Merge branch 'sc21' into 'main'

scripts for sc21

See merge request ADLR/megatron-lm!298

examples/sc21/CONFIG.sh

+#!/bin/bash
+
+
+# SLURM options.
+export SLURM_PARTITION=<slurm partition, used to feed -p option in slurm>
+export SLURM_ACCOUNT=<slurm account, used to feed -A option in slurm>
+
+
+# Source code.
+export MEGATRON_CODE_DIR=<megatron source code directory>
+
+
+# This variable is used to mount the relevant part of the filesystem
+# inside the docker container. Note that the `MEGATRON_CODE_DIR` and the
+# launch directory already get mounted; this variable should be used to
+# mount the directories that contain the data and tokenizer files.
+export DOCKER_MOUNT_DIR=<megatron dataset and bpe tokenizer vocab path>
+
+
+# Data and tokenizer files.
+MEGATRON_DATA=<path to megatron processed data>
+BPE_VOCAB_FILE=<path to bpe vocab file>
+BPE_MERGE_FILE=<path to bpe merges file>
+
+
+# Megatron input parameters.
+# `MEGATRON_EXTRA_PARAMS` can be used to provide any extra parameters
+# that are not listed here. 
+export MEGATRON_PARAMS=" ${MEGATRON_EXTRA_PARAMS} \
+	--tensor-model-parallel-size ${TP} \
+	--pipeline-model-parallel-size ${PP} \
+	--micro-batch-size ${MBS} \
+	--global-batch-size ${GBS} \
+        --num-layers ${NLS} \
+        --hidden-size ${HS} \
+        --num-attention-heads ${NAH} \
+	--DDP-impl ${DDP} \
+	--data-path ${MEGATRON_DATA} \
+	--vocab-file ${BPE_VOCAB_FILE} \
+	--merge-file ${BPE_MERGE_FILE} \
+        --log-interval 5 \
+        --seq-length 2048 \
+        --max-position-embeddings 2048 \
+        --train-iters 500 \
+        --lr-decay-iters 320 \
+        --lr 0.0001 \
+	--min-lr 0.00001 \
+        --lr-decay-style cosine \
+        --lr-warmup-fraction 0.01 \
+        --split 969,30,1 \
+        --eval-iters 100 \
+        --eval-interval 1000 \
+        --clip-grad 1.0 \
+        --fp16 \
+	--loss-scale 8192 "
+
+

examples/sc21/README.md

+# Reproducing Figures in SC21 Paper
+
+
+This directory contains some of the scripts that were used to produce the
+results in the [Megatron paper](https://arxiv.org/pdf/2104.04473.pdf) that is
+to appear at [SuperComputing 2021](https://sc21.supercomputing.org/). These
+scripts use [Slurm](https://slurm.schedmd.com/documentation.html) with the
+[pyxis plugin](https://github.com/NVIDIA/pyxis), but can be modified for other
+schedulers as well.
+
+
+## Setup
+
+All the cluster-dependent variables are in [`CONFIG.sh`](./CONFIG.sh). Please
+update the unspecified values (in angle brackets `<...>`) before launching any
+scripts.
+
+
+
+## Scripts
+
+Below is a list of scripts that can be used to reproduce various figures in our
+[paper](https://arxiv.org/pdf/2104.04473.pdf):
+
+* [run_table_1.sh](./run_table_1.sh): Table 1 showing weak-scaling throughput
+for GPT models ranging from 1 billion to 1 trillion parameters.
+* [run_figure_11.sh](./run_figure_11.sh): Figure 11 showing the weak-scaling
+performance of pipeline parallelism.
+* [run_figure_12.sh](./run_figure_12.sh): Figure 12 showing the effect of
+the interleaved schedule on a 175B GPT model.
+* [run_figure_13.sh](./run_figure_13.sh): Figure 13 showing the effect of
+different degrees of pipeline and tensor model parallelism on a model with
+162.2 billion parameters.
+* [run_figure_14.sh](./run_figure_14.sh): Figure 14 showing the effect of
+different degrees of data and pipeline model parallelism on a model with
+5.9 billion parameters.
+* [run_figure_15.sh](./run_figure_15.sh): Figure 15 showing the effect of
+different degrees of data and tensor model parallelism on a model with
+5.9 billion parameters.
+* [run_figure_16.sh](./run_figure_16.sh): Figure 16 showing the effect of
+microbatch size.
+* [run_figure_17.sh](./run_figure_17.sh): Figure 17 showing the effect of
+activation recomputation.
+* [run_figure_18.sh](./run_figure_18.sh): Figure 18 showing the effect of
+the scatter-gather communication optimization.

examples/sc21/SBATCH.sh

+#!/bin/bash
+
+
+sbatch -p ${SLURM_PARTITION} \
+       -A ${SLURM_ACCOUNT} \
+       --job-name=${JOB_NAME} \
+       --nodes=${NNODES} \
+       --export=MEGATRON_CODE_DIR,MEGATRON_PARAMS,DOCKER_MOUNT_DIR SRUN.sh
+
+exit 0
+
+
+

examples/sc21/SRUN.sh

+#!/bin/bash
+
+#SBATCH -t 0:30:00 --exclusive --mem=0 --overcommit --ntasks-per-node=8
+
+
+THIS_DIR=`pwd`
+DATETIME=`date +'date_%y-%m-%d_time_%H-%M-%S'`
+mkdir -p ${THIS_DIR}/logs
+
+
+CMD="python -u ${MEGATRON_CODE_DIR}/pretrain_gpt.py ${MEGATRON_PARAMS}"
+
+
+srun -l \
+     --container-image "nvcr.io#nvidia/pytorch:20.12-py3" \
+     --container-mounts "${THIS_DIR}:${THIS_DIR},${MEGATRON_CODE_DIR}:${MEGATRON_CODE_DIR},${DOCKER_MOUNT_DIR}:${DOCKER_MOUNT_DIR}" \
+     --output=${THIS_DIR}/logs/%x_%j_$DATETIME.log sh -c "${CMD}"
+

examples/sc21/run_figure_11.sh

+#!/bin/bash
+
+# ================================
+# Choose the case to run.
+# ================================
+
+# Pipeline-parallel size options = [1, 2, 4, 8].
+PP=1
+
+# Batch size (global batch size) options = [8, 128].
+GBS=8
+
+
+
+
+
+# Set pipeline-parallel size options.
+NLS=$((3*PP))
+NNODES=${PP}
+
+
+# Other params.
+TP=8
+MBS=1
+HS=20480
+NAH=128
+DDP=local
+MEGATRON_EXTRA_PARAMS="--checkpoint-activations "
+
+
+# Name of the job.
+export JOB_NAME=results_figure_11_pipeline_parallel_size_${PP}_batch_size_${GBS}
+
+
+# Import the configs.
+. `pwd`/CONFIG.sh
+
+
+# Submit the job.
+. `pwd`/SBATCH.sh
+
+
+exit 0
+
+
+

examples/sc21/run_figure_12.sh

+#!/bin/bash
+
+# ================================
+# Choose the case to run.
+# ================================
+
+# Interleaved schedule options = [YES, NO].
+INTERLEAVED=YES
+
+# Batch size (global batch size) options = [12, 24, 36, ..., 60].
+GBS=12
+
+
+
+
+
+# Set interleaved schedule options.
+if [ ${INTERLEAVED} == "YES" ]; then
+    MEGATRON_EXTRA_PARAMS="--checkpoint-activations --num-layers-per-virtual-pipeline-stage 2 "
+elif [ ${INTERLEAVED} == "NO" ]; then
+    MEGATRON_EXTRA_PARAMS="--checkpoint-activations "
+else
+    echo "Invalid configuration"
+    exit 1
+fi
+
+
+# Other params.
+TP=8
+PP=12
+MBS=1
+NLS=96
+HS=12288
+NAH=96
+DDP=local
+NNODES=12
+
+
+# Name of the job.
+export JOB_NAME=results_figure_12_interleaved_${INTERLEAVED}_batch_size_${GBS}
+
+
+# Import the configs.
+. `pwd`/CONFIG.sh
+
+
+# Submit the job.
+. `pwd`/SBATCH.sh
+
+
+exit 0
+
+
+

examples/sc21/run_figure_13.sh

+#!/bin/bash
+
+# ================================
+# Choose the case to run.
+# ================================
+
+# Pipeline-parallel size options = [2, 4, 8, 16, 32].
+PP=2
+
+# Batch size (global batch size) options = [32, 128].
+GBS=32
+
+
+
+
+
+# Set pipeline-parallel and tensor-parallel size options.
+TP=$((64/PP))
+
+
+# Other params.
+MBS=1
+NLS=32
+HS=20480
+NAH=128
+DDP=local
+MEGATRON_EXTRA_PARAMS="--checkpoint-activations "
+NNODES=8
+
+
+# Name of the job.
+export JOB_NAME=results_figure_13_pipeline_parallel_size_${PP}_tensor_parallel_size_${TP}_batch_size_${GBS}
+
+
+# Import the configs.
+. `pwd`/CONFIG.sh
+
+
+# Submit the job.
+. `pwd`/SBATCH.sh
+
+
+exit 0
+
+
+

examples/sc21/run_figure_14.sh

+#!/bin/bash
+
+# ================================
+# Choose the case to run.
+# ================================
+
+# Pipeline-parallel size options = [2, 4, 8, 16, 32].
+PP=2
+
+# Batch size (global batch size) options = [32, 512].
+GBS=32
+
+
+
+
+
+# Set pipeline-parallel and data-parallel size options.
+DP=$((64/PP))
+
+
+# Other params.
+TP=1
+MBS=1
+NLS=32
+HS=3840
+NAH=32
+DDP=local
+MEGATRON_EXTRA_PARAMS="--checkpoint-activations "
+NNODES=8
+
+
+# Name of the job.
+export JOB_NAME=results_figure_14_pipeline_parallel_size_${PP}_data_parallel_size_${DP}_batch_size_${GBS}
+
+
+# Import the configs.
+. `pwd`/CONFIG.sh
+
+
+# Submit the job.
+. `pwd`/SBATCH.sh
+
+
+exit 0
+
+
+

examples/sc21/run_figure_15.sh

+#!/bin/bash
+
+# ================================
+# Choose the case to run.
+# ================================
+
+# Tensor-parallel size options = [2, 4, 8, 16, 32].
+TP=2
+
+# Batch size (global batch size) options = [32, 128, 512].
+GBS=32
+
+
+
+
+
+# Set tensor-parallel and data-parallel size options.
+DP=$((64/TP))
+
+
+# Other params.
+PP=1
+MBS=1
+NLS=32
+HS=3840
+NAH=32
+DDP=local
+MEGATRON_EXTRA_PARAMS="--checkpoint-activations "
+NNODES=8
+
+
+# Name of the job.
+export JOB_NAME=results_figure_15_tensor_parallel_size_${TP}_data_parallel_size_${DP}_batch_size_${GBS}
+
+
+# Import the configs.
+. `pwd`/CONFIG.sh
+
+
+# Submit the job.
+. `pwd`/SBATCH.sh
+
+
+exit 0
+
+
+

examples/sc21/run_figure_16.sh

+#!/bin/bash
+
+# ================================
+# Choose the case to run.
+# ================================
+
+# Microbatch size options = [1, 2, 4, 8].
+MBS=1
+
+# Batch size (global batch size) options = [128, 512].
+GBS=128
+
+
+
+
+
+# Other params.
+TP=8
+PP=8
+NLS=32
+HS=15360
+NAH=128
+DDP=local
+MEGATRON_EXTRA_PARAMS="--checkpoint-activations "
+NNODES=8
+
+
+# Name of the job.
+export JOB_NAME=results_figure_16_microbatch_size_${MBS}_batch_size_${GBS}
+
+
+# Import the configs.
+. `pwd`/CONFIG.sh
+
+
+# Submit the job.
+. `pwd`/SBATCH.sh
+
+
+exit 0
+
+
+

examples/sc21/run_figure_17.sh

+#!/bin/bash
+
+# ================================
+# Choose the case to run.
+# ================================
+
+# Activation recomputation options = [YES, NO].
+ACTIVATION_RECOMPUTATION=YES
+
+# Batch size (global batch size) options = [1, 2, 4, ..., 256].
+GBS=1
+
+
+
+
+
+# Set activation recomputation.
+if [ ${ACTIVATION_RECOMPUTATION} == "YES" ]; then
+    MEGATRON_EXTRA_PARAMS="--checkpoint-activations "
+elif [ ${ACTIVATION_RECOMPUTATION} == "NO" ]; then
+    MEGATRON_EXTRA_PARAMS=""
+else
+    echo "Invalid configuration"
+    exit 1
+fi
+
+
+# Other params.
+TP=8
+PP=16
+MBS=1
+NLS=80
+HS=12288
+NAH=96
+DDP=local
+NNODES=16
+
+
+# Name of the job.
+export JOB_NAME=results_figure_17_activation_recomputation_${ACTIVATION_RECOMPUTATION}_batch_size_${GBS}
+
+
+# Import the configs.
+. `pwd`/CONFIG.sh
+
+
+# Submit the job.
+. `pwd`/SBATCH.sh
+
+
+exit 0
+
+
+

examples/sc21/run_figure_18.sh

+#!/bin/bash
+
+# ================================
+# Choose the case to run.
+# ================================
+
+# Scatter-gather communication optimization options = [YES, NO].
+SCATTER_GATHER=YES
+
+# Batch size (global batch size) options = [12, 24, 36, ..., 60].
+GBS=12
+
+
+
+
+
+# Set scatter-gather communication optimization options.
+if [ ${SCATTER_GATHER} == "YES" ]; then
+    MEGATRON_EXTRA_PARAMS="--checkpoint-activations --num-layers-per-virtual-pipeline-stage 2 "
+elif [ ${SCATTER_GATHER} == "NO" ]; then
+    MEGATRON_EXTRA_PARAMS="--checkpoint-activations --num-layers-per-virtual-pipeline-stage 2 --no-scatter-gather-tensors-in-pipeline "
+else
+    echo "Invalid configuration"
+    exit 1
+fi
+
+
+# Other params.
+TP=8
+PP=12
+MBS=1
+NLS=96
+HS=12288
+NAH=96
+DDP=local
+NNODES=12
+
+
+# Name of the job.
+export JOB_NAME=results_figure_18_scatter_gather_${SCATTER_GATHER}_batch_size_${GBS}
+
+
+# Import the configs.
+. `pwd`/CONFIG.sh
+
+
+# Submit the job.
+. `pwd`/SBATCH.sh
+
+
+exit 0
+
+
+

examples/sc21/run_table_1.sh

+#!/bin/bash
+
+# ================================
+# Choose the case to run.
+# ================================
+# model size options = [1.7B, 3.6B, 7.5B, 18B, 39B, 76B, 145B, 310B, 530B, 1T]
+MODEL_SIZE=1.7B
+
+
+
+
+
+
+if [ ${MODEL_SIZE} == "1.7B" ]; then
+    TP=1
+    PP=1
+    MBS=16
+    GBS=512
+    NLS=24
+    HS=2304
+    NAH=24
+    DDP=torch
+    NNODES=4
+    MEGATRON_EXTRA_PARAMS="--checkpoint-activations "
+elif [ ${MODEL_SIZE} == "3.6B" ]; then
+    TP=2
+    PP=1
+    MBS=16
+    GBS=512
+    NLS=30
+    HS=3072
+    NAH=32
+    DDP=torch
+    NNODES=8
+    MEGATRON_EXTRA_PARAMS="--checkpoint-activations "
+elif [ ${MODEL_SIZE} == "7.5B" ]; then
+    TP=4
+    PP=1
+    MBS=16
+    GBS=512
+    NLS=36
+    HS=4096
+    NAH=32
+    DDP=torch
+    NNODES=16
+    MEGATRON_EXTRA_PARAMS="--checkpoint-activations "
+elif [ ${MODEL_SIZE} == "18B" ]; then
+    TP=8
+    PP=1
+    MBS=8
+    GBS=1024
+    NLS=40
+    HS=6144
+    NAH=48
+    DDP=torch
+    NNODES=32
+    MEGATRON_EXTRA_PARAMS="--checkpoint-activations "
+elif [ ${MODEL_SIZE} == "39B" ]; then
+    TP=8
+    PP=2
+    MBS=4
+    GBS=1536
+    NLS=48
+    HS=8192
+    NAH=64
+    DDP=local
+    NNODES=64
+    MEGATRON_EXTRA_PARAMS="--checkpoint-activations "
+elif [ ${MODEL_SIZE} == "76B" ]; then
+    TP=8
+    PP=4
+    MBS=2
+    GBS=1792
+    NLS=60
+    HS=10240
+    NAH=80
+    DDP=local
+    NNODES=128
+    MEGATRON_EXTRA_PARAMS="--checkpoint-activations --num-layers-per-virtual-pipeline-stage 5"
+elif [ ${MODEL_SIZE} == "145B" ]; then
+    TP=8
+    PP=8
+    MBS=2
+    GBS=2304
+    NLS=80
+    HS=12288
+    NAH=96
+    DDP=local
+    NNODES=192
+    MEGATRON_EXTRA_PARAMS="--checkpoint-activations --num-layers-per-virtual-pipeline-stage 5 "
+elif [ ${MODEL_SIZE} == "310B" ]; then
+    TP=8
+    PP=16
+    MBS=1
+    GBS=2160
+    NLS=96
+    HS=16384
+    NAH=128
+    DDP=local
+    NNODES=240
+    MEGATRON_EXTRA_PARAMS="--checkpoint-activations --num-layers-per-virtual-pipeline-stage 3 "
+elif [ ${MODEL_SIZE} == "530B" ]; then
+    TP=8
+    PP=35
+    MBS=1
+    GBS=2520
+    NLS=105
+    HS=20480
+    NAH=128
+    DDP=local
+    NNODES=315
+    MEGATRON_EXTRA_PARAMS="--checkpoint-activations --num-layers-per-virtual-pipeline-stage 1 "
+elif [ ${MODEL_SIZE} == "1T" ]; then
+    TP=8
+    PP=64
+    MBS=1
+    GBS=3072
+    NLS=128
+    HS=25600
+    NAH=160
+    DDP=local
+    NNODES=384
+    MEGATRON_EXTRA_PARAMS="--checkpoint-activations "
+else
+    echo "Invalid configuration"
+    exit 1
+fi
+
+
+# Name of the job
+export JOB_NAME=results_table_1_model_size_${MODEL_SIZE}
+
+
+# Import the configs.
+. `pwd`/CONFIG.sh
+
+
+# Submit the job.
+. `pwd`/SBATCH.sh
+
+
+exit 0
+
+
+