diff --git a/scripts/run_1.14G_dp16_tp16_pp1_acc4_mbs2_seq32768_zero1_tpmodeALL_vocab32k.sh b/scripts/run_1.14G_dp16_tp16_pp1_acc4_mbs2_seq32768_zero1_tpmodeALL_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..421e56772a6931f44a882d7e102dec0cb4d8cdcc --- /dev/null +++ b/scripts/run_1.14G_dp16_tp16_pp1_acc4_mbs2_seq32768_zero1_tpmodeALL_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp16_tp16_pp1_acc4_mbs2_seq32768_zero1_tpmodeALL_vocab32k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=32 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.14G_dp16_tp16_pp1_acc4_mbs2_seq32768_zero1_tpmodeALL_vocab32k.yaml diff --git a/scripts/run_1.14G_dp2_tp128_pp1_acc128_mbs8_seq2048_zero1_tpmodeALL_vocab32k.sh b/scripts/run_1.14G_dp2_tp128_pp1_acc128_mbs8_seq2048_zero1_tpmodeALL_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..309af6c52a7cba868b91e4a2fd5e178b96536885 --- /dev/null +++ b/scripts/run_1.14G_dp2_tp128_pp1_acc128_mbs8_seq2048_zero1_tpmodeALL_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp2_tp128_pp1_acc128_mbs8_seq2048_zero1_tpmodeALL_vocab32k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=32 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.14G_dp2_tp128_pp1_acc128_mbs8_seq2048_zero1_tpmodeALL_vocab32k.yaml diff --git a/scripts/run_1.14G_dp2_tp128_pp1_acc32_mbs8_seq2048_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp2_tp128_pp1_acc32_mbs8_seq2048_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..69174c1f7b47650a9a23b351eec3181538041824 --- /dev/null +++ b/scripts/run_1.14G_dp2_tp128_pp1_acc32_mbs8_seq2048_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp2_tp128_pp1_acc32_mbs8_seq2048_zero1_tpmodeRED_vocab32k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=32 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.14G_dp2_tp128_pp1_acc32_mbs8_seq2048_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp2_tp16_pp1_acc4_mbs16_seq8192_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp2_tp16_pp1_acc4_mbs16_seq8192_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..a358eb66c5718d36fff1ad4c96a4ca57f3db28f6 --- /dev/null +++ b/scripts/run_1.14G_dp2_tp16_pp1_acc4_mbs16_seq8192_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp2_tp16_pp1_acc4_mbs16_seq8192_zero1_tpmodeRED_vocab32k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=4 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.14G_dp2_tp16_pp1_acc4_mbs16_seq8192_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp2_tp4_pp1_acc64_mbs16_seq2048_zero1_tpmodeALL_vocab32k.sh b/scripts/run_1.14G_dp2_tp4_pp1_acc64_mbs16_seq2048_zero1_tpmodeALL_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..1ae79f35a8f7ac87f55b994600b817e1ae60620b --- /dev/null +++ b/scripts/run_1.14G_dp2_tp4_pp1_acc64_mbs16_seq2048_zero1_tpmodeALL_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp2_tp4_pp1_acc64_mbs16_seq2048_zero1_tpmodeALL_vocab32k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=1 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.14G_dp2_tp4_pp1_acc64_mbs16_seq2048_zero1_tpmodeALL_vocab32k.yaml diff --git a/scripts/run_1.14G_dp2_tp4_pp1_acc8_mbs8_seq32768_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp2_tp4_pp1_acc8_mbs8_seq32768_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..029e7e48ab557312f5531a1724041470a434dd14 --- /dev/null +++ b/scripts/run_1.14G_dp2_tp4_pp1_acc8_mbs8_seq32768_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp2_tp4_pp1_acc8_mbs8_seq32768_zero1_tpmodeRED_vocab32k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=1 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.14G_dp2_tp4_pp1_acc8_mbs8_seq32768_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp2_tp64_pp1_acc128_mbs2_seq8192_zero1_tpmodeALL_vocab32k.sh b/scripts/run_1.14G_dp2_tp64_pp1_acc128_mbs2_seq8192_zero1_tpmodeALL_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..144ac08f74f3fe083b7bd3c1e6ad58d3d7061eb4 --- /dev/null +++ b/scripts/run_1.14G_dp2_tp64_pp1_acc128_mbs2_seq8192_zero1_tpmodeALL_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp2_tp64_pp1_acc128_mbs2_seq8192_zero1_tpmodeALL_vocab32k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=16 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.14G_dp2_tp64_pp1_acc128_mbs2_seq8192_zero1_tpmodeALL_vocab32k.yaml diff --git a/scripts/run_1.14G_dp2_tp8_pp1_acc1_mbs256_seq8192_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp2_tp8_pp1_acc1_mbs256_seq8192_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..53211548b8920996a9734a65574b0dc67d146d13 --- /dev/null +++ b/scripts/run_1.14G_dp2_tp8_pp1_acc1_mbs256_seq8192_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp2_tp8_pp1_acc1_mbs256_seq8192_zero1_tpmodeRED_vocab32k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=2 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.14G_dp2_tp8_pp1_acc1_mbs256_seq8192_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp4_tp2_pp1_acc16_mbs8_seq8192_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp4_tp2_pp1_acc16_mbs8_seq8192_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..2d122916be7a37a398216716151de0d1cf9f425c --- /dev/null +++ b/scripts/run_1.14G_dp4_tp2_pp1_acc16_mbs8_seq8192_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp4_tp2_pp1_acc16_mbs8_seq8192_zero1_tpmodeRED_vocab32k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=1 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.14G_dp4_tp2_pp1_acc16_mbs8_seq8192_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp4_tp2_pp1_acc8_mbs1_seq32768_zero1_tpmodeALL_vocab32k.sh b/scripts/run_1.14G_dp4_tp2_pp1_acc8_mbs1_seq32768_zero1_tpmodeALL_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..fcf2ad221fd9d7035edc4e5f273c15597f6e8366 --- /dev/null +++ b/scripts/run_1.14G_dp4_tp2_pp1_acc8_mbs1_seq32768_zero1_tpmodeALL_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp4_tp2_pp1_acc8_mbs1_seq32768_zero1_tpmodeALL_vocab32k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=1 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.14G_dp4_tp2_pp1_acc8_mbs1_seq32768_zero1_tpmodeALL_vocab32k.yaml diff --git a/scripts/run_1.14G_dp4_tp8_pp1_acc16_mbs2_seq8192_zero1_tpmodeALL_vocab32k.sh b/scripts/run_1.14G_dp4_tp8_pp1_acc16_mbs2_seq8192_zero1_tpmodeALL_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..a69ec4a587f4a9f72f81abc5a75d9fdcc5e5a776 --- /dev/null +++ b/scripts/run_1.14G_dp4_tp8_pp1_acc16_mbs2_seq8192_zero1_tpmodeALL_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp4_tp8_pp1_acc16_mbs2_seq8192_zero1_tpmodeALL_vocab32k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=4 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.14G_dp4_tp8_pp1_acc16_mbs2_seq8192_zero1_tpmodeALL_vocab32k.yaml diff --git a/scripts/run_1.14G_dp8_tp1_pp2_acc4_mbs1_seq32768_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp8_tp1_pp2_acc4_mbs1_seq32768_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..ca779a9b8c97bc520632c584c3b351df5e78b6e1 --- /dev/null +++ b/scripts/run_1.14G_dp8_tp1_pp2_acc4_mbs1_seq32768_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp8_tp1_pp2_acc4_mbs1_seq32768_zero1_tpmodeRED_vocab32k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=2 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.14G_dp8_tp1_pp2_acc4_mbs1_seq32768_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp8_tp1_pp2_acc64_mbs1_seq2048_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp8_tp1_pp2_acc64_mbs1_seq2048_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..96231948e2fd974f9d2e9d36cd5096941549b3c1 --- /dev/null +++ b/scripts/run_1.14G_dp8_tp1_pp2_acc64_mbs1_seq2048_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp8_tp1_pp2_acc64_mbs1_seq2048_zero1_tpmodeRED_vocab32k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=2 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.14G_dp8_tp1_pp2_acc64_mbs1_seq2048_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp8_tp2_pp1_acc2_mbs8_seq32768_zero1_tpmodeALL_vocab32k.sh b/scripts/run_1.14G_dp8_tp2_pp1_acc2_mbs8_seq32768_zero1_tpmodeALL_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..151563f49bc4b0c38d399a4fcf0e7e124eb3850a --- /dev/null +++ b/scripts/run_1.14G_dp8_tp2_pp1_acc2_mbs8_seq32768_zero1_tpmodeALL_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp8_tp2_pp1_acc2_mbs8_seq32768_zero1_tpmodeALL_vocab32k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=2 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.14G_dp8_tp2_pp1_acc2_mbs8_seq32768_zero1_tpmodeALL_vocab32k.yaml diff --git a/scripts/run_1.14G_dp8_tp64_pp1_acc16_mbs16_seq2048_zero1_tpmodeALL_vocab32k.sh b/scripts/run_1.14G_dp8_tp64_pp1_acc16_mbs16_seq2048_zero1_tpmodeALL_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..600980df1ab4d35c752386ef5d752e86866332b4 --- /dev/null +++ b/scripts/run_1.14G_dp8_tp64_pp1_acc16_mbs16_seq2048_zero1_tpmodeALL_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp8_tp64_pp1_acc16_mbs16_seq2048_zero1_tpmodeALL_vocab32k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=64 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.14G_dp8_tp64_pp1_acc16_mbs16_seq2048_zero1_tpmodeALL_vocab32k.yaml diff --git a/scripts/run_1.14G_dp8_tp8_pp1_acc4_mbs4_seq32768_zero1_tpmodeALL_vocab32k.sh b/scripts/run_1.14G_dp8_tp8_pp1_acc4_mbs4_seq32768_zero1_tpmodeALL_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..6eff7ab56490a1573b77954840736eed9fc281e2 --- /dev/null +++ b/scripts/run_1.14G_dp8_tp8_pp1_acc4_mbs4_seq32768_zero1_tpmodeALL_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp8_tp8_pp1_acc4_mbs4_seq32768_zero1_tpmodeALL_vocab32k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=8 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.14G_dp8_tp8_pp1_acc4_mbs4_seq32768_zero1_tpmodeALL_vocab32k.yaml diff --git a/scripts/run_1.34G_dp128_tp4_pp1_acc1_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp128_tp4_pp1_acc1_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..6cb01ca4798088699c3013a0da1015641e45d6f8 --- /dev/null +++ b/scripts/run_1.34G_dp128_tp4_pp1_acc1_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,124 @@ +#!/bin/bash +#SBATCH --job-name=bench_1.34G_dp128_tp4_pp1_acc1_mbs2_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=64 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_1.34G_dp128_tp4_pp1_acc1_mbs2_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp128_tp4_pp1_acc1_mbs2_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp128_tp4_pp1_acc1_mbs2_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_1.34G_dp16_tp16_pp1_acc4_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp16_tp16_pp1_acc4_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..c08c7d9cb51cc98b6a1f016a02a440216d147af4 --- /dev/null +++ b/scripts/run_1.34G_dp16_tp16_pp1_acc4_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,124 @@ +#!/bin/bash +#SBATCH --job-name=bench_1.34G_dp16_tp16_pp1_acc4_mbs4_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=32 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_1.34G_dp16_tp16_pp1_acc4_mbs4_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp16_tp16_pp1_acc4_mbs4_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp16_tp16_pp1_acc4_mbs4_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_1.34G_dp16_tp16_pp1_acc8_mbs2_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp16_tp16_pp1_acc8_mbs2_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..c24d096afe1c706cf4528534fa5a56c46351e367 --- /dev/null +++ b/scripts/run_1.34G_dp16_tp16_pp1_acc8_mbs2_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,124 @@ +#!/bin/bash +#SBATCH --job-name=bench_1.34G_dp16_tp16_pp1_acc8_mbs2_seq4096_zero0_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=32 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_1.34G_dp16_tp16_pp1_acc8_mbs2_seq4096_zero0_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp16_tp16_pp1_acc8_mbs2_seq4096_zero0_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp16_tp16_pp1_acc8_mbs2_seq4096_zero0_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_1.34G_dp16_tp1_pp1_acc2_mbs16_seq2048_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp16_tp1_pp1_acc2_mbs16_seq2048_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..cc83e63c3edc2808f7ed2c707a336d16abc6117b --- /dev/null +++ b/scripts/run_1.34G_dp16_tp1_pp1_acc2_mbs16_seq2048_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp16_tp1_pp1_acc2_mbs16_seq2048_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=2 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp16_tp1_pp1_acc2_mbs16_seq2048_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp16_tp2_pp1_acc128_mbs1_seq2048_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp16_tp2_pp1_acc128_mbs1_seq2048_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..6b55b948953ea994e0e568e52fcf2f8721febbdb --- /dev/null +++ b/scripts/run_1.34G_dp16_tp2_pp1_acc128_mbs1_seq2048_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp16_tp2_pp1_acc128_mbs1_seq2048_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=4 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp16_tp2_pp1_acc128_mbs1_seq2048_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp16_tp8_pp1_acc64_mbs2_seq2048_zero1_tpmodeALL_vocab131k.sh b/scripts/run_1.34G_dp16_tp8_pp1_acc64_mbs2_seq2048_zero1_tpmodeALL_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..0962828adeb2fc61bcc14f89607377c07fe21a29 --- /dev/null +++ b/scripts/run_1.34G_dp16_tp8_pp1_acc64_mbs2_seq2048_zero1_tpmodeALL_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp16_tp8_pp1_acc64_mbs2_seq2048_zero1_tpmodeALL_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=16 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp16_tp8_pp1_acc64_mbs2_seq2048_zero1_tpmodeALL_vocab131k.yaml diff --git a/scripts/run_1.34G_dp1_tp16_pp1_acc1_mbs256_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp1_tp16_pp1_acc1_mbs256_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..db0e8d4948edbba97914af3b1cecc16c543b54a2 --- /dev/null +++ b/scripts/run_1.34G_dp1_tp16_pp1_acc1_mbs256_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,124 @@ +#!/bin/bash +#SBATCH --job-name=bench_1.34G_dp1_tp16_pp1_acc1_mbs256_seq4096_zero0_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=2 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_1.34G_dp1_tp16_pp1_acc1_mbs256_seq4096_zero0_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp1_tp16_pp1_acc1_mbs256_seq4096_zero0_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp1_tp16_pp1_acc1_mbs256_seq4096_zero0_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_1.34G_dp1_tp8_pp1_acc256_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp1_tp8_pp1_acc256_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..6a7886dc25e5e65dd3542f62358fee2440c51e9f --- /dev/null +++ b/scripts/run_1.34G_dp1_tp8_pp1_acc256_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,96 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp1_tp8_pp1_acc256_mbs1_seq4096_zero0_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=1 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=1 + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then + # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else + # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +# Disable wandb +export WANDB_MODE=disabled + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=00:10:00 --job-name=bench_1.34G_dp1_tp8_pp1_acc256_mbs1_seq4096_zero0_tpmodeRED_vocab131k" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp1_tp8_pp1_acc256_mbs1_seq4096_zero0_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp2_tp128_pp1_acc32_mbs2_seq32768_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp2_tp128_pp1_acc32_mbs2_seq32768_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..9c0b4e49fc574a8aa9c06e6fd18db686560eb400 --- /dev/null +++ b/scripts/run_1.34G_dp2_tp128_pp1_acc32_mbs2_seq32768_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp2_tp128_pp1_acc32_mbs2_seq32768_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=32 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp2_tp128_pp1_acc32_mbs2_seq32768_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp2_tp16_pp1_acc1_mbs64_seq8192_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp2_tp16_pp1_acc1_mbs64_seq8192_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..8dfb8f2cb30ee721fd114f1afdf753b097666e28 --- /dev/null +++ b/scripts/run_1.34G_dp2_tp16_pp1_acc1_mbs64_seq8192_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp2_tp16_pp1_acc1_mbs64_seq8192_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=4 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp2_tp16_pp1_acc1_mbs64_seq8192_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp2_tp16_pp1_acc2_mbs8_seq32768_zero1_tpmodeALL_vocab131k.sh b/scripts/run_1.34G_dp2_tp16_pp1_acc2_mbs8_seq32768_zero1_tpmodeALL_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..58c30bbc7d180aa95df641afece0dcceb20c7ff9 --- /dev/null +++ b/scripts/run_1.34G_dp2_tp16_pp1_acc2_mbs8_seq32768_zero1_tpmodeALL_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp2_tp16_pp1_acc2_mbs8_seq32768_zero1_tpmodeALL_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=4 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp2_tp16_pp1_acc2_mbs8_seq32768_zero1_tpmodeALL_vocab131k.yaml diff --git a/scripts/run_1.34G_dp2_tp16_pp1_acc4_mbs256_seq2048_zero1_tpmodeALL_vocab131k.sh b/scripts/run_1.34G_dp2_tp16_pp1_acc4_mbs256_seq2048_zero1_tpmodeALL_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..8ddaf2130d70f6af1a4176ac527c86f2f766a223 --- /dev/null +++ b/scripts/run_1.34G_dp2_tp16_pp1_acc4_mbs256_seq2048_zero1_tpmodeALL_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp2_tp16_pp1_acc4_mbs256_seq2048_zero1_tpmodeALL_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=4 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp2_tp16_pp1_acc4_mbs256_seq2048_zero1_tpmodeALL_vocab131k.yaml diff --git a/scripts/run_1.34G_dp2_tp1_pp16_acc128_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp2_tp1_pp16_acc128_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..22a52c936c5a614d9686e97bcbc38ad729800e37 --- /dev/null +++ b/scripts/run_1.34G_dp2_tp1_pp16_acc128_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_1.34G_dp2_tp1_pp16_acc128_mbs1_seq4096_zero0_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:40:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=4 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_1.34G_dp2_tp1_pp16_acc128_mbs1_seq4096_zero0_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp2_tp1_pp16_acc128_mbs1_seq4096_zero0_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp2_tp1_pp16_acc128_mbs1_seq4096_zero0_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_1.34G_dp2_tp256_pp1_acc1_mbs256_seq8192_zero1_tpmodeALL_vocab131k.sh b/scripts/run_1.34G_dp2_tp256_pp1_acc1_mbs256_seq8192_zero1_tpmodeALL_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..4c9bcb2a2e9bffab4f83fe7daec715927d544449 --- /dev/null +++ b/scripts/run_1.34G_dp2_tp256_pp1_acc1_mbs256_seq8192_zero1_tpmodeALL_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp2_tp256_pp1_acc1_mbs256_seq8192_zero1_tpmodeALL_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=64 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp2_tp256_pp1_acc1_mbs256_seq8192_zero1_tpmodeALL_vocab131k.yaml diff --git a/scripts/run_1.34G_dp2_tp256_pp1_acc4_mbs64_seq2048_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp2_tp256_pp1_acc4_mbs64_seq2048_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..b5c0cfcb546ab7994a459db368a146614ae0e8b3 --- /dev/null +++ b/scripts/run_1.34G_dp2_tp256_pp1_acc4_mbs64_seq2048_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp2_tp256_pp1_acc4_mbs64_seq2048_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=64 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp2_tp256_pp1_acc4_mbs64_seq2048_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp2_tp256_pp1_acc8_mbs32_seq8192_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp2_tp256_pp1_acc8_mbs32_seq8192_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..e35ae674904c0b72530c9533361a1352bb4586a5 --- /dev/null +++ b/scripts/run_1.34G_dp2_tp256_pp1_acc8_mbs32_seq8192_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp2_tp256_pp1_acc8_mbs32_seq8192_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=64 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp2_tp256_pp1_acc8_mbs32_seq8192_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp2_tp8_pp32_acc64_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp2_tp8_pp32_acc64_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..20b4d644b74e5d87c1b71fefadf937efae70a2b5 --- /dev/null +++ b/scripts/run_1.34G_dp2_tp8_pp32_acc64_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_1.34G_dp2_tp8_pp32_acc64_mbs2_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:40:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=64 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_1.34G_dp2_tp8_pp32_acc64_mbs2_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp2_tp8_pp32_acc64_mbs2_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp2_tp8_pp32_acc64_mbs2_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_1.34G_dp4_tp16_pp2_acc8_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp4_tp16_pp2_acc8_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..e4e4152ab3247395d11c958da4165300de077b9a --- /dev/null +++ b/scripts/run_1.34G_dp4_tp16_pp2_acc8_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_1.34G_dp4_tp16_pp2_acc8_mbs8_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=16 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=INFO # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_1.34G_dp4_tp16_pp2_acc8_mbs8_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp4_tp16_pp2_acc8_mbs8_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp4_tp16_pp2_acc8_mbs8_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_1.34G_dp4_tp1_pp2_acc128_mbs4_seq2048_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp4_tp1_pp2_acc128_mbs4_seq2048_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..9a4cca08329d598f678cf823f36f90e83273db24 --- /dev/null +++ b/scripts/run_1.34G_dp4_tp1_pp2_acc128_mbs4_seq2048_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp4_tp1_pp2_acc128_mbs4_seq2048_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=1 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp4_tp1_pp2_acc128_mbs4_seq2048_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp64_tp1_pp2_acc1_mbs2_seq32768_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp64_tp1_pp2_acc1_mbs2_seq32768_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..c6bb9b64b033ed491b2e5dfd3d818e4ae369a364 --- /dev/null +++ b/scripts/run_1.34G_dp64_tp1_pp2_acc1_mbs2_seq32768_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp64_tp1_pp2_acc1_mbs2_seq32768_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=16 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp64_tp1_pp2_acc1_mbs2_seq32768_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp64_tp1_pp2_acc2_mbs1_seq32768_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp64_tp1_pp2_acc2_mbs1_seq32768_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..72f305cbd63dc4d7a14974037ba37fa686ddf9bf --- /dev/null +++ b/scripts/run_1.34G_dp64_tp1_pp2_acc2_mbs1_seq32768_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp64_tp1_pp2_acc2_mbs1_seq32768_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=16 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp64_tp1_pp2_acc2_mbs1_seq32768_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp64_tp2_pp1_acc1_mbs4_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp64_tp2_pp1_acc1_mbs4_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..13185e484af3264a97851ec385926112647ce69b --- /dev/null +++ b/scripts/run_1.34G_dp64_tp2_pp1_acc1_mbs4_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,161 @@ +#!/bin/bash +#SBATCH --job-name=bench_1.34G_dp64_tp2_pp1_acc1_mbs4_seq4096_zero0_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:40:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=normal + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=16 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e +echo "Running script: $0" + + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_1.34G_dp64_tp2_pp1_acc1_mbs4_seq4096_zero0_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp64_tp2_pp1_acc1_mbs4_seq4096_zero0_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp64_tp2_pp1_acc1_mbs4_seq4096_zero0_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_1.34G_dp64_tp2_pp1_acc2_mbs16_seq2048_zero1_tpmodeALL_vocab131k.sh b/scripts/run_1.34G_dp64_tp2_pp1_acc2_mbs16_seq2048_zero1_tpmodeALL_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..c8ef848de9319ad711b8e42117167f3f8be27ad6 --- /dev/null +++ b/scripts/run_1.34G_dp64_tp2_pp1_acc2_mbs16_seq2048_zero1_tpmodeALL_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp64_tp2_pp1_acc2_mbs16_seq2048_zero1_tpmodeALL_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=16 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp64_tp2_pp1_acc2_mbs16_seq2048_zero1_tpmodeALL_vocab131k.yaml diff --git a/scripts/run_1.34G_dp64_tp8_pp1_acc2_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp64_tp8_pp1_acc2_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..2303b389d1a49502284e3716f7d68b3103406b9c --- /dev/null +++ b/scripts/run_1.34G_dp64_tp8_pp1_acc2_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,124 @@ +#!/bin/bash +#SBATCH --job-name=bench_1.34G_dp64_tp8_pp1_acc2_mbs2_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=64 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_1.34G_dp64_tp8_pp1_acc2_mbs2_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp64_tp8_pp1_acc2_mbs2_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp64_tp8_pp1_acc2_mbs2_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_1.34G_dp8_tp1_pp1_acc8_mbs32_seq2048_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp8_tp1_pp1_acc8_mbs32_seq2048_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..2de4ae09cf790d61b11973d82999c1062af63789 --- /dev/null +++ b/scripts/run_1.34G_dp8_tp1_pp1_acc8_mbs32_seq2048_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp8_tp1_pp1_acc8_mbs32_seq2048_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=1 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp8_tp1_pp1_acc8_mbs32_seq2048_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp8_tp1_pp2_acc4_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp8_tp1_pp2_acc4_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..4ab7bc983736aca4acb1f0352369268c5e3469e8 --- /dev/null +++ b/scripts/run_1.34G_dp8_tp1_pp2_acc4_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_1.34G_dp8_tp1_pp2_acc4_mbs8_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=2 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=INFO # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_1.34G_dp8_tp1_pp2_acc4_mbs8_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp8_tp1_pp2_acc4_mbs8_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp8_tp1_pp2_acc4_mbs8_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_1.34G_dp8_tp2_pp1_acc4_mbs16_seq8192_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp8_tp2_pp1_acc4_mbs16_seq8192_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..c9e80363bef2aed7b8dc0c0dd7f7d445124a2b0b --- /dev/null +++ b/scripts/run_1.34G_dp8_tp2_pp1_acc4_mbs16_seq8192_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp8_tp2_pp1_acc4_mbs16_seq8192_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=2 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp8_tp2_pp1_acc4_mbs16_seq8192_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp8_tp2_pp4_acc8_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp8_tp2_pp4_acc8_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..a10eab2d7c1fadf83471c1edf6248e0ab3d3e0b4 --- /dev/null +++ b/scripts/run_1.34G_dp8_tp2_pp4_acc8_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_1.34G_dp8_tp2_pp4_acc8_mbs4_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=8 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=INFO # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_1.34G_dp8_tp2_pp4_acc8_mbs4_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp8_tp2_pp4_acc8_mbs4_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp8_tp2_pp4_acc8_mbs4_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_1.34G_dp8_tp32_pp1_acc32_mbs2_seq8192_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp8_tp32_pp1_acc32_mbs2_seq8192_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..ece2a09b423f9b48549463fde863591cfca2a58b --- /dev/null +++ b/scripts/run_1.34G_dp8_tp32_pp1_acc32_mbs2_seq8192_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp8_tp32_pp1_acc32_mbs2_seq8192_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=32 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp8_tp32_pp1_acc32_mbs2_seq8192_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp8_tp64_pp1_acc8_mbs32_seq2048_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp8_tp64_pp1_acc8_mbs32_seq2048_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..ccf12edf5a133ec019320750882ed58d99196518 --- /dev/null +++ b/scripts/run_1.34G_dp8_tp64_pp1_acc8_mbs32_seq2048_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp8_tp64_pp1_acc8_mbs32_seq2048_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=64 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp8_tp64_pp1_acc8_mbs32_seq2048_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_138M_dp512_tp1_pp1_acc1_mbs2_seq4096_zero0_l12_h768_heads12.sh b/scripts/run_138M_dp512_tp1_pp1_acc1_mbs2_seq4096_zero0_l12_h768_heads12.sh new file mode 100644 index 0000000000000000000000000000000000000000..ab75d330e0bb7e8c80b0ff05fb3d6068de8b5be7 --- /dev/null +++ b/scripts/run_138M_dp512_tp1_pp1_acc1_mbs2_seq4096_zero0_l12_h768_heads12.sh @@ -0,0 +1,57 @@ +#!/bin/bash + +#SBATCH --job-name=bench_138M_dp512_tp1_pp1_acc1_mbs2_seq4096_zero0_l12_h768_heads12 # Job name +#SBATCH --time=00:15:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=64 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_138M_dp512_tp1_pp1_acc1_mbs2_seq4096_zero0_l12_h768_heads12.yaml diff --git a/scripts/run_3.57G_dp1_tp16_pp2_acc16_mbs16_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_3.57G_dp1_tp16_pp2_acc16_mbs16_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..d4895fad07835a19f64871a8e2a00e98fc8368a5 --- /dev/null +++ b/scripts/run_3.57G_dp1_tp16_pp2_acc16_mbs16_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_3.57G_dp1_tp16_pp2_acc16_mbs16_seq4096_zero0_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:40:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=4 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_3.57G_dp1_tp16_pp2_acc16_mbs16_seq4096_zero0_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_3.57G_dp1_tp16_pp2_acc16_mbs16_seq4096_zero0_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_3.57G_dp1_tp16_pp2_acc16_mbs16_seq4096_zero0_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_3.57G_dp2_tp1_pp32_acc32_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_3.57G_dp2_tp1_pp32_acc32_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..41d31bddce24e0ebf93c7d84aa8c72111b9a9f1f --- /dev/null +++ b/scripts/run_3.57G_dp2_tp1_pp32_acc32_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_3.57G_dp2_tp1_pp32_acc32_mbs4_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=8 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=INFO # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_3.57G_dp2_tp1_pp32_acc32_mbs4_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_3.57G_dp2_tp1_pp32_acc32_mbs4_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_3.57G_dp2_tp1_pp32_acc32_mbs4_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_3.57G_dp32_tp4_pp1_acc2_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_3.57G_dp32_tp4_pp1_acc2_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..eb9acc977d41accdc23b96d5c5d90dcf3ab82c0a --- /dev/null +++ b/scripts/run_3.57G_dp32_tp4_pp1_acc2_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,124 @@ +#!/bin/bash +#SBATCH --job-name=bench_3.57G_dp32_tp4_pp1_acc2_mbs4_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=16 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_3.57G_dp32_tp4_pp1_acc2_mbs4_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_3.57G_dp32_tp4_pp1_acc2_mbs4_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_3.57G_dp32_tp4_pp1_acc2_mbs4_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_3.57G_dp32_tp8_pp1_acc2_mbs64_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_3.57G_dp32_tp8_pp1_acc2_mbs64_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..c5e9436c25023417bbb5f4ca7691ee0a57652416 --- /dev/null +++ b/scripts/run_3.57G_dp32_tp8_pp1_acc2_mbs64_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,161 @@ +#!/bin/bash +#SBATCH --job-name=bench_3.57G_dp32_tp8_pp1_acc2_mbs64_seq4096_zero0_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:40:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=normal + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=32 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e +echo "Running script: $0" + + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_3.57G_dp32_tp8_pp1_acc2_mbs64_seq4096_zero0_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_3.57G_dp32_tp8_pp1_acc2_mbs64_seq4096_zero0_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_3.57G_dp32_tp8_pp1_acc2_mbs64_seq4096_zero0_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_3.57G_dp64_tp1_pp2_acc2_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_3.57G_dp64_tp1_pp2_acc2_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..f4b47302c47473d3c0698b8d8d7e5e65650dc1c9 --- /dev/null +++ b/scripts/run_3.57G_dp64_tp1_pp2_acc2_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_3.57G_dp64_tp1_pp2_acc2_mbs2_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=16 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=INFO # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_3.57G_dp64_tp1_pp2_acc2_mbs2_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_3.57G_dp64_tp1_pp2_acc2_mbs2_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_3.57G_dp64_tp1_pp2_acc2_mbs2_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_469G_dp16_tp4_pp2_acc16_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_469G_dp16_tp4_pp2_acc16_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..5c7268d24943d26606365bfc84786293e044bf28 --- /dev/null +++ b/scripts/run_469G_dp16_tp4_pp2_acc16_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_469G_dp16_tp4_pp2_acc16_mbs1_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=16 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=INFO # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_469G_dp16_tp4_pp2_acc16_mbs1_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_469G_dp16_tp4_pp2_acc16_mbs1_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_469G_dp16_tp4_pp2_acc16_mbs1_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_469G_dp1_tp8_pp8_acc256_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_469G_dp1_tp8_pp8_acc256_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..0cdcf519f9da174bf9194df94974f3d41e1c65d8 --- /dev/null +++ b/scripts/run_469G_dp1_tp8_pp8_acc256_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,73 @@ +#!/bin/bash + +#SBATCH --job-name=bench_469G_dp1_tp8_pp8_acc256_mbs1_seq4096_zero0_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high +#SBATCH --exclude=ip-26-0-160-192,ip-26-0-171-102 + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=8 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +# Disable wandb +export WANDB_MODE=disabled + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + + +# Print GPU topology information +echo "=== GPU Topology ===" +nvidia-smi topo -m +echo "==================" + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_469G_dp1_tp8_pp8_acc256_mbs1_seq4096_zero0_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_469G_dp2_tp2_pp2_acc4_mbs32_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_469G_dp2_tp2_pp2_acc4_mbs32_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..acb7fdf819248fea0a7a19de9bc25ed4fda029ea --- /dev/null +++ b/scripts/run_469G_dp2_tp2_pp2_acc4_mbs32_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_469G_dp2_tp2_pp2_acc4_mbs32_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=1 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=INFO # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_469G_dp2_tp2_pp2_acc4_mbs32_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_469G_dp2_tp2_pp2_acc4_mbs32_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_469G_dp2_tp2_pp2_acc4_mbs32_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_469G_dp2_tp2_pp8_acc64_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_469G_dp2_tp2_pp8_acc64_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..d16bbbf8324f7cd32c8538ee43a66adaaa61ef56 --- /dev/null +++ b/scripts/run_469G_dp2_tp2_pp8_acc64_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_469G_dp2_tp2_pp8_acc64_mbs2_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=4 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=INFO # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_469G_dp2_tp2_pp8_acc64_mbs2_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_469G_dp2_tp2_pp8_acc64_mbs2_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_469G_dp2_tp2_pp8_acc64_mbs2_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_469G_dp2_tp4_pp2_acc8_mbs16_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_469G_dp2_tp4_pp2_acc8_mbs16_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..c06f404d79137a5f32209ab77d4f4bb7010fd18f --- /dev/null +++ b/scripts/run_469G_dp2_tp4_pp2_acc8_mbs16_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_469G_dp2_tp4_pp2_acc8_mbs16_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=2 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=INFO # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_469G_dp2_tp4_pp2_acc8_mbs16_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_469G_dp2_tp4_pp2_acc8_mbs16_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_469G_dp2_tp4_pp2_acc8_mbs16_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_469G_dp32_tp1_pp4_acc8_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_469G_dp32_tp1_pp4_acc8_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..1160399b4c06caa49406c7e4dc7e64c7dcdf115b --- /dev/null +++ b/scripts/run_469G_dp32_tp1_pp4_acc8_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_469G_dp32_tp1_pp4_acc8_mbs1_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=16 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=INFO # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_469G_dp32_tp1_pp4_acc8_mbs1_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_469G_dp32_tp1_pp4_acc8_mbs1_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_469G_dp32_tp1_pp4_acc8_mbs1_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_469G_dp8_tp64_pp1_acc1_mbs32_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_469G_dp8_tp64_pp1_acc1_mbs32_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..3e63d7259291de87014eb2bad5c5d5d1210cf692 --- /dev/null +++ b/scripts/run_469G_dp8_tp64_pp1_acc1_mbs32_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_469G_dp8_tp64_pp1_acc1_mbs32_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:40:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=64 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_469G_dp8_tp64_pp1_acc1_mbs32_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_469G_dp8_tp64_pp1_acc1_mbs32_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_469G_dp8_tp64_pp1_acc1_mbs32_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_8.86G_dp16_tp16_pp1_acc2_mbs8_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_8.86G_dp16_tp16_pp1_acc2_mbs8_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..7213afffdb8aa603f114b9129330fd9c5579faf2 --- /dev/null +++ b/scripts/run_8.86G_dp16_tp16_pp1_acc2_mbs8_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,124 @@ +#!/bin/bash +#SBATCH --job-name=bench_8.86G_dp16_tp16_pp1_acc2_mbs8_seq4096_zero0_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=32 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_8.86G_dp16_tp16_pp1_acc2_mbs8_seq4096_zero0_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_8.86G_dp16_tp16_pp1_acc2_mbs8_seq4096_zero0_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_8.86G_dp16_tp16_pp1_acc2_mbs8_seq4096_zero0_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_8.86G_dp4_tp2_pp16_acc16_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_8.86G_dp4_tp2_pp16_acc16_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..660fd63d781254b67faaf4d78f1d1863e219b6ab --- /dev/null +++ b/scripts/run_8.86G_dp4_tp2_pp16_acc16_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_8.86G_dp4_tp2_pp16_acc16_mbs4_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=16 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=INFO # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_8.86G_dp4_tp2_pp16_acc16_mbs4_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_8.86G_dp4_tp2_pp16_acc16_mbs4_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_8.86G_dp4_tp2_pp16_acc16_mbs4_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_8.86G_dp4_tp4_pp1_acc2_mbs32_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_8.86G_dp4_tp4_pp1_acc2_mbs32_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..4e2a700dfae9d9c6e4ff0f0c3a39142515c04049 --- /dev/null +++ b/scripts/run_8.86G_dp4_tp4_pp1_acc2_mbs32_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,124 @@ +#!/bin/bash +#SBATCH --job-name=bench_8.86G_dp4_tp4_pp1_acc2_mbs32_seq4096_zero0_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=2 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_8.86G_dp4_tp4_pp1_acc2_mbs32_seq4096_zero0_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_8.86G_dp4_tp4_pp1_acc2_mbs32_seq4096_zero0_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_8.86G_dp4_tp4_pp1_acc2_mbs32_seq4096_zero0_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_8.86G_dp8_tp8_pp8_acc16_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_8.86G_dp8_tp8_pp8_acc16_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..b0aa6328b21b4b2c3a97833ec4f48aea8bc7904d --- /dev/null +++ b/scripts/run_8.86G_dp8_tp8_pp8_acc16_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_8.86G_dp8_tp8_pp8_acc16_mbs2_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:40:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=64 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_8.86G_dp8_tp8_pp8_acc16_mbs2_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_8.86G_dp8_tp8_pp8_acc16_mbs2_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_8.86G_dp8_tp8_pp8_acc16_mbs2_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_80G_dp128_tp2_pp1_acc1_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_80G_dp128_tp2_pp1_acc1_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..3b03eeffc0280f30d85816db1cacba5c6bf71ded --- /dev/null +++ b/scripts/run_80G_dp128_tp2_pp1_acc1_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,124 @@ +#!/bin/bash +#SBATCH --job-name=bench_80G_dp128_tp2_pp1_acc1_mbs2_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=32 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_80G_dp128_tp2_pp1_acc1_mbs2_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp128_tp2_pp1_acc1_mbs2_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp128_tp2_pp1_acc1_mbs2_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_80G_dp16_tp2_pp16_acc16_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_80G_dp16_tp2_pp16_acc16_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..773f4de3f06289ad63a8fb992686a356279eef8b --- /dev/null +++ b/scripts/run_80G_dp16_tp2_pp16_acc16_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_80G_dp16_tp2_pp16_acc16_mbs1_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:40:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=64 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_80G_dp16_tp2_pp16_acc16_mbs1_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp16_tp2_pp16_acc16_mbs1_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp16_tp2_pp16_acc16_mbs1_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_80G_dp16_tp8_pp1_acc4_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_80G_dp16_tp8_pp1_acc4_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..d6181304854e26cfa7bef2ae0b21910c943df7aa --- /dev/null +++ b/scripts/run_80G_dp16_tp8_pp1_acc4_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,124 @@ +#!/bin/bash +#SBATCH --job-name=bench_80G_dp16_tp8_pp1_acc4_mbs4_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=16 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_80G_dp16_tp8_pp1_acc4_mbs4_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp16_tp8_pp1_acc4_mbs4_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp16_tp8_pp1_acc4_mbs4_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_80G_dp1_tp2_pp16_acc2_mbs128_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_80G_dp1_tp2_pp16_acc2_mbs128_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..d20c5cb306595ac9449c4612f4804aef6ec1d824 --- /dev/null +++ b/scripts/run_80G_dp1_tp2_pp16_acc2_mbs128_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,73 @@ +#!/bin/bash + +#SBATCH --job-name=bench_80G_dp1_tp2_pp16_acc2_mbs128_seq4096_zero0_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high +#SBATCH --exclude=ip-26-0-160-192,ip-26-0-171-102 + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=4 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +# Disable wandb +export WANDB_MODE=disabled + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + + +# Print GPU topology information +echo "=== GPU Topology ===" +nvidia-smi topo -m +echo "==================" + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp1_tp2_pp16_acc2_mbs128_seq4096_zero0_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_80G_dp1_tp4_pp8_acc256_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_80G_dp1_tp4_pp8_acc256_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..74ecf5e8d773ab9a8b0cba8f00284d56accf82a3 --- /dev/null +++ b/scripts/run_80G_dp1_tp4_pp8_acc256_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,161 @@ +#!/bin/bash +#SBATCH --job-name=bench_80G_dp1_tp4_pp8_acc256_mbs1_seq4096_zero0_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:40:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=normal + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=4 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e +echo "Running script: $0" + + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_80G_dp1_tp4_pp8_acc256_mbs1_seq4096_zero0_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp1_tp4_pp8_acc256_mbs1_seq4096_zero0_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp1_tp4_pp8_acc256_mbs1_seq4096_zero0_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_80G_dp2_tp16_pp4_acc8_mbs16_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_80G_dp2_tp16_pp4_acc8_mbs16_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..710a7299b92dec360ed8c1b581a806623e2f8d22 --- /dev/null +++ b/scripts/run_80G_dp2_tp16_pp4_acc8_mbs16_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_80G_dp2_tp16_pp4_acc8_mbs16_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=16 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=INFO # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_80G_dp2_tp16_pp4_acc8_mbs16_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp2_tp16_pp4_acc8_mbs16_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp2_tp16_pp4_acc8_mbs16_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_80G_dp2_tp16_pp8_acc128_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_80G_dp2_tp16_pp8_acc128_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..6949d8d6c12e2525e1088c1cdda0ca5a181382e1 --- /dev/null +++ b/scripts/run_80G_dp2_tp16_pp8_acc128_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_80G_dp2_tp16_pp8_acc128_mbs1_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=32 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=INFO # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_80G_dp2_tp16_pp8_acc128_mbs1_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp2_tp16_pp8_acc128_mbs1_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp2_tp16_pp8_acc128_mbs1_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_80G_dp2_tp1_pp16_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_80G_dp2_tp1_pp16_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..99b5b346c3ea93d3bf52879b96728798596a4b80 --- /dev/null +++ b/scripts/run_80G_dp2_tp1_pp16_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_80G_dp2_tp1_pp16_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=4 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=INFO # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_80G_dp2_tp1_pp16_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp2_tp1_pp16_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp2_tp1_pp16_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_80G_dp2_tp2_pp32_acc128_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_80G_dp2_tp2_pp32_acc128_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..d75b6d438ca48009e3d3a04f3086fb840f6c4e65 --- /dev/null +++ b/scripts/run_80G_dp2_tp2_pp32_acc128_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_80G_dp2_tp2_pp32_acc128_mbs1_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=16 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=INFO # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_80G_dp2_tp2_pp32_acc128_mbs1_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp2_tp2_pp32_acc128_mbs1_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp2_tp2_pp32_acc128_mbs1_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_80G_dp2_tp2_pp4_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_80G_dp2_tp2_pp4_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..4279c11d125308bc55549839788d6775406eb142 --- /dev/null +++ b/scripts/run_80G_dp2_tp2_pp4_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_80G_dp2_tp2_pp4_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=2 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=INFO # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_80G_dp2_tp2_pp4_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp2_tp2_pp4_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp2_tp2_pp4_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_80G_dp32_tp1_pp4_acc8_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_80G_dp32_tp1_pp4_acc8_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..a43c9af7f729258c7f3e1b62a59c5d3f4ff74b87 --- /dev/null +++ b/scripts/run_80G_dp32_tp1_pp4_acc8_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_80G_dp32_tp1_pp4_acc8_mbs1_seq4096_zero0_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:40:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=16 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_80G_dp32_tp1_pp4_acc8_mbs1_seq4096_zero0_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp32_tp1_pp4_acc8_mbs1_seq4096_zero0_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp32_tp1_pp4_acc8_mbs1_seq4096_zero0_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_80G_dp8_tp4_pp2_acc2_mbs16_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_80G_dp8_tp4_pp2_acc2_mbs16_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..2dd7d8b4ea4c0b4b4402e1415993060b4440596a --- /dev/null +++ b/scripts/run_80G_dp8_tp4_pp2_acc2_mbs16_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_80G_dp8_tp4_pp2_acc2_mbs16_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=8 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=INFO # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_80G_dp8_tp4_pp2_acc2_mbs16_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp8_tp4_pp2_acc2_mbs16_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp8_tp4_pp2_acc2_mbs16_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi