diff --git a/scripts/run_1.14G_dp16_tp16_pp1_acc8_mbs16_seq2048_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp16_tp16_pp1_acc8_mbs16_seq2048_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..9766a49207d03f528a8d1826012c7bc326f6800d --- /dev/null +++ b/scripts/run_1.14G_dp16_tp16_pp1_acc8_mbs16_seq2048_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp16_tp16_pp1_acc8_mbs16_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_dp16_tp16_pp1_acc8_mbs16_seq2048_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp256_tp1_pp2_acc1_mbs2_seq8192_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp256_tp1_pp2_acc1_mbs2_seq8192_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..e069bac8b2ee3ae101c39fdb61e667a9536755ad --- /dev/null +++ b/scripts/run_1.14G_dp256_tp1_pp2_acc1_mbs2_seq8192_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp256_tp1_pp2_acc1_mbs2_seq8192_zero1_tpmodeRED_vocab32k # 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 + +# # 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_dp256_tp1_pp2_acc1_mbs2_seq8192_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp2_tp16_pp1_acc1_mbs16_seq32768_zero1_tpmodeALL_vocab32k.sh b/scripts/run_1.14G_dp2_tp16_pp1_acc1_mbs16_seq32768_zero1_tpmodeALL_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..943da5939d5da4c66fd156931527208a767405fa --- /dev/null +++ b/scripts/run_1.14G_dp2_tp16_pp1_acc1_mbs16_seq32768_zero1_tpmodeALL_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp2_tp16_pp1_acc1_mbs16_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=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_acc1_mbs16_seq32768_zero1_tpmodeALL_vocab32k.yaml diff --git a/scripts/run_1.14G_dp2_tp256_pp1_acc2_mbs8_seq32768_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp2_tp256_pp1_acc2_mbs8_seq32768_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..4893e3eff03fa02d49f55db8b610a7db40cc4c43 --- /dev/null +++ b/scripts/run_1.14G_dp2_tp256_pp1_acc2_mbs8_seq32768_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp2_tp256_pp1_acc2_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=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_dp2_tp256_pp1_acc2_mbs8_seq32768_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp2_tp4_pp1_acc32_mbs8_seq8192_zero1_tpmodeALL_vocab32k.sh b/scripts/run_1.14G_dp2_tp4_pp1_acc32_mbs8_seq8192_zero1_tpmodeALL_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..26898bdb7b7a2fa704541b5f4cea4afc5909b217 --- /dev/null +++ b/scripts/run_1.14G_dp2_tp4_pp1_acc32_mbs8_seq8192_zero1_tpmodeALL_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp2_tp4_pp1_acc32_mbs8_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=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_acc32_mbs8_seq8192_zero1_tpmodeALL_vocab32k.yaml diff --git a/scripts/run_1.14G_dp2_tp64_pp1_acc16_mbs64_seq2048_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp2_tp64_pp1_acc16_mbs64_seq2048_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..cc2f51abae90617b46ea7bfdd2a52a66720fe3fa --- /dev/null +++ b/scripts/run_1.14G_dp2_tp64_pp1_acc16_mbs64_seq2048_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp2_tp64_pp1_acc16_mbs64_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=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_acc16_mbs64_seq2048_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp2_tp64_pp1_acc4_mbs64_seq2048_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp2_tp64_pp1_acc4_mbs64_seq2048_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..2409020f970b322decddac5c4108e4096f59a9f0 --- /dev/null +++ b/scripts/run_1.14G_dp2_tp64_pp1_acc4_mbs64_seq2048_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp2_tp64_pp1_acc4_mbs64_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=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_acc4_mbs64_seq2048_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp2_tp8_pp1_acc16_mbs16_seq2048_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp2_tp8_pp1_acc16_mbs16_seq2048_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..79e0dd21c079970eea5984ce3a5ff9ca8433a90a --- /dev/null +++ b/scripts/run_1.14G_dp2_tp8_pp1_acc16_mbs16_seq2048_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp2_tp8_pp1_acc16_mbs16_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_dp2_tp8_pp1_acc16_mbs16_seq2048_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp32_tp16_pp1_acc4_mbs1_seq8192_zero1_tpmodeALL_vocab32k.sh b/scripts/run_1.14G_dp32_tp16_pp1_acc4_mbs1_seq8192_zero1_tpmodeALL_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..51940289086798b7d3e7c2702d158bd29db5e53f --- /dev/null +++ b/scripts/run_1.14G_dp32_tp16_pp1_acc4_mbs1_seq8192_zero1_tpmodeALL_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp32_tp16_pp1_acc4_mbs1_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=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_dp32_tp16_pp1_acc4_mbs1_seq8192_zero1_tpmodeALL_vocab32k.yaml diff --git a/scripts/run_1.14G_dp32_tp2_pp1_acc4_mbs4_seq8192_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp32_tp2_pp1_acc4_mbs4_seq8192_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..f6d5a46cfa554ba3cf4034bb157314a9f01d39d3 --- /dev/null +++ b/scripts/run_1.14G_dp32_tp2_pp1_acc4_mbs4_seq8192_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp32_tp2_pp1_acc4_mbs4_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=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_dp32_tp2_pp1_acc4_mbs4_seq8192_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp4_tp128_pp1_acc2_mbs64_seq8192_zero1_tpmodeALL_vocab32k.sh b/scripts/run_1.14G_dp4_tp128_pp1_acc2_mbs64_seq8192_zero1_tpmodeALL_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..86dd604dd16acc5f3077c8d640f2a10fef449adf --- /dev/null +++ b/scripts/run_1.14G_dp4_tp128_pp1_acc2_mbs64_seq8192_zero1_tpmodeALL_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp4_tp128_pp1_acc2_mbs64_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=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_dp4_tp128_pp1_acc2_mbs64_seq8192_zero1_tpmodeALL_vocab32k.yaml diff --git a/scripts/run_1.14G_dp4_tp128_pp1_acc4_mbs8_seq8192_zero1_tpmodeALL_vocab32k.sh b/scripts/run_1.14G_dp4_tp128_pp1_acc4_mbs8_seq8192_zero1_tpmodeALL_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..1c58f20430521f82db3bff2a6ace977b967a43f0 --- /dev/null +++ b/scripts/run_1.14G_dp4_tp128_pp1_acc4_mbs8_seq8192_zero1_tpmodeALL_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp4_tp128_pp1_acc4_mbs8_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=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_dp4_tp128_pp1_acc4_mbs8_seq8192_zero1_tpmodeALL_vocab32k.yaml diff --git a/scripts/run_1.14G_dp4_tp1_pp2_acc2_mbs4_seq32768_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp4_tp1_pp2_acc2_mbs4_seq32768_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..9023eb8357989036b5bda99d765621c448cd50fc --- /dev/null +++ b/scripts/run_1.14G_dp4_tp1_pp2_acc2_mbs4_seq32768_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp4_tp1_pp2_acc2_mbs4_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_dp4_tp1_pp2_acc2_mbs4_seq32768_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp4_tp2_pp1_acc4_mbs8_seq8192_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp4_tp2_pp1_acc4_mbs8_seq8192_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..1fac3195f1f7d333b7641bc47cfce7fe935cc03a --- /dev/null +++ b/scripts/run_1.14G_dp4_tp2_pp1_acc4_mbs8_seq8192_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp4_tp2_pp1_acc4_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_acc4_mbs8_seq8192_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp4_tp32_pp1_acc16_mbs8_seq8192_zero1_tpmodeALL_vocab32k.sh b/scripts/run_1.14G_dp4_tp32_pp1_acc16_mbs8_seq8192_zero1_tpmodeALL_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..f18396e1e7d6c98c8a6ca5c7c903496d0a23eb35 --- /dev/null +++ b/scripts/run_1.14G_dp4_tp32_pp1_acc16_mbs8_seq8192_zero1_tpmodeALL_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp4_tp32_pp1_acc16_mbs8_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_dp4_tp32_pp1_acc16_mbs8_seq8192_zero1_tpmodeALL_vocab32k.yaml diff --git a/scripts/run_1.14G_dp4_tp64_pp1_acc1_mbs32_seq32768_zero1_tpmodeALL_vocab32k.sh b/scripts/run_1.14G_dp4_tp64_pp1_acc1_mbs32_seq32768_zero1_tpmodeALL_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..1685da58a52771e111e2d6ba2b8de1a8394dc78c --- /dev/null +++ b/scripts/run_1.14G_dp4_tp64_pp1_acc1_mbs32_seq32768_zero1_tpmodeALL_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp4_tp64_pp1_acc1_mbs32_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_dp4_tp64_pp1_acc1_mbs32_seq32768_zero1_tpmodeALL_vocab32k.yaml diff --git a/scripts/run_1.14G_dp64_tp4_pp1_acc2_mbs1_seq32768_zero1_tpmodeALL_vocab32k.sh b/scripts/run_1.14G_dp64_tp4_pp1_acc2_mbs1_seq32768_zero1_tpmodeALL_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..f582e468a8897686e6324444e8318795843fe7d8 --- /dev/null +++ b/scripts/run_1.14G_dp64_tp4_pp1_acc2_mbs1_seq32768_zero1_tpmodeALL_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp64_tp4_pp1_acc2_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=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_dp64_tp4_pp1_acc2_mbs1_seq32768_zero1_tpmodeALL_vocab32k.yaml diff --git a/scripts/run_1.14G_dp8_tp16_pp1_acc8_mbs2_seq8192_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp8_tp16_pp1_acc8_mbs2_seq8192_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..cb2d80a21044486cc19819685b4b0e544fc517ac --- /dev/null +++ b/scripts/run_1.14G_dp8_tp16_pp1_acc8_mbs2_seq8192_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp8_tp16_pp1_acc8_mbs2_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=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_dp8_tp16_pp1_acc8_mbs2_seq8192_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp8_tp32_pp1_acc4_mbs4_seq8192_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp8_tp32_pp1_acc4_mbs4_seq8192_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..cd5c277d898740d07cbc0819863989861578048d --- /dev/null +++ b/scripts/run_1.14G_dp8_tp32_pp1_acc4_mbs4_seq8192_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp8_tp32_pp1_acc4_mbs4_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=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_dp8_tp32_pp1_acc4_mbs4_seq8192_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp8_tp8_pp1_acc1_mbs4_seq32768_zero1_tpmodeALL_vocab32k.sh b/scripts/run_1.14G_dp8_tp8_pp1_acc1_mbs4_seq32768_zero1_tpmodeALL_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..33964725e31cc5085b00687502d0a08e4c8b059b --- /dev/null +++ b/scripts/run_1.14G_dp8_tp8_pp1_acc1_mbs4_seq32768_zero1_tpmodeALL_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp8_tp8_pp1_acc1_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_acc1_mbs4_seq32768_zero1_tpmodeALL_vocab32k.yaml diff --git a/scripts/run_1.14G_dp8_tp8_pp1_acc2_mbs32_seq2048_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp8_tp8_pp1_acc2_mbs32_seq2048_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..f9a5beefe3b8d53c7cc1bfe60c1ee635653dc5a5 --- /dev/null +++ b/scripts/run_1.14G_dp8_tp8_pp1_acc2_mbs32_seq2048_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp8_tp8_pp1_acc2_mbs32_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=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_acc2_mbs32_seq2048_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.34G_dp16_tp16_pp1_acc32_mbs4_seq2048_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp16_tp16_pp1_acc32_mbs4_seq2048_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..753326191a7946f6e852b634491c0c5b5fe27794 --- /dev/null +++ b/scripts/run_1.34G_dp16_tp16_pp1_acc32_mbs4_seq2048_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp16_tp16_pp1_acc32_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=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_dp16_tp16_pp1_acc32_mbs4_seq2048_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp16_tp1_pp2_acc32_mbs4_seq2048_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp16_tp1_pp2_acc32_mbs4_seq2048_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..6120393a5c449ff133cdd394a25187a2175eaf0c --- /dev/null +++ b/scripts/run_1.34G_dp16_tp1_pp2_acc32_mbs4_seq2048_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp16_tp1_pp2_acc32_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=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_tp1_pp2_acc32_mbs4_seq2048_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp16_tp2_pp16_acc16_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp16_tp2_pp16_acc16_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..3f3464b232778936989911e54fcefc17fb1f6a55 --- /dev/null +++ b/scripts/run_1.34G_dp16_tp2_pp16_acc16_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_1.34G_dp16_tp2_pp16_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=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 * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=INFO # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_1.34G_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_1.34G_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_1.34G_dp16_tp2_pp16_acc16_mbs1_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_1.34G_dp16_tp2_pp1_acc16_mbs8_seq2048_zero1_tpmodeALL_vocab131k.sh b/scripts/run_1.34G_dp16_tp2_pp1_acc16_mbs8_seq2048_zero1_tpmodeALL_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..76bb2d2d072ce694571165b535c6c02fc19440e5 --- /dev/null +++ b/scripts/run_1.34G_dp16_tp2_pp1_acc16_mbs8_seq2048_zero1_tpmodeALL_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp16_tp2_pp1_acc16_mbs8_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_dp16_tp2_pp1_acc16_mbs8_seq2048_zero1_tpmodeALL_vocab131k.yaml diff --git a/scripts/run_1.34G_dp16_tp32_pp1_acc64_mbs2_seq2048_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp16_tp32_pp1_acc64_mbs2_seq2048_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..122fca70adb7591fbb2fbc725c94fb1c788ca0c5 --- /dev/null +++ b/scripts/run_1.34G_dp16_tp32_pp1_acc64_mbs2_seq2048_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp16_tp32_pp1_acc64_mbs2_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_dp16_tp32_pp1_acc64_mbs2_seq2048_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp16_tp8_pp1_acc8_mbs1_seq32768_zero1_tpmodeALL_vocab131k.sh b/scripts/run_1.34G_dp16_tp8_pp1_acc8_mbs1_seq32768_zero1_tpmodeALL_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..dde285be14cdd1feabe9531fb9d051f32fd4912d --- /dev/null +++ b/scripts/run_1.34G_dp16_tp8_pp1_acc8_mbs1_seq32768_zero1_tpmodeALL_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp16_tp8_pp1_acc8_mbs1_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=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_acc8_mbs1_seq32768_zero1_tpmodeALL_vocab131k.yaml diff --git a/scripts/run_1.34G_dp1_tp2_pp16_acc2_mbs128_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp1_tp2_pp16_acc2_mbs128_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..818079bcb75d406e9237ceadac3696932c871054 --- /dev/null +++ b/scripts/run_1.34G_dp1_tp2_pp16_acc2_mbs128_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,73 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_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_1.34G_dp1_tp2_pp16_acc2_mbs128_seq4096_zero0_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp256_tp1_pp1_acc1_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp256_tp1_pp1_acc1_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..1da483321882adc30b1b88b94434965dad8c511d --- /dev/null +++ b/scripts/run_1.34G_dp256_tp1_pp1_acc1_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,161 @@ +#!/bin/bash +#SBATCH --job-name=bench_1.34G_dp256_tp1_pp1_acc1_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=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_1.34G_dp256_tp1_pp1_acc1_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_dp256_tp1_pp1_acc1_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_dp256_tp1_pp1_acc1_mbs1_seq4096_zero0_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_1.34G_dp2_tp1_pp8_acc8_mbs16_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp2_tp1_pp8_acc8_mbs16_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..4e43667902bc11431a442ddd867e2e874a56b936 --- /dev/null +++ b/scripts/run_1.34G_dp2_tp1_pp8_acc8_mbs16_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_1.34G_dp2_tp1_pp8_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_1.34G_dp2_tp1_pp8_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_1.34G_dp2_tp1_pp8_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_1.34G_dp2_tp1_pp8_acc8_mbs16_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_1.34G_dp2_tp256_pp1_acc16_mbs4_seq8192_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp2_tp256_pp1_acc16_mbs4_seq8192_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..097ce69227066902ba8050c80322ceab8eee55b1 --- /dev/null +++ b/scripts/run_1.34G_dp2_tp256_pp1_acc16_mbs4_seq8192_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp2_tp256_pp1_acc16_mbs4_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_acc16_mbs4_seq8192_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp2_tp2_pp2_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp2_tp2_pp2_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..13df7f7dd61cebe324bb584adfc287b4026219a7 --- /dev/null +++ b/scripts/run_1.34G_dp2_tp2_pp2_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_1.34G_dp2_tp2_pp2_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=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_1.34G_dp2_tp2_pp2_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_1.34G_dp2_tp2_pp2_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_1.34G_dp2_tp2_pp2_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_1.34G_dp2_tp32_pp1_acc16_mbs4_seq32768_zero1_tpmodeALL_vocab131k.sh b/scripts/run_1.34G_dp2_tp32_pp1_acc16_mbs4_seq32768_zero1_tpmodeALL_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..5d371dc7af964bffe62cd60bdcbd9736e445569a --- /dev/null +++ b/scripts/run_1.34G_dp2_tp32_pp1_acc16_mbs4_seq32768_zero1_tpmodeALL_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp2_tp32_pp1_acc16_mbs4_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=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_1.34G_dp2_tp32_pp1_acc16_mbs4_seq32768_zero1_tpmodeALL_vocab131k.yaml diff --git a/scripts/run_1.34G_dp2_tp4_pp1_acc4_mbs16_seq32768_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp2_tp4_pp1_acc4_mbs16_seq32768_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..6ca1a5e3a34afce8a8654ac147b5b12ff5830c7b --- /dev/null +++ b/scripts/run_1.34G_dp2_tp4_pp1_acc4_mbs16_seq32768_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp2_tp4_pp1_acc4_mbs16_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=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_dp2_tp4_pp1_acc4_mbs16_seq32768_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp32_tp1_pp1_acc16_mbs4_seq2048_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp32_tp1_pp1_acc16_mbs4_seq2048_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..feb9613fe77069e44508739be01c0f911cd8ffb8 --- /dev/null +++ b/scripts/run_1.34G_dp32_tp1_pp1_acc16_mbs4_seq2048_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp32_tp1_pp1_acc16_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=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_dp32_tp1_pp1_acc16_mbs4_seq2048_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp32_tp8_pp1_acc4_mbs16_seq2048_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp32_tp8_pp1_acc4_mbs16_seq2048_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..dbfc9bf8843a0cdb263593d719c7e94cca3dc53c --- /dev/null +++ b/scripts/run_1.34G_dp32_tp8_pp1_acc4_mbs16_seq2048_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp32_tp8_pp1_acc4_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=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_dp32_tp8_pp1_acc4_mbs16_seq2048_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp32_tp8_pp2_acc4_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp32_tp8_pp2_acc4_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..538734a26e905838cb7e64fc43d2e509d6a0bae6 --- /dev/null +++ b/scripts/run_1.34G_dp32_tp8_pp2_acc4_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_1.34G_dp32_tp8_pp2_acc4_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_dp32_tp8_pp2_acc4_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_dp32_tp8_pp2_acc4_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_dp32_tp8_pp2_acc4_mbs2_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_1.34G_dp4_tp128_pp1_acc32_mbs4_seq2048_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp4_tp128_pp1_acc32_mbs4_seq2048_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..7265321a906b7d2f40c2f0b6760d081a42aa5e9b --- /dev/null +++ b/scripts/run_1.34G_dp4_tp128_pp1_acc32_mbs4_seq2048_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp4_tp128_pp1_acc32_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=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_dp4_tp128_pp1_acc32_mbs4_seq2048_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp4_tp32_pp1_acc8_mbs64_seq2048_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp4_tp32_pp1_acc8_mbs64_seq2048_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..2dd25d0ad074f1bcbc0ba142491e0f467c8012e9 --- /dev/null +++ b/scripts/run_1.34G_dp4_tp32_pp1_acc8_mbs64_seq2048_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp4_tp32_pp1_acc8_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=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_dp4_tp32_pp1_acc8_mbs64_seq2048_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp4_tp4_pp1_acc32_mbs4_seq8192_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp4_tp4_pp1_acc32_mbs4_seq8192_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..749fdf464a7c190e827b111a19aace7df86b2ae0 --- /dev/null +++ b/scripts/run_1.34G_dp4_tp4_pp1_acc32_mbs4_seq8192_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp4_tp4_pp1_acc32_mbs4_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_dp4_tp4_pp1_acc32_mbs4_seq8192_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp4_tp64_pp1_acc128_mbs1_seq8192_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp4_tp64_pp1_acc128_mbs1_seq8192_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..f7fde6687b8bae07d4a598a3a255ab132a087d06 --- /dev/null +++ b/scripts/run_1.34G_dp4_tp64_pp1_acc128_mbs1_seq8192_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp4_tp64_pp1_acc128_mbs1_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_dp4_tp64_pp1_acc128_mbs1_seq8192_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp4_tp8_pp1_acc8_mbs64_seq2048_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp4_tp8_pp1_acc8_mbs64_seq2048_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..f52532b1b2e77ef7973f9593c6f542f1fb92359e --- /dev/null +++ b/scripts/run_1.34G_dp4_tp8_pp1_acc8_mbs64_seq2048_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp4_tp8_pp1_acc8_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=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_dp4_tp8_pp1_acc8_mbs64_seq2048_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp8_tp1_pp1_acc16_mbs4_seq2048_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp8_tp1_pp1_acc16_mbs4_seq2048_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..b021bfc3436a47ac12cb7173e246c922b448cf46 --- /dev/null +++ b/scripts/run_1.34G_dp8_tp1_pp1_acc16_mbs4_seq2048_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp8_tp1_pp1_acc16_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_dp8_tp1_pp1_acc16_mbs4_seq2048_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp8_tp1_pp2_acc8_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp8_tp1_pp2_acc8_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..19d176d31744f6f6d904d9883bf630e3cc0139b6 --- /dev/null +++ b/scripts/run_1.34G_dp8_tp1_pp2_acc8_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_1.34G_dp8_tp1_pp2_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=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_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_tp1_pp2_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_tp1_pp2_acc8_mbs4_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_1.34G_dp8_tp32_pp1_acc1_mbs16_seq8192_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp8_tp32_pp1_acc1_mbs16_seq8192_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..5849dd1c523a81d2eb17c0c6034aa2132ceefe18 --- /dev/null +++ b/scripts/run_1.34G_dp8_tp32_pp1_acc1_mbs16_seq8192_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp8_tp32_pp1_acc1_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=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_acc1_mbs16_seq8192_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp8_tp4_pp1_acc16_mbs4_seq8192_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp8_tp4_pp1_acc16_mbs4_seq8192_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..e732b7ed5bb7f68029987b40bb03acc9e6dab372 --- /dev/null +++ b/scripts/run_1.34G_dp8_tp4_pp1_acc16_mbs4_seq8192_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp8_tp4_pp1_acc16_mbs4_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_dp8_tp4_pp1_acc16_mbs4_seq8192_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp8_tp64_pp1_acc16_mbs4_seq2048_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp8_tp64_pp1_acc16_mbs4_seq2048_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..daa79de0cabf1ff53cfdb0c9f9bc1cb552687f4e --- /dev/null +++ b/scripts/run_1.34G_dp8_tp64_pp1_acc16_mbs4_seq2048_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp8_tp64_pp1_acc16_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=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_acc16_mbs4_seq2048_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp8_tp8_pp1_acc4_mbs16_seq2048_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp8_tp8_pp1_acc4_mbs16_seq2048_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..5c6d70ecf7732cc937ff7bac92d170915edb83a3 --- /dev/null +++ b/scripts/run_1.34G_dp8_tp8_pp1_acc4_mbs16_seq2048_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp8_tp8_pp1_acc4_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=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_1.34G_dp8_tp8_pp1_acc4_mbs16_seq2048_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_3.57G_dp1_tp1_pp16_acc15_mbs1_seq4096_zero0_tpmodeRED_vocab131k_cache.sh b/scripts/run_3.57G_dp1_tp1_pp16_acc15_mbs1_seq4096_zero0_tpmodeRED_vocab131k_cache.sh new file mode 100644 index 0000000000000000000000000000000000000000..43533f394ae4de5fc18b421880cb56673db14e14 --- /dev/null +++ b/scripts/run_3.57G_dp1_tp1_pp16_acc15_mbs1_seq4096_zero0_tpmodeRED_vocab131k_cache.sh @@ -0,0 +1,161 @@ +#!/bin/bash +#SBATCH --job-name=bench_3.57G_dp1_tp1_pp16_acc15_mbs1_seq4096_zero0_tpmodeRED_vocab131k_cache # 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=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 +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_dp1_tp1_pp16_acc15_mbs1_seq4096_zero0_tpmodeRED_vocab131k_cache" + 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_tp1_pp16_acc15_mbs1_seq4096_zero0_tpmodeRED_vocab131k_cache.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_tp1_pp16_acc15_mbs1_seq4096_zero0_tpmodeRED_vocab131k_cache.yaml +fi diff --git a/scripts/run_3.57G_dp1_tp1_pp8_acc7_mbs1_seq4096_zero0_tpmodeALL_vocab131k_cache.sh b/scripts/run_3.57G_dp1_tp1_pp8_acc7_mbs1_seq4096_zero0_tpmodeALL_vocab131k_cache.sh new file mode 100644 index 0000000000000000000000000000000000000000..8efcabae57b7c66981583416e95080d0ae81d4b9 --- /dev/null +++ b/scripts/run_3.57G_dp1_tp1_pp8_acc7_mbs1_seq4096_zero0_tpmodeALL_vocab131k_cache.sh @@ -0,0 +1,161 @@ +#!/bin/bash +#SBATCH --job-name=bench_3.57G_dp1_tp1_pp8_acc7_mbs1_seq4096_zero0_tpmodeALL_vocab131k_cache # 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=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 +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_dp1_tp1_pp8_acc7_mbs1_seq4096_zero0_tpmodeALL_vocab131k_cache" + 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_tp1_pp8_acc7_mbs1_seq4096_zero0_tpmodeALL_vocab131k_cache.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_tp1_pp8_acc7_mbs1_seq4096_zero0_tpmodeALL_vocab131k_cache.yaml +fi diff --git a/scripts/run_3.57G_dp1_tp32_pp1_acc1_mbs90_seq4096_zero0_tpmodeALL_vocab131k_cache.sh b/scripts/run_3.57G_dp1_tp32_pp1_acc1_mbs90_seq4096_zero0_tpmodeALL_vocab131k_cache.sh new file mode 100644 index 0000000000000000000000000000000000000000..a4bf543d894b90297e00998787622577bb7fd357 --- /dev/null +++ b/scripts/run_3.57G_dp1_tp32_pp1_acc1_mbs90_seq4096_zero0_tpmodeALL_vocab131k_cache.sh @@ -0,0 +1,161 @@ +#!/bin/bash +#SBATCH --job-name=bench_3.57G_dp1_tp32_pp1_acc1_mbs90_seq4096_zero0_tpmodeALL_vocab131k_cache # 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 +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_dp1_tp32_pp1_acc1_mbs90_seq4096_zero0_tpmodeALL_vocab131k_cache" + 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_tp32_pp1_acc1_mbs90_seq4096_zero0_tpmodeALL_vocab131k_cache.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_tp32_pp1_acc1_mbs90_seq4096_zero0_tpmodeALL_vocab131k_cache.yaml +fi diff --git a/scripts/run_3.57G_dp1_tp8_pp1_acc1_mbs20_seq4096_zero0_tpmodeALL_vocab131k_cache.sh b/scripts/run_3.57G_dp1_tp8_pp1_acc1_mbs20_seq4096_zero0_tpmodeALL_vocab131k_cache.sh new file mode 100644 index 0000000000000000000000000000000000000000..8495a526c63e187f4eee9c1dad10286857707cf0 --- /dev/null +++ b/scripts/run_3.57G_dp1_tp8_pp1_acc1_mbs20_seq4096_zero0_tpmodeALL_vocab131k_cache.sh @@ -0,0 +1,161 @@ +#!/bin/bash +#SBATCH --job-name=bench_3.57G_dp1_tp8_pp1_acc1_mbs20_seq4096_zero0_tpmodeALL_vocab131k_cache # 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=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 +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_dp1_tp8_pp1_acc1_mbs20_seq4096_zero0_tpmodeALL_vocab131k_cache" + 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_tp8_pp1_acc1_mbs20_seq4096_zero0_tpmodeALL_vocab131k_cache.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_tp8_pp1_acc1_mbs20_seq4096_zero0_tpmodeALL_vocab131k_cache.yaml +fi diff --git a/scripts/run_3.57G_dp2_tp16_pp8_acc8_mbs16_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_3.57G_dp2_tp16_pp8_acc8_mbs16_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..c8dc539885648806795fdd6c9aab215697468cde --- /dev/null +++ b/scripts/run_3.57G_dp2_tp16_pp8_acc8_mbs16_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_3.57G_dp2_tp16_pp8_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=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_3.57G_dp2_tp16_pp8_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_3.57G_dp2_tp16_pp8_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_3.57G_dp2_tp16_pp8_acc8_mbs16_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_3.57G_dp32_tp1_pp1_acc4_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_3.57G_dp32_tp1_pp1_acc4_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..b7f7b87c90a207cec67a3e539bd28504b8c22653 --- /dev/null +++ b/scripts/run_3.57G_dp32_tp1_pp1_acc4_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,161 @@ +#!/bin/bash +#SBATCH --job-name=bench_3.57G_dp32_tp1_pp1_acc4_mbs2_seq4096_zero1_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_3.57G_dp32_tp1_pp1_acc4_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_dp32_tp1_pp1_acc4_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_dp32_tp1_pp1_acc4_mbs2_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_3.57G_dp8_tp1_pp2_acc2_mbs16_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_3.57G_dp8_tp1_pp2_acc2_mbs16_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..5073fe4e652e3f280b42e578f0eac1c6ed4fecaa --- /dev/null +++ b/scripts/run_3.57G_dp8_tp1_pp2_acc2_mbs16_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,73 @@ +#!/bin/bash + +#SBATCH --job-name=bench_3.57G_dp8_tp1_pp2_acc2_mbs16_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=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 + +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_3.57G_dp8_tp1_pp2_acc2_mbs16_seq4096_zero0_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_469G_dp1_tp16_pp4_acc8_mbs32_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_469G_dp1_tp16_pp4_acc8_mbs32_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..f6c7e815e61a5617012a1415240463d7ffb771ff --- /dev/null +++ b/scripts/run_469G_dp1_tp16_pp4_acc8_mbs32_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,73 @@ +#!/bin/bash + +#SBATCH --job-name=bench_469G_dp1_tp16_pp4_acc8_mbs32_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_tp16_pp4_acc8_mbs32_seq4096_zero0_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_469G_dp2_tp2_pp8_acc128_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_469G_dp2_tp2_pp8_acc128_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..a3f81efdbe39a9cb34f23a4dbbc9d293e0a5b97e --- /dev/null +++ b/scripts/run_469G_dp2_tp2_pp8_acc128_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_469G_dp2_tp2_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=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_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_469G_dp2_tp2_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_469G_dp2_tp2_pp8_acc128_mbs1_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_469G_dp2_tp8_pp1_acc1_mbs128_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_469G_dp2_tp8_pp1_acc1_mbs128_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..07b16de2e3407052a55fe682aa00e9463df1a4c2 --- /dev/null +++ b/scripts/run_469G_dp2_tp8_pp1_acc1_mbs128_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,71 @@ +#!/bin/bash + +#SBATCH --job-name=bench_469G_dp2_tp8_pp1_acc1_mbs128_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high +#SBATCH --reservation=nouamane_weekend_32 +#SBATCH --exclude=ip-26-0-160-242,ip-26-0-161-[138,142,153,178,221],ip-26-0-162-[46,79,233],ip-26-0-163-[43,58,134,147,158,220,226,236],ip-26-0-164-[0,18,45,75,187,207,236],ip-26-0-165-24,ip-26-0-166-[15,36,125,214,244],ip-26-0-167-[9,51,175,177,217,245],ip-26-0-168-[30,34,52],ip-26-0-169-[86,207,239,247],ip-26-0-170-[132,143,160],ip-26-0-171-[21,56,62,88,168,230,249],ip-26-0-172-[57,73],ip-26-0-173-[121,246],ip-26-0-174-[100,186,196,240],ip-26-0-175-[34,132,165,170,241] + +#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 + +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 + + +# 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_dp2_tp8_pp1_acc1_mbs128_seq4096_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_469G_dp2_tp8_pp8_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_469G_dp2_tp8_pp8_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..9b1d58073c7109e95cb43ce1b7b8fd48e59955a0 --- /dev/null +++ b/scripts/run_469G_dp2_tp8_pp8_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_469G_dp2_tp8_pp8_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=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_dp2_tp8_pp8_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_469G_dp2_tp8_pp8_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_469G_dp2_tp8_pp8_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_469G_dp32_tp2_pp2_acc8_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_469G_dp32_tp2_pp2_acc8_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..086b46b62b9d09671c7456a00a042f3ffb399f5c --- /dev/null +++ b/scripts/run_469G_dp32_tp2_pp2_acc8_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_469G_dp32_tp2_pp2_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_tp2_pp2_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_tp2_pp2_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_tp2_pp2_acc8_mbs1_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_469G_dp8_tp32_pp2_acc2_mbs16_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_469G_dp8_tp32_pp2_acc2_mbs16_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..06b93d871ae988ab7a4622ff863297eae8b6becd --- /dev/null +++ b/scripts/run_469G_dp8_tp32_pp2_acc2_mbs16_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_469G_dp8_tp32_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=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 * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=INFO # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_469G_dp8_tp32_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_469G_dp8_tp32_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_469G_dp8_tp32_pp2_acc2_mbs16_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_5.5G_dp16_tp8_pp4_acc1_mbs1_seq2048_zero0_tpmodeRED_l32_h4096_heads32.sh b/scripts/run_5.5G_dp16_tp8_pp4_acc1_mbs1_seq2048_zero0_tpmodeRED_l32_h4096_heads32.sh new file mode 100644 index 0000000000000000000000000000000000000000..a4e0a68acd6d84c34d672475e5ca8d70a807b347 --- /dev/null +++ b/scripts/run_5.5G_dp16_tp8_pp4_acc1_mbs1_seq2048_zero0_tpmodeRED_l32_h4096_heads32.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_5.5G_dp16_tp8_pp4_acc1_mbs1_seq2048_zero0_tpmodeRED_l32_h4096_heads32 # 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 + +# # 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_5.5G_dp16_tp8_pp4_acc1_mbs1_seq2048_zero0_tpmodeRED_l32_h4096_heads32.yaml diff --git a/scripts/run_8.86G_dp2_tp16_pp16_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_8.86G_dp2_tp16_pp16_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..74d1d46e045f52c9a1e8f7e97d6d89de0bb02e6f --- /dev/null +++ b/scripts/run_8.86G_dp2_tp16_pp16_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_8.86G_dp2_tp16_pp16_acc16_mbs8_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_dp2_tp16_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_8.86G_dp2_tp16_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_8.86G_dp2_tp16_pp16_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_8.86G_dp2_tp8_pp1_acc4_mbs32_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_8.86G_dp2_tp8_pp1_acc4_mbs32_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..738efabf11ab162a088a58540f1b438804a465cb --- /dev/null +++ b/scripts/run_8.86G_dp2_tp8_pp1_acc4_mbs32_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,124 @@ +#!/bin/bash +#SBATCH --job-name=bench_8.86G_dp2_tp8_pp1_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=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_dp2_tp8_pp1_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_8.86G_dp2_tp8_pp1_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_8.86G_dp2_tp8_pp1_acc4_mbs32_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_8.86G_dp32_tp1_pp2_acc1_mbs8_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_8.86G_dp32_tp1_pp2_acc1_mbs8_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..f6a84e1f5a7ea28b5e89dafe923553f0523f1c09 --- /dev/null +++ b/scripts/run_8.86G_dp32_tp1_pp2_acc1_mbs8_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,73 @@ +#!/bin/bash + +#SBATCH --job-name=bench_8.86G_dp32_tp1_pp2_acc1_mbs8_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_8.86G_dp32_tp1_pp2_acc1_mbs8_seq4096_zero0_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_8.86G_dp4_tp2_pp8_acc64_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_8.86G_dp4_tp2_pp8_acc64_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..b28981ade072bf633348cc4ba28acdd154c8d743 --- /dev/null +++ b/scripts/run_8.86G_dp4_tp2_pp8_acc64_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_8.86G_dp4_tp2_pp8_acc64_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=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 * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_8.86G_dp4_tp2_pp8_acc64_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_8.86G_dp4_tp2_pp8_acc64_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_8.86G_dp4_tp2_pp8_acc64_mbs1_seq4096_zero0_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_8.86G_dp8_tp16_pp1_acc1_mbs32_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_8.86G_dp8_tp16_pp1_acc1_mbs32_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..1590e2c9560ff29423b6e4b62924014b90b8b61e --- /dev/null +++ b/scripts/run_8.86G_dp8_tp16_pp1_acc1_mbs32_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,124 @@ +#!/bin/bash +#SBATCH --job-name=bench_8.86G_dp8_tp16_pp1_acc1_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=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_8.86G_dp8_tp16_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_8.86G_dp8_tp16_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_8.86G_dp8_tp16_pp1_acc1_mbs32_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_80G_dp2_tp32_pp2_acc1_mbs128_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_80G_dp2_tp32_pp2_acc1_mbs128_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..00005f2381a2a16c8cf722d15bcce45ddb6759da --- /dev/null +++ b/scripts/run_80G_dp2_tp32_pp2_acc1_mbs128_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_80G_dp2_tp32_pp2_acc1_mbs128_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_tp32_pp2_acc1_mbs128_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_tp32_pp2_acc1_mbs128_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_tp32_pp2_acc1_mbs128_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_80G_dp2_tp32_pp4_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_80G_dp2_tp32_pp4_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..d004317903d95b66ecf9463b81be0a84e8d0c6f7 --- /dev/null +++ b/scripts/run_80G_dp2_tp32_pp4_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_80G_dp2_tp32_pp4_acc16_mbs8_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=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 * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_80G_dp2_tp32_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_tp32_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_tp32_pp4_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_80G_dp2_tp8_pp4_acc64_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_80G_dp2_tp8_pp4_acc64_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..6ea1eebac0f2847a36c7f85f9f09079a793bf64f --- /dev/null +++ b/scripts/run_80G_dp2_tp8_pp4_acc64_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_80G_dp2_tp8_pp4_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=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_dp2_tp8_pp4_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_80G_dp2_tp8_pp4_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_80G_dp2_tp8_pp4_acc64_mbs2_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_80G_dp32_tp2_pp2_acc8_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_80G_dp32_tp2_pp2_acc8_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..24c5be9f9dd2c4cd478cee3dc90f3947aad4237d --- /dev/null +++ b/scripts/run_80G_dp32_tp2_pp2_acc8_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_80G_dp32_tp2_pp2_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_tp2_pp2_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_tp2_pp2_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_tp2_pp2_acc8_mbs1_seq4096_zero0_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_80G_dp4_tp1_pp16_acc64_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_80G_dp4_tp1_pp16_acc64_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..094a4731742a8dd9ddea36c9a1376d0382000b90 --- /dev/null +++ b/scripts/run_80G_dp4_tp1_pp16_acc64_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,73 @@ +#!/bin/bash + +#SBATCH --job-name=bench_80G_dp4_tp1_pp16_acc64_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_80G_dp4_tp1_pp16_acc64_mbs1_seq4096_zero0_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_80G_dp4_tp32_pp2_acc4_mbs16_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_80G_dp4_tp32_pp2_acc4_mbs16_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..6ed5c7f021742960e1e92b8445e531edc6622e64 --- /dev/null +++ b/scripts/run_80G_dp4_tp32_pp2_acc4_mbs16_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_80G_dp4_tp32_pp2_acc4_mbs16_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=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 * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_80G_dp4_tp32_pp2_acc4_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_dp4_tp32_pp2_acc4_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_dp4_tp32_pp2_acc4_mbs16_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_80G_dp4_tp32_pp2_acc8_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_80G_dp4_tp32_pp2_acc8_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..a08493db825b929d9f4cfea8fed1ee286e4ce66d --- /dev/null +++ b/scripts/run_80G_dp4_tp32_pp2_acc8_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_80G_dp4_tp32_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=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_dp4_tp32_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_80G_dp4_tp32_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_80G_dp4_tp32_pp2_acc8_mbs8_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_dp16_tp1_pp1_acc1_seq2048.sh b/scripts/run_dp16_tp1_pp1_acc1_seq2048.sh new file mode 100644 index 0000000000000000000000000000000000000000..376e95b3f953fad213ae71b9dbd9d7a53ff9d089 --- /dev/null +++ b/scripts/run_dp16_tp1_pp1_acc1_seq2048.sh @@ -0,0 +1,57 @@ +#!/bin/bash + +#SBATCH --job-name=bench_dp16_tp1_pp1_acc1_seq2048 # Job name +#SBATCH --time=00:15:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%x-%j.out + +#SBATCH --nodes=2 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=80 # 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_dp16_tp1_pp1_acc1_seq2048.yaml