Diff with nvidia main

examples/finetune_bigcode_model.slurm

@@ -0,0 +1,144 @@
+#!/bin/bash
+#SBATCH --job-name=starcoderpy
+#SBATCH --nodes=64
+#SBATCH --ntasks-per-node=1
+#SBATCH --exclusive
+#SBATCH --gres=gpu:8
+#SBATCH --partition=production-cluster
+#SBATCH --output=/fsx/leandro/logs/starcoderpy/bcs-%x-%j.out
+
+set -x -e
+source /admin/home/leandro/.bashrc
+
+conda activate megatron
+
+echo "START TIME: $(date)"
+
+# File Path setup
+SCRIPT_REPO=/fsx/leandro/git/Megatron-LM-BC
+pushd $SCRIPT_REPO
+
+LOG_PATH=$SCRIPT_REPO/main_log.txt
+
+# Training setup
+GPUS_PER_NODE=8
+MASTER_ADDR=$(scontrol show hostnames $SLURM_JOB_NODELIST | head -n 1)
+MASTER_PORT=6000
+NNODES=$SLURM_NNODES
+NODE_RANK=$SLURM_PROCID
+WORLD_SIZE=$(($GPUS_PER_NODE*$NNODES))
+
+# File path setup
+STARCODER_PATH=/fsx/boomcode/starcoder/
+CHECKPOINT_PATH=/fsx/boomcode/starcoderpy/$SLURM_JOB_ID
+TOKENIZER_FILE=/fsx/boomcode/tokenizer-starcoder/tokenizer.json
+WEIGHTS_TRAIN=/fsx/boomcode/datamix_python/train_data_paths.txt.tmp
+WEIGHTS_VALID=/fsx/boomcode/datamix_python/valid_data_paths.txt.tmp
+DATA_PATH=/fsx/boomcode/tokenized/python/
+mkdir -p $CHECKPOINT_PATH/tensorboard
+
+GPT_ARGS="\
+        --tensor-model-parallel-size 4 \
+       --pipeline-model-parallel-size 4 \
+       --sequence-parallel \
+       --num-layers 40 \
+       --hidden-size 6144 \
+       --num-attention-heads 48 \
+       --attention-head-type multiquery \
+       --init-method-std 0.01275 \
+       --seq-length 8192 \
+       --max-position-embeddings 8192 \
+       --attention-dropout 0.1 \
+       --hidden-dropout 0.1 \
+       --micro-batch-size 1 \
+       --global-batch-size 512 \
+       --lr 0.00005 \
+       --min-lr 0.000005 \
+       --train-iters 258500 \
+       --lr-decay-iters 8500 \
+       --lr-decay-style cosine \
+       --lr-warmup-iters 500 \
+       --weight-decay .1 \
+       --adam-beta2 .95 \
+       --clip-grad 1.0 \
+       --bf16 \
+       --use-flash-attn \
+       --fim-rate 0.5 \
+       --log-interval 10 \
+       --save-interval 2500 \
+       --eval-interval 100 \
+       --eval-iters 10 \
+       --valid-num-workers 0 \
+       --override-opt_param-scheduler \
+       --no-load-optim \
+       --no-load-rng \
+       --finetune \
+"
+
+# --dataloader-type cyclic\
+TENSORBOARD_ARGS="--tensorboard-dir ${CHECKPOINT_PATH}/tensorboard"
+
+CMD=" \
+    $SCRIPT_REPO/pretrain_gpt.py \
+    $GPT_ARGS \
+    --tokenizer-type TokenizerFromFile \
+    --tokenizer-file $TOKENIZER_FILE \
+    --save $CHECKPOINT_PATH \
+    --load $STARCODER_PATH \
+    --train-weighted-split-paths-path $WEIGHTS_TRAIN \
+    --valid-weighted-split-paths-path $WEIGHTS_VALID \
+    --structured-logs \
+    --structured-logs-dir $CHECKPOINT_PATH/logs \
+    $TENSORBOARD_ARGS \
+    --wandb-entity-name lvwerra \
+    --wandb-project-name starcoder-py \
+    "
+
+# --data-path $DATA_PATH\gpt2-preprocessed_content_document 
+
+export LAUNCHER="python -u -m torch.distributed.run \
+    --nproc_per_node $GPUS_PER_NODE \
+    --nnodes $NNODES \
+    --rdzv_endpoint $MASTER_ADDR:$MASTER_PORT \
+    --rdzv_backend c10d \
+    --max_restarts 0 \
+    --tee 3 \
+    "
+
+echo $CMD
+
+# hide duplicated errors using this hack - will be properly fixed in pt-1.12
+# export TORCHELASTIC_ERROR_FILE=/tmp/torch-elastic-error.json
+
+# force crashing on nccl issues like hanging broadcast
+export NCCL_ASYNC_ERROR_HANDLING=1
+# export NCCL_DEBUG=INFO
+# export NCCL_DEBUG_SUBSYS=COLL
+# export NCCL_SOCKET_NTHREADS=1
+# export NCCL_NSOCKS_PERTHREAD=1
+# export CUDA_LAUNCH_BLOCKING=1
+
+# AWS specific
+export NCCL_PROTO=simple
+export RDMAV_FORK_SAFE=1
+export FI_EFA_FORK_SAFE=1
+export FI_EFA_USE_DEVICE_RDMA=1
+export FI_PROVIDER=efa
+export FI_LOG_LEVEL=1
+export NCCL_IB_DISABLE=1
+export NCCL_SOCKET_IFNAME=ens
+
+export CUDA_HOME=/usr/local/cuda-11.6
+
+# srun error handling:
+# --wait=60: wait 60 sec after the first task terminates before terminating all remaining tasks
+# --kill-on-bad-exit=1: terminate a step if any task exits with a non-zero exit code
+SRUN_ARGS=" \
+    --wait=60 \
+    --kill-on-bad-exit=1 \
+    "
+
+# py-spy top -s -i -n -- $LAUNCHER --node_rank $SLURM_PROCID --role $SLURMD_NODENAME: $CMD
+clear; srun $SRUN_ARGS --jobid $SLURM_JOB_ID bash -c "$LAUNCHER --node_rank \$SLURM_PROCID --role \$SLURMD_NODENAME: $CMD" 2>&1 | tee $LOG_PATH
+
+echo "END TIME: $(date)"

examples/finetune_starcoderplus.slurm

@@ -0,0 +1,141 @@
+#!/bin/bash
+#SBATCH --job-name=starcoderplus
+#SBATCH --nodes=64
+#SBATCH --ntasks-per-node=1
+#SBATCH --exclusive
+#SBATCH --gres=gpu:8
+#SBATCH --partition=production-cluster
+#SBATCH --output=/fsx/leandro/logs/starcoderplus/bcs-%x-%j.out
+
+set -x -e
+source /admin/home/leandro/.bashrc
+
+conda activate megatron
+
+echo "START TIME: $(date)"
+
+# File Path setup
+SCRIPT_REPO=/fsx/leandro/git/Megatron-LM-BC
+pushd $SCRIPT_REPO
+
+LOG_PATH=$SCRIPT_REPO/main_log.txt
+
+# Training setup
+GPUS_PER_NODE=8
+MASTER_ADDR=$(scontrol show hostnames $SLURM_JOB_NODELIST | head -n 1)
+MASTER_PORT=6000
+NNODES=$SLURM_NNODES
+NODE_RANK=$SLURM_PROCID
+WORLD_SIZE=$(($GPUS_PER_NODE*$NNODES))
+
+# File path setup
+STARCODER_PATH=/fsx/boomcode/starcoder/
+CHECKPOINT_PATH=/fsx/boomcode/starcoderplus/$SLURM_JOB_ID
+TOKENIZER_FILE=/fsx/boomcode/tokenizer-starcoder/tokenizer.json
+WEIGHTS_TRAIN=/fsx/boomcode/datamix/train_data_paths.txt.tmp
+WEIGHTS_VALID=/fsx/boomcode/datamix/valid_data_paths.txt.tmp
+
+mkdir -p $CHECKPOINT_PATH/tensorboard
+
+GPT_ARGS="\
+        --tensor-model-parallel-size 4 \
+       --pipeline-model-parallel-size 4 \
+       --sequence-parallel \
+       --num-layers 40 \
+       --hidden-size 6144 \
+       --num-attention-heads 48 \
+       --attention-head-type multiquery \
+       --init-method-std 0.01275 \
+       --seq-length 8192 \
+       --max-position-embeddings 8192 \
+       --attention-dropout 0.1 \
+       --hidden-dropout 0.1 \
+       --micro-batch-size 1 \
+       --global-batch-size 512 \
+       --lr 0.0001 \
+       --min-lr 0.00001 \
+       --train-iters 400000 \
+       --lr-decay-iters 150000 \
+       --lr-decay-style cosine \
+       --lr-warmup-iters 1000 \
+       --weight-decay .1 \
+       --adam-beta2 .95 \
+       --clip-grad 1.0 \
+       --bf16 \
+       --use-flash-attn \
+       --fim-rate 0.5 \
+       --log-interval 10 \
+       --save-interval 2500 \
+       --eval-interval 2500 \
+       --eval-iters 2 \
+       --valid-num-workers 0 \
+       --override-opt_param-scheduler \
+       --no-load-optim \
+       --no-load-rng \
+       --finetune \
+"
+
+TENSORBOARD_ARGS="--tensorboard-dir ${CHECKPOINT_PATH}/tensorboard"
+
+CMD=" \
+    $SCRIPT_REPO/pretrain_gpt.py \
+    $GPT_ARGS \
+    --tokenizer-type TokenizerFromFile \
+    --tokenizer-file $TOKENIZER_FILE \
+    --save $CHECKPOINT_PATH \
+    --load $STARCODER_PATH \
+    --train-weighted-split-paths-path $WEIGHTS_TRAIN \
+    --valid-weighted-split-paths-path $WEIGHTS_VALID \
+    --structured-logs \
+    --structured-logs-dir $CHECKPOINT_PATH/logs \
+    $TENSORBOARD_ARGS \
+    --wandb-entity-name lvwerra \
+    --wandb-project-name starcoder-plus \
+    "
+
+export LAUNCHER="python -u -m torch.distributed.run \
+    --nproc_per_node $GPUS_PER_NODE \
+    --nnodes $NNODES \
+    --rdzv_endpoint $MASTER_ADDR:$MASTER_PORT \
+    --rdzv_backend c10d \
+    --max_restarts 0 \
+    --tee 3 \
+    "
+
+echo $CMD
+
+# hide duplicated errors using this hack - will be properly fixed in pt-1.12
+# export TORCHELASTIC_ERROR_FILE=/tmp/torch-elastic-error.json
+
+# force crashing on nccl issues like hanging broadcast
+export NCCL_ASYNC_ERROR_HANDLING=1
+# export NCCL_DEBUG=INFO
+# export NCCL_DEBUG_SUBSYS=COLL
+# export NCCL_SOCKET_NTHREADS=1
+# export NCCL_NSOCKS_PERTHREAD=1
+# export CUDA_LAUNCH_BLOCKING=1
+
+# AWS specific
+export NCCL_PROTO=simple
+export RDMAV_FORK_SAFE=1
+export FI_EFA_FORK_SAFE=1
+export FI_EFA_USE_DEVICE_RDMA=1
+export FI_PROVIDER=efa
+export FI_LOG_LEVEL=1
+export NCCL_IB_DISABLE=1
+export NCCL_SOCKET_IFNAME=ens
+
+export CUDA_HOME=/usr/local/cuda-11.6
+
+# srun error handling:
+# --wait=60: wait 60 sec after the first task terminates before terminating all remaining tasks
+# --kill-on-bad-exit=1: terminate a step if any task exits with a non-zero exit code
+SRUN_ARGS=" \
+    --wait=60 \
+    --kill-on-bad-exit=1 \
+    "
+
+# py-spy top -s -i -n -- $LAUNCHER --node_rank $SLURM_PROCID --role $SLURMD_NODENAME: $CMD
+clear; srun $SRUN_ARGS --jobid $SLURM_JOB_ID bash -c "$LAUNCHER --node_rank \$SLURM_PROCID --role \$SLURMD_NODENAME: $CMD" 2>&1 | tee $LOG_PATH
+
+echo "END TIME: $(date)"