
export NVTE_FLASH_ATTN=1 # 走cutlass
# export NVTE_FLASH_ATTN_TRITON=1 # 走triton_fa
# --transformer-impl transformer_engine # 走core用这两组参数
# --use-mcore-models
# --transformer-impl local # 走legacy用这两组参数
# --use-legacy-models
TRAINING_ARGS=(
--transformer-impl local # 走legacy用这两组参数
--use-legacy-models
--micro-batch-size 1
--global-batch-size 960 #32 #512 #256 # 64 #240 #60 #512 #64
--train-iters 100
--weight-decay 0.1
--adam-beta1 0.9
--adam-beta2 0.95
--init-method-std 0.006
--clip-grad 1.0
--bf16
# --fp16 # 开启fp16需要指定loss-scale
# --loss-scale 1024
--use-distributed-optimizer
--disable-bias-linear
--attention-dropout 0
--hidden-dropout 0
# --no-gradient-accumulation-fusion
--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 # 在dp阶段通信中, 梯度或参数将被直接平均, 而不是先求和(到一个设备)再平均
# --recompute-activations
# --recompute-granularity full # 开启重计算降低显存增加耗时
# --recompute-num-layers 1 #0 #
# --recompute-method block
--overlap-grad-reduce # 重叠ddp grad reduce
# --tp-comm-overlap # tensor parallel comm和gemm重叠, 启动core
# --tp-comm-overlap-rs-dgrad # reduce-scatter和dgrad gemm重叠, 启动core
--use-flash-attn-cutlass
)
# export TORCHINDUCTOR_COORDINATE_DESCENT_TUNING=1
# export TORCHINDUCTOR_BENCHMARK_FUSION=1
# export TORCHINDUCTOR_BENCHMARK_MULTI_TEMPLATES=1
# export TORCHINDUCTOR_MAX_AUTOTUNE=1
# export TORCHINDUCTOR_CACHE_DIR=./cache
# --use-flash-attn-cutlass # cutlass fa
# --use-flash-attn-triton # triton fa
# --use-flash-attn-torch # torch fa
MODEL_PARALLEL_ARGS=(
--sequence-parallel
--tensor-model-parallel-size 4
--pipeline-model-parallel-size 8
# --context-parallel-size 2
# --num-layers-per-virtual-pipeline-stage 5
# --microbatch-group-size-per-virtual-pipeline-stage 1
# --no-overlap-p2p-communication # 开启后
)
DATA_ARGS=(
--data-path $DATA_PATH
--seq-length 4096 #8192 #4096
--split 949,50,1
--tokenizer-type Llama3Tokenizer
--tokenizer-model /public/home/wangxj/Downloads/model_weights/llama3.2/tokenizer.model
# --tokenizer-model /data/model_weights/llama3.2/tokenizer.model
)
EVAL_AND_LOGGING_ARGS=(
--log-interval 1
--log-throughput
--save-interval 1000
--eval-interval 1000
--save $SAVE_PATH
--load $SAVE_PATH
--eval-iters 10
--tensorboard-dir $TENSORBOARD_LOGS_PATH
)
# FINETUNE_ARGS=(
# # --finetune
# # --pretrained-checkpoint $CHECKPOINT_PATH
# --load $CHECKPOINT_PATH
# --no-load-optim
# --no-load-rng
# )
PROFILE_ARGS=(
--profile
--profile-step-start 4
--profile-step-end 5
--use-pytorch-profiler
--profile-ranks 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
--profile-dir prof_data
)
RANK=$OMPI_COMM_WORLD_RANK
LOCAL_RANK=$OMPI_COMM_WORLD_LOCAL_RANK
WORLD_SIZE=$OMPI_COMM_WORLD_SIZE
DIST_URL=${1}
DIST_PORT=34577
DISTRIBUTED_ARGS=(
--rank ${RANK}
--world-size ${WORLD_SIZE}
--local-rank ${LOCAL_RANK}
--dist-url tcp://${DIST_URL}:${DIST_PORT}
)
APP="python -u pretrain_gpt.py \
${GPT_MODEL_ARGS[@]} \
${TRAINING_ARGS[@]} \
${MODEL_PARALLEL_ARGS[@]} \
${DATA_ARGS[@]} \
${EVAL_AND_LOGGING_ARGS[@]} \
${DISTRIBUTED_ARGS[@]} \
"
# 开启profile
# ${PROFILE_ARGS[@]} \
# export HIP_VISIBLE_DEVICES=0,7 # # 4,5,6,7 #,
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 # # 4,5,6,7 #,
# export CUDA_VISIBLE_DEVICES=4,5,6,7 # 0,1,2,3,
# ${APP}
case ${LOCAL_RANK} in
[0])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
# hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
numactl --cpunodebind=0 --membind=0 ${APP}
;;
[1])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
# hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
numactl --cpunodebind=1 --membind=1 ${APP}
;;
[2])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
# hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
numactl --cpunodebind=2 --membind=2 ${APP}
;;
[3])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
numactl --cpunodebind=3 --membind=3 ${APP}
# hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
;;
[4])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
numactl --cpunodebind=4 --membind=4 ${APP}
# hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
;;
[5])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
numactl --cpunodebind=5 --membind=5 ${APP}
# hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
;;
[6])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
numactl --cpunodebind=6 --membind=6 ${APP}
# hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
;;
[7])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
numactl --cpunodebind=7 --membind=7 ${APP}
# hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
;;
esac
\ No newline at end of file
examples/llama/hostfile 0 → 100644
View file @ 1e8185f4
node036 slots=8
node061 slots=8
node062 slots=8
node063 slots=8
node064 slots=8
node065 slots=8
node066 slots=8
node067 slots=8
examples/llama/run_llama2_13B.sh 0 → 100755
View file @ 1e8185f4
for para in $*
do
if [[ $para == --profiling* ]];then
profiling=${para#*=}
fi
done
# Those variables need to modify
GPUS="8" # how many gpus to use
# DTK_ENV="/opt/dtk/env.sh"
DTK_ENV="/public/home/wangxj/Downloads/blas/dtk-25.04.1-rc1/env.sh" # where env.sh of dtk
# NCCL_ENV="/workspace/dcu_megatron/requirements/nccl_wz/env.sh" # where env.sh of nccl (requirements/nccl_wz/env.sh or requirements/nccl_zz/env.sh)
NCCL_ENV="/public/home/wangxj/Projects/dcu_megatron/requirements/nccl_wz/env.sh"
HOST="localhost" # hostname
PORT="11451" # port id
# DATA_PATH="/data/datasets/oscar-1GB-head/oscar-1GB_head-llama2_text_document" # path to oscar-1GB_head-llama2_text_document
DATA_PATH="/public/home/wangxj/Downloads/datasets/oscar-1GB-head/oscar-1GB_head-llama2_text_document"
# TOKENIZER_MODEL_PATH="/data/model_weights/llama2_7b_hf/tokenizer.model" # path to tokenizer.model
TOKENIZER_MODEL_PATH="/public/home/wangxj/Downloads/model_weights/llama2_7b_hf/tokenizer.model"
CHECKPOINT_PATH="./ckpt" # path to ckpt
# Runs Llama2 7B model
mpirun -np ${GPUS} --hostfile hostfile \
--allow-run-as-root \
--bind-to none \
--mca plm_rsh_no_tree_spawn 1 \
bash -c "
source ${DTK_ENV} && \
source ${NCCL_ENV} && \
./train_llama2_13b_1nodes.sh \
${HOST} \
${PORT} \
--data_path=$DATA_PATH \
--tokenizer_path=$TOKENIZER_MODEL_PATH \
--checkpoint_path=$CHECKPOINT_PATH \
--profiling=$profiling" > ./log/log-$((${GPUS} / 8))nodes-`date +%F-%H%M`.log 2>&1
wait
\ No newline at end of file
examples/llama/run_llama2_70B.sh 0 → 100755
View file @ 1e8185f4
for para in $*
do
if [[ $para == --profiling* ]];then
profiling=${para#*=}
fi
done
# Those variables need to modify
GPUS="64" # how many gpus to use
DTK_ENV="/public/home/wangxj/Downloads/blas/dtk-25.04.1-rc1/env.sh" # where env.sh of dtk
# NCCL_ENV="/workspace/dcu_megatron/requirements/nccl_wz/env.sh" # where env.sh of nccl (requirements/nccl_wz/env.sh or requirements/nccl_zz/env.sh)
NCCL_ENV="/public/home/wangxj/Projects/dcu_megatron/requirements/nccl_wz/env.sh"
HOST="node036" # hostname
PORT="11451" # port id
# DATA_PATH="/data/datasets/oscar-1GB-head/oscar-1GB_head-llama2_text_document" # path to oscar-1GB_head-llama2_text_document
DATA_PATH="/public/home/wangxj/Downloads/datasets/oscar-1GB-head/oscar-1GB_head-llama2_text_document"
# TOKENIZER_MODEL_PATH="/data/model_weights/llama2_7b_hf/tokenizer.model" # path to tokenizer.model
TOKENIZER_MODEL_PATH="/public/home/wangxj/Downloads/model_weights/llama2_7b_hf/tokenizer.model"
CHECKPOINT_PATH="./ckpt" # path to ckpt
# Runs Llama2 7B model
mpirun -np ${GPUS} --hostfile hostfile \
--allow-run-as-root \
--bind-to none \
--mca plm_rsh_no_tree_spawn 1 \
bash -c "
source ${DTK_ENV} && \
source ${NCCL_ENV} && \
./train_llama2_70b_8nodes.sh \
${HOST} \
${PORT} \
--data_path=$DATA_PATH \
--tokenizer_path=$TOKENIZER_MODEL_PATH \
--checkpoint_path=$CHECKPOINT_PATH \
--profiling=$profiling" > ./log/log-$((${GPUS} / 8))nodes-`date +%F-%H%M`.log 2>&1
wait
\ No newline at end of file
examples/llama/run_llama2_7B.sh
View file @ 1e8185f4
...@@ -6,29 +6,33 @@ do ...@@ -6,29 +6,33 @@ do
done done
# Those variables need to modify # Those variables need to modify
GPUS="" # how many gpus to use GPUS="8" # how many gpus to use
DTK_ENV="" # where env.sh of dtk # DTK_ENV="/opt/dtk/env.sh"
NCCL_ENV="" # where env.sh of nccl (requirements/nccl_wz/env.sh or requirements/nccl_zz/env.sh) DTK_ENV="/public/home/wangxj/Downloads/blas/dtk-25.04.1-rc1/env.sh" # where env.sh of dtk
HOST="" # hostname # NCCL_ENV="/workspace/dcu_megatron/requirements/nccl_wz/env.sh" # where env.sh of nccl (requirements/nccl_wz/env.sh or requirements/nccl_zz/env.sh)
PORT="" # port id NCCL_ENV="/public/home/wangxj/Projects/dcu_megatron/requirements/nccl_wz/env.sh"
DATA_PATH="" # path to oscar-1GB_head-llama2_text_document HOST="localhost" # hostname
TOKENIZER_MODEL_PATH="" # path to tokenizer.model PORT="11451" # port id
CHECKPOINT_PATH="" # path to ckpt # DATA_PATH="/data/datasets/oscar-1GB-head/oscar-1GB_head-llama2_text_document" # path to oscar-1GB_head-llama2_text_document
DATA_PATH="/public/home/wangxj/Downloads/datasets/oscar-1GB-head/oscar-1GB_head-llama2_text_document"
# TOKENIZER_MODEL_PATH="/data/model_weights/llama2_7b_hf/tokenizer.model" # path to tokenizer.model
TOKENIZER_MODEL_PATH="/public/home/wangxj/Downloads/model_weights/llama2_7b_hf/tokenizer.model"
CHECKPOINT_PATH="./ckpt" # path to ckpt
# Runs Llama2 7B model # Runs Llama2 7B model
mpirun -np ${GPUS} --hostfile hostfile_llama2_7B \ mpirun -np ${GPUS} --hostfile hostfile \
--allow-run-as-root \ --allow-run-as-root \
--bind-to none \ --bind-to none \
--mca plm_rsh_no_tree_spawn 1 \ --mca plm_rsh_no_tree_spawn 1 \
bash -c " bash -c "
source ${DTK_ENV} && \ source ${DTK_ENV} && \
source ${NCCL_ENV} && \ source ${NCCL_ENV} && \
./train_llama2_7b_$((${GPUS} / 8))nodes.sh \ ./train_llama2_7b_1nodes.sh \
${HOST} \ ${HOST} \
${PORT} \ ${PORT} \
--data_path=$DATA_PATH \ --data_path=$DATA_PATH \
--tokenizer_path=$TOKENIZER_MODEL_PATH \ --tokenizer_path=$TOKENIZER_MODEL_PATH \
--checkpoint_path=$CHECKPOINT_PATH \ --checkpoint_path=$CHECKPOINT_PATH \
--profiling=$profiling" > log-$((${GPUS} / 8))nodes-`date +%F-%H%M`.log 2>&1 --profiling=$profiling" > ./log/log-$((${GPUS} / 8))nodes-`date +%F-%H%M`.log 2>&1
wait wait
\ No newline at end of file
examples/llama/run_llama3_70B.sh 0 → 100755
View file @ 1e8185f4
for para in $*
do
if [[ $para == --profiling* ]];then
profiling=${para#*=}
fi
done
# Those variables need to modify
GPUS="64" # how many gpus to use
DTK_ENV="/public/home/wangxj/Downloads/blas/dtk-25.04.1-rc1/env.sh" # where env.sh of dtk
# NCCL_ENV="/workspace/dcu_megatron/requirements/nccl_wz/env.sh" # where env.sh of nccl (requirements/nccl_wz/env.sh or requirements/nccl_zz/env.sh)
NCCL_ENV="/public/home/wangxj/Projects/dcu_megatron/requirements/nccl_wz/env.sh"
HOST="node036" # hostname
PORT="11451" # port id
# DATA_PATH="/data/datasets/oscar-1GB-head/oscar-1GB_head-llama2_text_document" # path to oscar-1GB_head-llama2_text_document
DATA_PATH="/public/home/wangxj/Downloads/datasets/oscar-1GB-head/oscar-1GB_head-llama3.2_text_document"
# TOKENIZER_MODEL_PATH="/data/model_weights/llama2_7b_hf/tokenizer.model" # path to tokenizer.model
TOKENIZER_MODEL_PATH="/public/home/wangxj/Downloads/model_weights/llama3.2/tokenizer.model"
CHECKPOINT_PATH="./ckpt" # path to ckpt
# Runs Llama2 7B model
mpirun -np ${GPUS} --hostfile hostfile \
--allow-run-as-root \
--bind-to none \
--mca plm_rsh_no_tree_spawn 1 \
bash -c "
source ${DTK_ENV} && \
source ${NCCL_ENV} && \
./train_llama3_70b_8nodes.sh \
${HOST} \
${PORT} \
--data_path=$DATA_PATH \
--tokenizer_path=$TOKENIZER_MODEL_PATH \
--checkpoint_path=$CHECKPOINT_PATH \
--profiling=$profiling" > ./log/log-$((${GPUS} / 8))nodes-`date +%F-%H%M`.log 2>&1
wait
\ No newline at end of file
examples/llama/train_llama2_13b_1nodes.sh 0 → 100755
View file @ 1e8185f4
#!/bin/bash
# 使用blaslt和rocblas
export export LD_LIBRARY_PATH=/public/home/wangxj/Downloads/blas/hipblaslt-install0519/lib:$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 #10 # 4 # 20
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 40
--hidden-size 5120
--ffn-hidden-size 13824
--num-attention-heads 40
--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 2
--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} ${APP}
\ No newline at end of file
examples/llama/train_llama2_70b_8nodes.sh 0 → 100755
View file @ 1e8185f4
#!/bin/bash
# 使用blaslt和rocblas
export export LD_LIBRARY_PATH=/public/home/wangxj/Downloads/blas/hipblaslt-install0519/lib:$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 #10 # 4
# 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 80 #80 #80 #40 # 20 #
--hidden-size 8192
--ffn-hidden-size 28672 # 28672
--num-attention-heads 64
--max-position-embeddings 8192
--group-query-attention
--num-query-groups 8
--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 512
--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 4
--pipeline-model-parallel-size 8
--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_tp4-pp2-cp1-tpoverlap-nosyns_20
--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} ${APP}
\ No newline at end of file
examples/llama/train_llama2_7b_1nodes.sh
View file @ 1e8185f4
#!/bin/bash #!/bin/bash
# 使用blaslt和rocblas
export export LD_LIBRARY_PATH=/public/home/wangxj/Downloads/blas/hipblaslt-install0519/lib:$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) INITIALIZATION_ARGS=( --num-workers 2)
for para in $* for para in $*
...@@ -30,7 +34,7 @@ DATA_PATH=${data_path} ...@@ -30,7 +34,7 @@ DATA_PATH=${data_path}
TOKENIZER_MODEL_PATH=${tokenizer_path} TOKENIZER_MODEL_PATH=${tokenizer_path}
CHECKPOINT_PATH=${checkpoint_path} CHECKPOINT_PATH=${checkpoint_path}
# default env # 运行环境参数
DIST_URL=${1} DIST_URL=${1}
DIST_PORT=${2} DIST_PORT=${2}
RANK=$OMPI_COMM_WORLD_RANK RANK=$OMPI_COMM_WORLD_RANK
...@@ -38,12 +42,18 @@ LOCAL_RANK=$OMPI_COMM_WORLD_LOCAL_RANK ...@@ -38,12 +42,18 @@ LOCAL_RANK=$OMPI_COMM_WORLD_LOCAL_RANK
WORLD_SIZE=$OMPI_COMM_WORLD_SIZE WORLD_SIZE=$OMPI_COMM_WORLD_SIZE
CURRENT_DIR="$( cd "$( dirname "$0" )" && pwd )" CURRENT_DIR="$( cd "$( dirname "$0" )" && pwd )"
MEGATRON_PATH=$( dirname $( dirname ${CURRENT_DIR})) MEGATRON_PATH=$( dirname $( dirname ${CURRENT_DIR}))
export PYTHONPATH=${MEGATRON_PATH}/Megatron-LM:$PYTHONPATH
# default env
export GLOG_minloglevel=3 export GLOG_minloglevel=3
export CUDA_DEVICE_MAX_CONNECTIONS=1 export CUDA_DEVICE_MAX_CONNECTIONS=1
export HSA_FORCE_FINE_GRAIN_PCIE=1 export HSA_FORCE_FINE_GRAIN_PCIE=1
export OMP_NUM_THREADS=1 export OMP_NUM_THREADS=1
export GPU_MAX_HW_QUEUES=10 export GPU_MAX_HW_QUEUES=10 # 4 # 20
export PYTHONPATH=${MEGATRON_PATH}/Megatron-LM:$PYTHONPATH
# 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 # torch控制多流转单流
export ALLREDUCE_STREAM_WITH_COMPUTE=1 export ALLREDUCE_STREAM_WITH_COMPUTE=1
...@@ -66,14 +76,14 @@ GPT_MODEL_ARGS=( ...@@ -66,14 +76,14 @@ GPT_MODEL_ARGS=(
--ffn-hidden-size 11008 --ffn-hidden-size 11008
--num-attention-heads 32 --num-attention-heads 32
--max-position-embeddings 4096 --max-position-embeddings 4096
--normalization LightopRMSNorm --normalization RMSNorm # Lightop
--position-embedding-type rope --position-embedding-type rope
--untie-embeddings-and-output-weights --untie-embeddings-and-output-weights
) )
TRAINING_ARGS=( TRAINING_ARGS=(
--transformer-impl local --transformer-impl transformer_engine
--use-legacy-models --use-mcore-models
--micro-batch-size 1 --micro-batch-size 1
--global-batch-size 256 --global-batch-size 256
--train-iters 50 --train-iters 50
...@@ -94,6 +104,8 @@ TRAINING_ARGS=( ...@@ -94,6 +104,8 @@ TRAINING_ARGS=(
--ckpt-format torch --ckpt-format torch
--ddp-average-in-collective --ddp-average-in-collective
--overlap-grad-reduce --overlap-grad-reduce
# --tp-comm-overlap
# --tp-comm-overlap-rs-dgrad
--use-flash-attn --use-flash-attn
) )
...@@ -159,4 +171,4 @@ elif [[ $profiling == "hip" ]]; then ...@@ -159,4 +171,4 @@ elif [[ $profiling == "hip" ]]; then
fi fi
#for hygon cpu #for hygon cpu
${MEGATRON_PATH}/requirements/launch_with_binding.sh ${LOCAL_RANK} ${APP} ${MEGATRON_PATH}/requirements/launch_with_binding.sh ${LOCAL_RANK} ${APP}
\ No newline at end of file
examples/llama/train_llama3_70b_8nodes.sh 0 → 100755
View file @ 1e8185f4
#!/bin/bash
# 使用blaslt和rocblas
export export LD_LIBRARY_PATH=/public/home/wangxj/Downloads/blas/hipblaslt-install0519/lib:$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=20 #10 # 4
# 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 80 #80 #80 #40 # 20 #
--hidden-size 8192
--ffn-hidden-size 28672 # 28672
--num-attention-heads 64
--max-position-embeddings 8192
--group-query-attention
--num-query-groups 8
--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 512
--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 4
--pipeline-model-parallel-size 8
--context-parallel-size 1
--use-distributed-optimizer
--sequence-parallel
)
DATA_ARGS=(
--tokenizer-type Llama3Tokenizer
--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_tp4-pp2-cp1-tpoverlap-nosyns_20
--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}
\ No newline at end of file
examples/qwen/hostfile 0 → 100644
View file @ 1e8185f4
node036 slots=8
node034 slots=8
\ No newline at end of file
examples/qwen/run_qwen1.5_14B.sh 0 → 100755
View file @ 1e8185f4
for para in $*
do
if [[ $para == --profiling* ]];then
profiling=${para#*=}
fi
done
# Those variables need to modify
GPUS="8" # how many gpus to use
# DTK_ENV="/opt/dtk/env.sh"
DTK_ENV="/public/home/wangxj/Downloads/blas/dtk-25.04.1-rc1/env.sh" # where env.sh of dtk
# NCCL_ENV="/workspace/dcu_megatron/requirements/nccl_wz/env.sh" # where env.sh of nccl (requirements/nccl_wz/env.sh or requirements/nccl_zz/env.sh)
NCCL_ENV="/public/home/wangxj/Projects/dcu_megatron/requirements/nccl_wz/env.sh"
HOST="localhost" # hostname
PORT="11451" # port id
# DATA_PATH="/data/datasets/oscar-1GB-head/oscar-1GB_head-llama2_text_document" # path to oscar-1GB_head-llama2_text_document
DATA_PATH="/public/home/wangxj/Downloads/datasets/oscar-1GB-head/oscar-1GB_head-qwen_text_document"
# TOKENIZER_MODEL_PATH="/data/model_weights/llama2_7b_hf/tokenizer.model" # path to tokenizer.model
TOKENIZER_MODEL_PATH="/public/home/wangxj/Downloads/model_weights/qwen1.5"
CHECKPOINT_PATH="./ckpt" # path to ckpt
# Runs Llama2 7B model
mpirun -np ${GPUS} --hostfile hostfile \
--allow-run-as-root \
--bind-to none \
--mca plm_rsh_no_tree_spawn 1 \
bash -c "
source ${DTK_ENV} && \
source ${NCCL_ENV} && \
./train_qwen1.5_14b_1nodes.sh \
${HOST} \
${PORT} \
--data_path=$DATA_PATH \
--tokenizer_path=$TOKENIZER_MODEL_PATH \
--checkpoint_path=$CHECKPOINT_PATH \
--profiling=$profiling" > ./log/log-$((${GPUS} / 8))nodes-`date +%F-%H%M`.log 2>&1
wait
\ No newline at end of file
examples/qwen/run_qwen1.5_32B.sh 0 → 100755
View file @ 1e8185f4
for para in $*
do
if [[ $para == --profiling* ]];then
profiling=${para#*=}
fi
done
# Those variables need to modify
GPUS="16" # how many gpus to use
# DTK_ENV="/opt/dtk/env.sh"
DTK_ENV="/public/home/wangxj/Downloads/blas/dtk-25.04.1-rc1/env.sh" # where env.sh of dtk
# NCCL_ENV="/workspace/dcu_megatron/requirements/nccl_wz/env.sh" # where env.sh of nccl (requirements/nccl_wz/env.sh or requirements/nccl_zz/env.sh)
NCCL_ENV="/public/home/wangxj/Projects/dcu_megatron/requirements/nccl_wz/env.sh"
HOST="node036" # hostname
PORT="11451" # port id
# DATA_PATH="/data/datasets/oscar-1GB-head/oscar-1GB_head-llama2_text_document" # path to oscar-1GB_head-llama2_text_document
DATA_PATH="/public/home/wangxj/Downloads/datasets/oscar-1GB-head/oscar-1GB_head-qwen_text_document"
# TOKENIZER_MODEL_PATH="/data/model_weights/llama2_7b_hf/tokenizer.model" # path to tokenizer.model
TOKENIZER_MODEL_PATH="/public/home/wangxj/Downloads/model_weights/qwen1.5"
CHECKPOINT_PATH="./ckpt" # path to ckpt
# Runs Llama2 7B model
mpirun -np ${GPUS} --hostfile hostfile \
--allow-run-as-root \
--bind-to none \
--mca plm_rsh_no_tree_spawn 1 \
bash -c "
source ${DTK_ENV} && \
source ${NCCL_ENV} && \
./train_qwen1.5_32b_2nodes.sh \
${HOST} \
${PORT} \
--data_path=$DATA_PATH \
--tokenizer_path=$TOKENIZER_MODEL_PATH \
--checkpoint_path=$CHECKPOINT_PATH \
--profiling=$profiling" > ./log/log-$((${GPUS} / 8))nodes-`date +%F-%H%M`.log 2>&1
wait
\ No newline at end of file
examples/qwen/train_qwen1.5_14b_1nodes.sh 0 → 100755
View file @ 1e8185f4
#!/bin/bash
# 使用blaslt和rocblas
export export LD_LIBRARY_PATH=/public/home/wangxj/Downloads/blas/hipblaslt-install0519/lib:$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 #10 # 4 # 20
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 40
--hidden-size 5120
--ffn-hidden-size 13696
--num-attention-heads 40
--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
--add-qkv-bias
--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 2
--pipeline-model-parallel-size 4
--context-parallel-size 1
--use-distributed-optimizer
--sequence-parallel
)
DATA_ARGS=(
--tokenizer-type QwenTokenizer
--merge-file ${TOKENIZER_MODEL_PATH}/merges.txt
--vocab-file ${TOKENIZER_MODEL_PATH}/vocab.json
--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
case ${LOCAL_RANK} in
[0])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
numactl --cpunodebind=0 --membind=0 ${APP}
;;
[1])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
numactl --cpunodebind=1 --membind=1 ${APP}
;;
[2])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
numactl --cpunodebind=2 --membind=2 ${APP}
;;
[3])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
numactl --cpunodebind=3 --membind=3 ${APP}
;;
[4])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
numactl --cpunodebind=4 --membind=4 ${APP}
;;
[5])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
numactl --cpunodebind=5 --membind=5 ${APP}
;;
[6])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
numactl --cpunodebind=6 --membind=6 ${APP}
;;
[7])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
numactl --cpunodebind=7 --membind=7 ${APP}
;;
esac
\ No newline at end of file
examples/qwen/train_qwen1.5_32b_2nodes.sh 0 → 100755
View file @ 1e8185f4
#!/bin/bash
# 使用blaslt和rocblas
export export LD_LIBRARY_PATH=/public/home/wangxj/Downloads/blas/hipblaslt-install0519/lib:$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 #10 # 4 # 20
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 64
--hidden-size 5120
--ffn-hidden-size 27392
--num-attention-heads 40
--max-position-embeddings 32768
--num-query-groups 8
--group-query-attention
--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
--add-qkv-bias
--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 4
--pipeline-model-parallel-size 4
--context-parallel-size 1
--use-distributed-optimizer
--sequence-parallel
)
DATA_ARGS=(
--tokenizer-type QwenTokenizer
--merge-file ${TOKENIZER_MODEL_PATH}/merges.txt
--vocab-file ${TOKENIZER_MODEL_PATH}/vocab.json
--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
case ${LOCAL_RANK} in
[0])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
numactl --cpunodebind=0 --membind=0 ${APP}
;;
[1])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
numactl --cpunodebind=1 --membind=1 ${APP}
;;
[2])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
numactl --cpunodebind=2 --membind=2 ${APP}
;;
[3])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
numactl --cpunodebind=3 --membind=3 ${APP}
;;
[4])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
numactl --cpunodebind=4 --membind=4 ${APP}
;;
[5])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
numactl --cpunodebind=5 --membind=5 ${APP}
;;
[6])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
numactl --cpunodebind=6 --membind=6 ${APP}
;;
[7])
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
numactl --cpunodebind=7 --membind=7 ${APP}
;;
esac
\ No newline at end of file
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment