#!/bin/bash
export CUDA_DEVICE_MAX_CONNECTIONS=1
export HCCL_DETERMINISITIC=True
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
GPUS_PER_NODE=8
MASTER_ADDR=localhost
MASTER_PORT=6079
NNODES=1
NODE_RANK=0
WORLD_SIZE=$(($GPUS_PER_NODE*$NNODES))
basepath=$(cd `dirname $0`; cd ../../../; pwd)
CKPT_SAVE_DIR="/data/ckpt"
CKPT_LOAD_DIR="/data/ci/Llama2-mcore-tp2/"
DATA_PATH="/data/ci/llama2_pack/alpaca"
TOKENIZER_MODEL="/data/llama-2-7b-hf"
TP=2
PP=1
CP=4
DISTRIBUTED_ARGS=(
--nproc_per_node $GPUS_PER_NODE
--nnodes $NNODES
--node_rank $NODE_RANK
--master_addr $MASTER_ADDR
--master_port $MASTER_PORT
)
DIST_ALGO=(
--tensor-model-parallel-size ${TP}
--pipeline-model-parallel-size ${PP}
--context-parallel-size ${CP}
--cp-window-size 1
--use-fused-ring-attention-update
--context-parallel-algo adaptive_cp_algo
--cp-attention-mask-type general
--adaptive-cp-manually-set-mask-list
--adaptive-cp-dynamic-attn-mask
--adaptive-cp-only-reschedule
--sequence-parallel
)
MODEL_ARGS=(
--use-mcore-models
--num-layers 32
--hidden-size 4096
--ffn-hidden-size 11008
--num-attention-heads 32
--seq-length 4096
--max-position-embeddings 4096
)
ACCELERATE_ARGS=(
--reuse-fp32-param
--overlap-grad-reduce
--overlap-param-gather
--use-distributed-optimizer
--recompute-activation-function
--recompute-activation-function-num-layers 1
)
TRAINING_ARGS=(
--tokenizer-type PretrainedFromHF
--tokenizer-name-or-path ${TOKENIZER_MODEL}
--tokenizer-not-use-fast
--micro-batch-size 1
--global-batch-size 8
--make-vocab-size-divisible-by 1
--lr 1.25e-6
--train-iters 15
--lr-decay-style cosine
--untie-embeddings-and-output-weights
--disable-bias-linear
--attention-dropout 0.0
--init-method-std 0.01
--hidden-dropout 0.0
--position-embedding-type rope
--normalization RMSNorm
--use-fused-rmsnorm
--swiglu
--use-flash-attn
--reset-position-ids
--no-masked-softmax-fusion
--attention-softmax-in-fp32
--min-lr 1.25e-7
--weight-decay 1e-1
--lr-warmup-fraction 0.01
--clip-grad 1.0
--adam-beta1 0.9
--initial-loss-scale 65536
--adam-beta2 0.95
--no-gradient-accumulation-fusion
--no-load-optim
--no-load-rng
--use-fused-swiglu
--use-fused-rotary-pos-emb
--overlap-grad-reduce
--bf16
--finetune
--stage sft
--is-instruction-dataset
)
DATA_ARGS=(
--data-path $DATA_PATH
--split 949,50,1
)
OUTPUT_ARGS=(
--log-interval 1
--save-interval 10000
--eval-interval 1000
--eval-iters 0
--log-throughput
)
torchrun ${DISTRIBUTED_ARGS[@]} $basepath/posttrain_gpt.py \
${DIST_ALGO[@]} \
${MODEL_ARGS[@]} \
${TRAINING_ARGS[@]} \
${DATA_ARGS[@]} \
${OUTPUT_ARGS[@]} \
${ACCELERATE_ARGS[@]} \
--load ${CKPT_LOAD_DIR} \
--save ${CKPT_SAVE_DIR} \
--distributed-backend nccl