diff --git a/scripts/run_1.14G_dp128_tp4_pp1_acc1_mbs4_seq2048_zero1_tpmodeALL_vocab32k.sh b/scripts/run_1.14G_dp128_tp4_pp1_acc1_mbs4_seq2048_zero1_tpmodeALL_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..e3bf0879acd2bc2ff4cff14a572911e9dc90cfe9 --- /dev/null +++ b/scripts/run_1.14G_dp128_tp4_pp1_acc1_mbs4_seq2048_zero1_tpmodeALL_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp128_tp4_pp1_acc1_mbs4_seq2048_zero1_tpmodeALL_vocab32k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=64 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.14G_dp128_tp4_pp1_acc1_mbs4_seq2048_zero1_tpmodeALL_vocab32k.yaml diff --git a/scripts/run_1.14G_dp16_tp16_pp1_acc4_mbs2_seq32768_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp16_tp16_pp1_acc4_mbs2_seq32768_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..bf75b9dcd59dd67c6a05ab16583739052d5ef45f --- /dev/null +++ b/scripts/run_1.14G_dp16_tp16_pp1_acc4_mbs2_seq32768_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp16_tp16_pp1_acc4_mbs2_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=32 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.14G_dp16_tp16_pp1_acc4_mbs2_seq32768_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp16_tp2_pp1_acc1_mbs32_seq8192_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp16_tp2_pp1_acc1_mbs32_seq8192_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..c6858df73e3691587c3f07b9488345d0d2dbbc1c --- /dev/null +++ b/scripts/run_1.14G_dp16_tp2_pp1_acc1_mbs32_seq8192_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp16_tp2_pp1_acc1_mbs32_seq8192_zero1_tpmodeRED_vocab32k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=4 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.14G_dp16_tp2_pp1_acc1_mbs32_seq8192_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp2_tp128_pp1_acc128_mbs8_seq2048_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp2_tp128_pp1_acc128_mbs8_seq2048_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..332ec368d56ac7185008602c5baf13cc21ea2478 --- /dev/null +++ b/scripts/run_1.14G_dp2_tp128_pp1_acc128_mbs8_seq2048_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp2_tp128_pp1_acc128_mbs8_seq2048_zero1_tpmodeRED_vocab32k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=32 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.14G_dp2_tp128_pp1_acc128_mbs8_seq2048_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp2_tp128_pp1_acc2_mbs32_seq32768_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp2_tp128_pp1_acc2_mbs32_seq32768_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..e4d688b2a4b15bb34b4a30cbe3355c9f9ca80840 --- /dev/null +++ b/scripts/run_1.14G_dp2_tp128_pp1_acc2_mbs32_seq32768_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp2_tp128_pp1_acc2_mbs32_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=32 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.14G_dp2_tp128_pp1_acc2_mbs32_seq32768_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp2_tp128_pp1_acc64_mbs4_seq8192_zero1_tpmodeALL_vocab32k.sh b/scripts/run_1.14G_dp2_tp128_pp1_acc64_mbs4_seq8192_zero1_tpmodeALL_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..fee7fcd3f5cc8d4efafcad707f10958175311c9c --- /dev/null +++ b/scripts/run_1.14G_dp2_tp128_pp1_acc64_mbs4_seq8192_zero1_tpmodeALL_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp2_tp128_pp1_acc64_mbs4_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=32 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.14G_dp2_tp128_pp1_acc64_mbs4_seq8192_zero1_tpmodeALL_vocab32k.yaml diff --git a/scripts/run_1.14G_dp2_tp8_pp1_acc4_mbs16_seq32768_zero1_tpmodeALL_vocab32k.sh b/scripts/run_1.14G_dp2_tp8_pp1_acc4_mbs16_seq32768_zero1_tpmodeALL_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..d4fbb3fdebb36d8aec25981db8cb7a2231bfca5e --- /dev/null +++ b/scripts/run_1.14G_dp2_tp8_pp1_acc4_mbs16_seq32768_zero1_tpmodeALL_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp2_tp8_pp1_acc4_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=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_acc4_mbs16_seq32768_zero1_tpmodeALL_vocab32k.yaml diff --git a/scripts/run_1.14G_dp32_tp2_pp1_acc16_mbs1_seq8192_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp32_tp2_pp1_acc16_mbs1_seq8192_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..304bd068328bfc9a9da63e48bfca668ebf82112d --- /dev/null +++ b/scripts/run_1.14G_dp32_tp2_pp1_acc16_mbs1_seq8192_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp32_tp2_pp1_acc16_mbs1_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_acc16_mbs1_seq8192_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp32_tp2_pp1_acc1_mbs16_seq2048_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp32_tp2_pp1_acc1_mbs16_seq2048_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..110a832349cdb093fe990995f6d6022661839bab --- /dev/null +++ b/scripts/run_1.14G_dp32_tp2_pp1_acc1_mbs16_seq2048_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp32_tp2_pp1_acc1_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=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_acc1_mbs16_seq2048_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp32_tp2_pp1_acc2_mbs8_seq2048_zero1_tpmodeALL_vocab32k.sh b/scripts/run_1.14G_dp32_tp2_pp1_acc2_mbs8_seq2048_zero1_tpmodeALL_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..9a3535cab79e2efee095aa04b15bc7c0aeff96d7 --- /dev/null +++ b/scripts/run_1.14G_dp32_tp2_pp1_acc2_mbs8_seq2048_zero1_tpmodeALL_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp32_tp2_pp1_acc2_mbs8_seq2048_zero1_tpmodeALL_vocab32k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=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_acc2_mbs8_seq2048_zero1_tpmodeALL_vocab32k.yaml diff --git a/scripts/run_1.14G_dp4_tp16_pp1_acc4_mbs32_seq2048_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp4_tp16_pp1_acc4_mbs32_seq2048_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..ec559c3605213f7fc69558b899d049217478f48b --- /dev/null +++ b/scripts/run_1.14G_dp4_tp16_pp1_acc4_mbs32_seq2048_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp4_tp16_pp1_acc4_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_dp4_tp16_pp1_acc4_mbs32_seq2048_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp4_tp2_pp1_acc2_mbs16_seq32768_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp4_tp2_pp1_acc2_mbs16_seq32768_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..c7733b4dcc1b5ef39b046af114eee334d4808af7 --- /dev/null +++ b/scripts/run_1.14G_dp4_tp2_pp1_acc2_mbs16_seq32768_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp4_tp2_pp1_acc2_mbs16_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_tp2_pp1_acc2_mbs16_seq32768_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp4_tp8_pp1_acc16_mbs8_seq2048_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp4_tp8_pp1_acc16_mbs8_seq2048_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..0d3f4891890d0d9889c602ede9ee3470612a4555 --- /dev/null +++ b/scripts/run_1.14G_dp4_tp8_pp1_acc16_mbs8_seq2048_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp4_tp8_pp1_acc16_mbs8_seq2048_zero1_tpmodeRED_vocab32k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=4 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.14G_dp4_tp8_pp1_acc16_mbs8_seq2048_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp64_tp1_pp2_acc4_mbs2_seq8192_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp64_tp1_pp2_acc4_mbs2_seq8192_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..17b1ec0e8d6101f6b04f83ad177115b94e336b1e --- /dev/null +++ b/scripts/run_1.14G_dp64_tp1_pp2_acc4_mbs2_seq8192_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp64_tp1_pp2_acc4_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_dp64_tp1_pp2_acc4_mbs2_seq8192_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp8_tp1_pp1_acc32_mbs8_seq2048_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp8_tp1_pp1_acc32_mbs8_seq2048_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..1f8c49726d5279000d24daaa79c7a1265ade4ee5 --- /dev/null +++ b/scripts/run_1.14G_dp8_tp1_pp1_acc32_mbs8_seq2048_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp8_tp1_pp1_acc32_mbs8_seq2048_zero1_tpmodeRED_vocab32k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=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_dp8_tp1_pp1_acc32_mbs8_seq2048_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp8_tp2_pp1_acc32_mbs2_seq8192_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp8_tp2_pp1_acc32_mbs2_seq8192_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..af60fa3291c6fe11a3b41e0204d65dc4d1d78aef --- /dev/null +++ b/scripts/run_1.14G_dp8_tp2_pp1_acc32_mbs2_seq8192_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp8_tp2_pp1_acc32_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=2 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.14G_dp8_tp2_pp1_acc32_mbs2_seq8192_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp8_tp32_pp1_acc16_mbs1_seq8192_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp8_tp32_pp1_acc16_mbs1_seq8192_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..a2bea0f061434c68040952f381cac84a876d7885 --- /dev/null +++ b/scripts/run_1.14G_dp8_tp32_pp1_acc16_mbs1_seq8192_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp8_tp32_pp1_acc16_mbs1_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_acc16_mbs1_seq8192_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.34G_dp128_tp2_pp1_acc1_mbs4_seq8192_zero1_tpmodeALL_vocab131k.sh b/scripts/run_1.34G_dp128_tp2_pp1_acc1_mbs4_seq8192_zero1_tpmodeALL_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..9698bdde681342587181da73757bc01d075485ce --- /dev/null +++ b/scripts/run_1.34G_dp128_tp2_pp1_acc1_mbs4_seq8192_zero1_tpmodeALL_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp128_tp2_pp1_acc1_mbs4_seq8192_zero1_tpmodeALL_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=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_dp128_tp2_pp1_acc1_mbs4_seq8192_zero1_tpmodeALL_vocab131k.yaml diff --git a/scripts/run_1.34G_dp16_tp16_pp1_acc16_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp16_tp16_pp1_acc16_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..e4a6b4759e50fce2f8c1732eecf556e421483c39 --- /dev/null +++ b/scripts/run_1.34G_dp16_tp16_pp1_acc16_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,124 @@ +#!/bin/bash +#SBATCH --job-name=bench_1.34G_dp16_tp16_pp1_acc16_mbs1_seq4096_zero0_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=32 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_1.34G_dp16_tp16_pp1_acc16_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_dp16_tp16_pp1_acc16_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_dp16_tp16_pp1_acc16_mbs1_seq4096_zero0_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_1.34G_dp16_tp4_pp1_acc1_mbs32_seq8192_zero1_tpmodeALL_vocab131k.sh b/scripts/run_1.34G_dp16_tp4_pp1_acc1_mbs32_seq8192_zero1_tpmodeALL_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..e54599f729a2e2349b5afa1fd008318a7783c678 --- /dev/null +++ b/scripts/run_1.34G_dp16_tp4_pp1_acc1_mbs32_seq8192_zero1_tpmodeALL_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp16_tp4_pp1_acc1_mbs32_seq8192_zero1_tpmodeALL_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=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_dp16_tp4_pp1_acc1_mbs32_seq8192_zero1_tpmodeALL_vocab131k.yaml diff --git a/scripts/run_1.34G_dp16_tp4_pp2_acc8_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp16_tp4_pp2_acc8_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..b83148ff19578d1b5c4922e62bf9336a453e6c38 --- /dev/null +++ b/scripts/run_1.34G_dp16_tp4_pp2_acc8_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_1.34G_dp16_tp4_pp2_acc8_mbs2_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=16 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=INFO # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_1.34G_dp16_tp4_pp2_acc8_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_dp16_tp4_pp2_acc8_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_dp16_tp4_pp2_acc8_mbs2_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_1.34G_dp1_tp1_pp8_acc1_mbs256_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp1_tp1_pp8_acc1_mbs256_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..789f7b09d8070b2ba602da70fd319e6d8bdf8e71 --- /dev/null +++ b/scripts/run_1.34G_dp1_tp1_pp8_acc1_mbs256_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,73 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp1_tp1_pp8_acc1_mbs256_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=1 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +set -x -e + +# 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_tp1_pp8_acc1_mbs256_seq4096_zero0_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp2_tp256_pp1_acc1_mbs64_seq32768_zero1_tpmodeALL_vocab131k.sh b/scripts/run_1.34G_dp2_tp256_pp1_acc1_mbs64_seq32768_zero1_tpmodeALL_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..f04a398f32be742b26ee554ab7c3b6e3b789ebed --- /dev/null +++ b/scripts/run_1.34G_dp2_tp256_pp1_acc1_mbs64_seq32768_zero1_tpmodeALL_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp2_tp256_pp1_acc1_mbs64_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=64 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp2_tp256_pp1_acc1_mbs64_seq32768_zero1_tpmodeALL_vocab131k.yaml diff --git a/scripts/run_1.34G_dp2_tp2_pp32_acc64_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp2_tp2_pp32_acc64_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..179bd25cf2b57bacb2e943fa82b5dbd7091fada8 --- /dev/null +++ b/scripts/run_1.34G_dp2_tp2_pp32_acc64_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_1.34G_dp2_tp2_pp32_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=16 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=INFO # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_1.34G_dp2_tp2_pp32_acc64_mbs2_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp2_tp2_pp32_acc64_mbs2_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp2_tp2_pp32_acc64_mbs2_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_1.34G_dp2_tp32_pp1_acc8_mbs32_seq8192_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp2_tp32_pp1_acc8_mbs32_seq8192_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..9c7e7bf8f77a4e85c41e8d61d11e9ded1e5b6a5f --- /dev/null +++ b/scripts/run_1.34G_dp2_tp32_pp1_acc8_mbs32_seq8192_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp2_tp32_pp1_acc8_mbs32_seq8192_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=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_acc8_mbs32_seq8192_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp2_tp4_pp1_acc32_mbs8_seq8192_zero1_tpmodeALL_vocab131k.sh b/scripts/run_1.34G_dp2_tp4_pp1_acc32_mbs8_seq8192_zero1_tpmodeALL_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..1b22b9b58ce4c1475fdf59fe7d5ec7b80bcfe81a --- /dev/null +++ b/scripts/run_1.34G_dp2_tp4_pp1_acc32_mbs8_seq8192_zero1_tpmodeALL_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp2_tp4_pp1_acc32_mbs8_seq8192_zero1_tpmodeALL_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=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_acc32_mbs8_seq8192_zero1_tpmodeALL_vocab131k.yaml diff --git a/scripts/run_1.34G_dp2_tp8_pp1_acc1_mbs16_seq32768_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp2_tp8_pp1_acc1_mbs16_seq32768_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..1204729c648e7d9dd114f1ff149d4b0bf94bdea8 --- /dev/null +++ b/scripts/run_1.34G_dp2_tp8_pp1_acc1_mbs16_seq32768_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp2_tp8_pp1_acc1_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=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_dp2_tp8_pp1_acc1_mbs16_seq32768_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp2_tp8_pp4_acc32_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp2_tp8_pp4_acc32_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..43506584af4daab68daa3c6adc3ed552079c4b98 --- /dev/null +++ b/scripts/run_1.34G_dp2_tp8_pp4_acc32_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_1.34G_dp2_tp8_pp4_acc32_mbs4_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=8 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=INFO # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_1.34G_dp2_tp8_pp4_acc32_mbs4_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp2_tp8_pp4_acc32_mbs4_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp2_tp8_pp4_acc32_mbs4_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_1.34G_dp32_tp16_pp1_acc1_mbs16_seq8192_zero1_tpmodeALL_vocab131k.sh b/scripts/run_1.34G_dp32_tp16_pp1_acc1_mbs16_seq8192_zero1_tpmodeALL_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..f30557deac521c33391d65150c684a9539b1e8da --- /dev/null +++ b/scripts/run_1.34G_dp32_tp16_pp1_acc1_mbs16_seq8192_zero1_tpmodeALL_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp32_tp16_pp1_acc1_mbs16_seq8192_zero1_tpmodeALL_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=64 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp32_tp16_pp1_acc1_mbs16_seq8192_zero1_tpmodeALL_vocab131k.yaml diff --git a/scripts/run_1.34G_dp4_tp16_pp1_acc128_mbs1_seq8192_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp4_tp16_pp1_acc128_mbs1_seq8192_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..cdc53eafb2d8fc2024b11bdee82d8195b6ff7ace --- /dev/null +++ b/scripts/run_1.34G_dp4_tp16_pp1_acc128_mbs1_seq8192_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp4_tp16_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=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_dp4_tp16_pp1_acc128_mbs1_seq8192_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp4_tp1_pp2_acc2_mbs4_seq32768_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp4_tp1_pp2_acc2_mbs4_seq32768_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..89a9cbafbaec7f1c49e0e018908b76a399f95045 --- /dev/null +++ b/scripts/run_1.34G_dp4_tp1_pp2_acc2_mbs4_seq32768_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp4_tp1_pp2_acc2_mbs4_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_dp4_tp1_pp2_acc2_mbs4_seq32768_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp4_tp1_pp2_acc4_mbs2_seq32768_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp4_tp1_pp2_acc4_mbs2_seq32768_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..ac34ddfb769a19f14ee7f81487b5c5d5e9937a07 --- /dev/null +++ b/scripts/run_1.34G_dp4_tp1_pp2_acc4_mbs2_seq32768_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp4_tp1_pp2_acc4_mbs2_seq32768_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=1 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp4_tp1_pp2_acc4_mbs2_seq32768_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp4_tp4_pp1_acc4_mbs8_seq32768_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp4_tp4_pp1_acc4_mbs8_seq32768_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..6f6b1bdd8e31e264135bc6bbbafc2f31ee890e7a --- /dev/null +++ b/scripts/run_1.34G_dp4_tp4_pp1_acc4_mbs8_seq32768_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp4_tp4_pp1_acc4_mbs8_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=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_acc4_mbs8_seq32768_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp4_tp64_pp1_acc1_mbs8_seq32768_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp4_tp64_pp1_acc1_mbs8_seq32768_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..bcb65175d8845f2f61eca5d53694fad339481e6f --- /dev/null +++ b/scripts/run_1.34G_dp4_tp64_pp1_acc1_mbs8_seq32768_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp4_tp64_pp1_acc1_mbs8_seq32768_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=32 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp4_tp64_pp1_acc1_mbs8_seq32768_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp4_tp64_pp1_acc2_mbs64_seq8192_zero1_tpmodeALL_vocab131k.sh b/scripts/run_1.34G_dp4_tp64_pp1_acc2_mbs64_seq8192_zero1_tpmodeALL_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..70673605a150af230813cb36fe11d83b5e4acf05 --- /dev/null +++ b/scripts/run_1.34G_dp4_tp64_pp1_acc2_mbs64_seq8192_zero1_tpmodeALL_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp4_tp64_pp1_acc2_mbs64_seq8192_zero1_tpmodeALL_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=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_acc2_mbs64_seq8192_zero1_tpmodeALL_vocab131k.yaml diff --git a/scripts/run_1.34G_dp64_tp1_pp2_acc8_mbs1_seq8192_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp64_tp1_pp2_acc8_mbs1_seq8192_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..6ad2d4d7b6c23ea1e57f40eaaf421971514368e2 --- /dev/null +++ b/scripts/run_1.34G_dp64_tp1_pp2_acc8_mbs1_seq8192_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp64_tp1_pp2_acc8_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=16 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp64_tp1_pp2_acc8_mbs1_seq8192_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp8_tp16_pp1_acc32_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp8_tp16_pp1_acc32_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..4bab362febb2ab688aa971203fdbcd01320031d1 --- /dev/null +++ b/scripts/run_1.34G_dp8_tp16_pp1_acc32_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,124 @@ +#!/bin/bash +#SBATCH --job-name=bench_1.34G_dp8_tp16_pp1_acc32_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 $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_1.34G_dp8_tp16_pp1_acc32_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_dp8_tp16_pp1_acc32_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_dp8_tp16_pp1_acc32_mbs1_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_1.34G_dp8_tp1_pp2_acc8_mbs8_seq2048_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp8_tp1_pp2_acc8_mbs8_seq2048_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..d22dcee062142769035757de8ad723e4a7eee8ec --- /dev/null +++ b/scripts/run_1.34G_dp8_tp1_pp2_acc8_mbs8_seq2048_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp8_tp1_pp2_acc8_mbs8_seq2048_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=2 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp8_tp1_pp2_acc8_mbs8_seq2048_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp8_tp64_pp1_acc2_mbs2_seq32768_zero1_tpmodeALL_vocab131k.sh b/scripts/run_1.34G_dp8_tp64_pp1_acc2_mbs2_seq32768_zero1_tpmodeALL_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..a78f49a9e5b691f8fb6fac4b749a9b78d2d3ccaf --- /dev/null +++ b/scripts/run_1.34G_dp8_tp64_pp1_acc2_mbs2_seq32768_zero1_tpmodeALL_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp8_tp64_pp1_acc2_mbs2_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=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_acc2_mbs2_seq32768_zero1_tpmodeALL_vocab131k.yaml diff --git a/scripts/run_187G_dp1_tp64_pp8_acc1_mbs1_seq2048_zero0_tpmodeRED_l126_h16384_heads128.sh b/scripts/run_187G_dp1_tp64_pp8_acc1_mbs1_seq2048_zero0_tpmodeRED_l126_h16384_heads128.sh new file mode 100644 index 0000000000000000000000000000000000000000..4a70d326670b1dc9ee05b7aec4e3c18e7f773729 --- /dev/null +++ b/scripts/run_187G_dp1_tp64_pp8_acc1_mbs1_seq2048_zero0_tpmodeRED_l126_h16384_heads128.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_187G_dp1_tp64_pp8_acc1_mbs1_seq2048_zero0_tpmodeRED_l126_h16384_heads128 # 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_187G_dp1_tp64_pp8_acc1_mbs1_seq2048_zero0_tpmodeRED_l126_h16384_heads128.yaml diff --git a/scripts/run_2.28G_dp16_tp2_pp1_acc4_mbs4_seq4096_zero1_tpmodeALL_l26_h2304_heads16.sh b/scripts/run_2.28G_dp16_tp2_pp1_acc4_mbs4_seq4096_zero1_tpmodeALL_l26_h2304_heads16.sh new file mode 100644 index 0000000000000000000000000000000000000000..52701647134d40157078e0d5881075e398a21b76 --- /dev/null +++ b/scripts/run_2.28G_dp16_tp2_pp1_acc4_mbs4_seq4096_zero1_tpmodeALL_l26_h2304_heads16.sh @@ -0,0 +1,57 @@ +#!/bin/bash + +#SBATCH --job-name=bench_2.28G_dp16_tp2_pp1_acc4_mbs4_seq4096_zero1_tpmodeALL_l26_h2304_heads16 # 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=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 + +# 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_2.28G_dp16_tp2_pp1_acc4_mbs4_seq4096_zero1_tpmodeALL_l26_h2304_heads16.yaml diff --git a/scripts/run_2.28G_dp1_tp8_pp1_acc1_mbs1_seq65536_zero0_tpmodeRED_l26_h2304_heads16.sh b/scripts/run_2.28G_dp1_tp8_pp1_acc1_mbs1_seq65536_zero0_tpmodeRED_l26_h2304_heads16.sh new file mode 100644 index 0000000000000000000000000000000000000000..b431855cd7f02108a2e9b159bd650ebe3e3841df --- /dev/null +++ b/scripts/run_2.28G_dp1_tp8_pp1_acc1_mbs1_seq65536_zero0_tpmodeRED_l26_h2304_heads16.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_2.28G_dp1_tp8_pp1_acc1_mbs1_seq65536_zero0_tpmodeRED_l26_h2304_heads16 # 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=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 specificsbat +# 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_2.28G_dp1_tp8_pp1_acc1_mbs1_seq65536_zero0_tpmodeRED_l26_h2304_heads16.yaml diff --git a/scripts/run_2.28G_dp8_tp4_pp1_acc4_mbs8_seq2048_zero1_tpmodeRED_l26_h2304_heads16.sh b/scripts/run_2.28G_dp8_tp4_pp1_acc4_mbs8_seq2048_zero1_tpmodeRED_l26_h2304_heads16.sh new file mode 100644 index 0000000000000000000000000000000000000000..5a7add3ac381c44ef10e7f1e21ccb951cdd76e3b --- /dev/null +++ b/scripts/run_2.28G_dp8_tp4_pp1_acc4_mbs8_seq2048_zero1_tpmodeRED_l26_h2304_heads16.sh @@ -0,0 +1,57 @@ +#!/bin/bash + +#SBATCH --job-name=bench_2.28G_dp8_tp4_pp1_acc4_mbs8_seq2048_zero1_tpmodeRED_l26_h2304_heads16 # 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=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 + +# 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_2.28G_dp8_tp4_pp1_acc4_mbs8_seq2048_zero1_tpmodeRED_l26_h2304_heads16.yaml diff --git a/scripts/run_3.57G_dp1_tp1_pp128_acc2_mbs128_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_3.57G_dp1_tp1_pp128_acc2_mbs128_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..5277a07180d870bb9efb6165ee80352065a96ee6 --- /dev/null +++ b/scripts/run_3.57G_dp1_tp1_pp128_acc2_mbs128_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,73 @@ +#!/bin/bash + +#SBATCH --job-name=bench_3.57G_dp1_tp1_pp128_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=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 + +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_dp1_tp1_pp128_acc2_mbs128_seq4096_zero0_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_3.57G_dp1_tp4_pp1_acc1_mbs7_seq4096_zero0_tpmodeALL_vocab131k_cache.sh b/scripts/run_3.57G_dp1_tp4_pp1_acc1_mbs7_seq4096_zero0_tpmodeALL_vocab131k_cache.sh new file mode 100644 index 0000000000000000000000000000000000000000..b1ffa81221301b41bc43d05db8bed2fcf8d58b64 --- /dev/null +++ b/scripts/run_3.57G_dp1_tp4_pp1_acc1_mbs7_seq4096_zero0_tpmodeALL_vocab131k_cache.sh @@ -0,0 +1,161 @@ +#!/bin/bash +#SBATCH --job-name=bench_3.57G_dp1_tp4_pp1_acc1_mbs7_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=4 +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_tp4_pp1_acc1_mbs7_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_tp4_pp1_acc1_mbs7_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_tp4_pp1_acc1_mbs7_seq4096_zero0_tpmodeALL_vocab131k_cache.yaml +fi diff --git a/scripts/run_3.57G_dp2_tp1_pp4_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_3.57G_dp2_tp1_pp4_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..52feaaa78801f80fd32a5d158ee4dfbe6074d340 --- /dev/null +++ b/scripts/run_3.57G_dp2_tp1_pp4_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_3.57G_dp2_tp1_pp4_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=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_3.57G_dp2_tp1_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_3.57G_dp2_tp1_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_3.57G_dp2_tp1_pp4_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_3.57G_dp2_tp4_pp4_acc64_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_3.57G_dp2_tp4_pp4_acc64_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..5e9bdb85f199b05b2120831625d3f6a7fb5d1bb7 --- /dev/null +++ b/scripts/run_3.57G_dp2_tp4_pp4_acc64_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_3.57G_dp2_tp4_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=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_3.57G_dp2_tp4_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_3.57G_dp2_tp4_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_3.57G_dp2_tp4_pp4_acc64_mbs2_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_3.57G_dp2_tp8_pp1_acc16_mbs8_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_3.57G_dp2_tp8_pp1_acc16_mbs8_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..0a4704e88f35785baef5d55948e66ed049759ce0 --- /dev/null +++ b/scripts/run_3.57G_dp2_tp8_pp1_acc16_mbs8_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,124 @@ +#!/bin/bash +#SBATCH --job-name=bench_3.57G_dp2_tp8_pp1_acc16_mbs8_seq4096_zero0_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=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_3.57G_dp2_tp8_pp1_acc16_mbs8_seq4096_zero0_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_3.57G_dp2_tp8_pp1_acc16_mbs8_seq4096_zero0_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_3.57G_dp2_tp8_pp1_acc16_mbs8_seq4096_zero0_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_3.57G_dp4_tp4_pp32_acc32_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_3.57G_dp4_tp4_pp32_acc32_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..d05b7121f762ce29dec814901143fbb3635677eb --- /dev/null +++ b/scripts/run_3.57G_dp4_tp4_pp32_acc32_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_3.57G_dp4_tp4_pp32_acc32_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_3.57G_dp4_tp4_pp32_acc32_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_dp4_tp4_pp32_acc32_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_dp4_tp4_pp32_acc32_mbs2_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_469G_dp1_tp2_pp8_acc256_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_469G_dp1_tp2_pp8_acc256_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..538889623b16dfcf9fdcc92de1b7ed4d9eed7f7e --- /dev/null +++ b/scripts/run_469G_dp1_tp2_pp8_acc256_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,73 @@ +#!/bin/bash + +#SBATCH --job-name=bench_469G_dp1_tp2_pp8_acc256_mbs1_seq4096_zero0_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high +#SBATCH --exclude=ip-26-0-160-192,ip-26-0-171-102 + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=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_469G_dp1_tp2_pp8_acc256_mbs1_seq4096_zero0_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_469G_dp64_tp2_pp2_acc1_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_469G_dp64_tp2_pp2_acc1_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..17fd4612b2287605e042cb8bf67c8cf59bd97c78 --- /dev/null +++ b/scripts/run_469G_dp64_tp2_pp2_acc1_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_469G_dp64_tp2_pp2_acc1_mbs4_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=32 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$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_dp64_tp2_pp2_acc1_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_469G_dp64_tp2_pp2_acc1_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_469G_dp64_tp2_pp2_acc1_mbs4_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_469G_dp8_tp1_pp2_acc4_mbs8_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_469G_dp8_tp1_pp2_acc4_mbs8_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..e0bef4bf7b386f9c818c04a16cb06542ef39b345 --- /dev/null +++ b/scripts/run_469G_dp8_tp1_pp2_acc4_mbs8_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,73 @@ +#!/bin/bash + +#SBATCH --job-name=bench_469G_dp8_tp1_pp2_acc4_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=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_469G_dp8_tp1_pp2_acc4_mbs8_seq4096_zero0_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_8.86G_dp128_tp2_pp2_acc2_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_8.86G_dp128_tp2_pp2_acc2_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..81464f1be2f1945d8d9d9f4e16dbbfb01ae64e90 --- /dev/null +++ b/scripts/run_8.86G_dp128_tp2_pp2_acc2_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,161 @@ +#!/bin/bash +#SBATCH --job-name=bench_8.86G_dp128_tp2_pp2_acc2_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=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 +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_8.86G_dp128_tp2_pp2_acc2_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_dp128_tp2_pp2_acc2_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_dp128_tp2_pp2_acc2_mbs1_seq4096_zero0_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_8.86G_dp16_tp2_pp2_acc16_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_8.86G_dp16_tp2_pp2_acc16_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..07bd0c2f077883fdbb53d333662a9e9a29d37777 --- /dev/null +++ b/scripts/run_8.86G_dp16_tp2_pp2_acc16_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_8.86G_dp16_tp2_pp2_acc16_mbs1_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=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_8.86G_dp16_tp2_pp2_acc16_mbs1_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_8.86G_dp16_tp2_pp2_acc16_mbs1_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_8.86G_dp16_tp2_pp2_acc16_mbs1_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_8.86G_dp16_tp32_pp1_acc16_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_8.86G_dp16_tp32_pp1_acc16_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..94ce765f267063e809f8aa700d5fd8a87d6fe457 --- /dev/null +++ b/scripts/run_8.86G_dp16_tp32_pp1_acc16_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_8.86G_dp16_tp32_pp1_acc16_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=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_dp16_tp32_pp1_acc16_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_dp16_tp32_pp1_acc16_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_dp16_tp32_pp1_acc16_mbs1_seq4096_zero0_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_8.86G_dp16_tp8_pp2_acc2_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_8.86G_dp16_tp8_pp2_acc2_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..1e67cdeffce623536bd07a6ed3f42fc19a70f365 --- /dev/null +++ b/scripts/run_8.86G_dp16_tp8_pp2_acc2_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_8.86G_dp16_tp8_pp2_acc2_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_8.86G_dp16_tp8_pp2_acc2_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_dp16_tp8_pp2_acc2_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_dp16_tp8_pp2_acc2_mbs8_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_8.86G_dp1_tp1_pp32_acc1_mbs256_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_8.86G_dp1_tp1_pp32_acc1_mbs256_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..4ae2eac9066c9aea581013a2c4ca613e13bd3e54 --- /dev/null +++ b/scripts/run_8.86G_dp1_tp1_pp32_acc1_mbs256_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,73 @@ +#!/bin/bash + +#SBATCH --job-name=bench_8.86G_dp1_tp1_pp32_acc1_mbs256_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_8.86G_dp1_tp1_pp32_acc1_mbs256_seq4096_zero0_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_8.86G_dp1_tp1_pp64_acc4_mbs64_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_8.86G_dp1_tp1_pp64_acc4_mbs64_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..79dd93f87160c73ab796db5cf56414be2a9637c4 --- /dev/null +++ b/scripts/run_8.86G_dp1_tp1_pp64_acc4_mbs64_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,73 @@ +#!/bin/bash + +#SBATCH --job-name=bench_8.86G_dp1_tp1_pp64_acc4_mbs64_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_dp1_tp1_pp64_acc4_mbs64_seq4096_zero0_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_8.86G_dp1_tp2_pp8_acc32_mbs8_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_8.86G_dp1_tp2_pp8_acc32_mbs8_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..0bfed57cf20de260829c87c60a73325614eeeb4f --- /dev/null +++ b/scripts/run_8.86G_dp1_tp2_pp8_acc32_mbs8_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,73 @@ +#!/bin/bash + +#SBATCH --job-name=bench_8.86G_dp1_tp2_pp8_acc32_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=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_8.86G_dp1_tp2_pp8_acc32_mbs8_seq4096_zero0_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_8.86G_dp1_tp4_pp2_acc8_mbs32_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_8.86G_dp1_tp4_pp2_acc8_mbs32_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..9c0e8ae9ceb7e0ed8cc0fdfb1968a859b4aa24fd --- /dev/null +++ b/scripts/run_8.86G_dp1_tp4_pp2_acc8_mbs32_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,73 @@ +#!/bin/bash + +#SBATCH --job-name=bench_8.86G_dp1_tp4_pp2_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=1 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +set -x -e + +# 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_dp1_tp4_pp2_acc8_mbs32_seq4096_zero0_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_8.86G_dp1_tp4_pp8_acc16_mbs16_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_8.86G_dp1_tp4_pp8_acc16_mbs16_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..6bbc91f8f0ccf5ad99f87a6eb0ef0ae4f54e37b8 --- /dev/null +++ b/scripts/run_8.86G_dp1_tp4_pp8_acc16_mbs16_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,73 @@ +#!/bin/bash + +#SBATCH --job-name=bench_8.86G_dp1_tp4_pp8_acc16_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=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_8.86G_dp1_tp4_pp8_acc16_mbs16_seq4096_zero0_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_8.86G_dp2_tp16_pp1_acc8_mbs16_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_8.86G_dp2_tp16_pp1_acc8_mbs16_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..2d44834e377a6016d8ff204f31dc880bd699ade1 --- /dev/null +++ b/scripts/run_8.86G_dp2_tp16_pp1_acc8_mbs16_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,161 @@ +#!/bin/bash +#SBATCH --job-name=bench_8.86G_dp2_tp16_pp1_acc8_mbs16_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_8.86G_dp2_tp16_pp1_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_8.86G_dp2_tp16_pp1_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_8.86G_dp2_tp16_pp1_acc8_mbs16_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_8.86G_dp4_tp8_pp16_acc16_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_8.86G_dp4_tp8_pp16_acc16_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..7df15543f04b172aa9b0605beaf8e1d5aaac0bf8 --- /dev/null +++ b/scripts/run_8.86G_dp4_tp8_pp16_acc16_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_8.86G_dp4_tp8_pp16_acc16_mbs4_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_dp4_tp8_pp16_acc16_mbs4_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_8.86G_dp4_tp8_pp16_acc16_mbs4_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_8.86G_dp4_tp8_pp16_acc16_mbs4_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_8.86G_dp64_tp1_pp4_acc4_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_8.86G_dp64_tp1_pp4_acc4_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..3e53cd27e309166b992cb9d3ccb8e7e6b674c4f0 --- /dev/null +++ b/scripts/run_8.86G_dp64_tp1_pp4_acc4_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,161 @@ +#!/bin/bash +#SBATCH --job-name=bench_8.86G_dp64_tp1_pp4_acc4_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_8.86G_dp64_tp1_pp4_acc4_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_dp64_tp1_pp4_acc4_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_dp64_tp1_pp4_acc4_mbs1_seq4096_zero0_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_8.86G_dp8_tp16_pp1_acc8_mbs4_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_8.86G_dp8_tp16_pp1_acc8_mbs4_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..5284702b3084547f08e9bc414e0729a4d310a7bc --- /dev/null +++ b/scripts/run_8.86G_dp8_tp16_pp1_acc8_mbs4_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,124 @@ +#!/bin/bash +#SBATCH --job-name=bench_8.86G_dp8_tp16_pp1_acc8_mbs4_seq4096_zero0_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=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_acc8_mbs4_seq4096_zero0_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_8.86G_dp8_tp16_pp1_acc8_mbs4_seq4096_zero0_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_8.86G_dp8_tp16_pp1_acc8_mbs4_seq4096_zero0_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_8.86G_dp8_tp1_pp8_acc8_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_8.86G_dp8_tp1_pp8_acc8_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..442c9b629749cf16507e7d250febfefa84a054bb --- /dev/null +++ b/scripts/run_8.86G_dp8_tp1_pp8_acc8_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_8.86G_dp8_tp1_pp8_acc8_mbs4_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=8 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=INFO # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_8.86G_dp8_tp1_pp8_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_8.86G_dp8_tp1_pp8_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_8.86G_dp8_tp1_pp8_acc8_mbs4_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_8.86G_dp8_tp2_pp4_acc4_mbs8_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_8.86G_dp8_tp2_pp4_acc4_mbs8_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..7662511541dc4b99f7407318ec5fcdb1cd9d4b2c --- /dev/null +++ b/scripts/run_8.86G_dp8_tp2_pp4_acc4_mbs8_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,73 @@ +#!/bin/bash + +#SBATCH --job-name=bench_8.86G_dp8_tp2_pp4_acc4_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_dp8_tp2_pp4_acc4_mbs8_seq4096_zero0_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_80G_dp1_tp32_pp2_acc1_mbs256_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_80G_dp1_tp32_pp2_acc1_mbs256_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..fa5d8bef8f8e40f8651591feedf526f555dd2036 --- /dev/null +++ b/scripts/run_80G_dp1_tp32_pp2_acc1_mbs256_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,73 @@ +#!/bin/bash + +#SBATCH --job-name=bench_80G_dp1_tp32_pp2_acc1_mbs256_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_dp1_tp32_pp2_acc1_mbs256_seq4096_zero0_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_80G_dp1_tp4_pp8_acc16_mbs16_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_80G_dp1_tp4_pp8_acc16_mbs16_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..7958c83cca77738266bbac3a6b4ba1268101a36b --- /dev/null +++ b/scripts/run_80G_dp1_tp4_pp8_acc16_mbs16_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,73 @@ +#!/bin/bash + +#SBATCH --job-name=bench_80G_dp1_tp4_pp8_acc16_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=4 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +# Disable wandb +export WANDB_MODE=disabled + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + + +# Print GPU topology information +echo "=== GPU Topology ===" +nvidia-smi topo -m +echo "==================" + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp1_tp4_pp8_acc16_mbs16_seq4096_zero0_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_80G_dp2_tp1_pp4_acc128_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_80G_dp2_tp1_pp4_acc128_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..7c45b36e3c2b81363095a700ca183f91769f7e7a --- /dev/null +++ b/scripts/run_80G_dp2_tp1_pp4_acc128_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,73 @@ +#!/bin/bash + +#SBATCH --job-name=bench_80G_dp2_tp1_pp4_acc128_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=1 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +set -x -e + +# 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_dp2_tp1_pp4_acc128_mbs1_seq4096_zero0_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_80G_dp32_tp16_pp1_acc1_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_80G_dp32_tp16_pp1_acc1_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..663a0ed9dffc5d3a2540027cf353a9c005968f6f --- /dev/null +++ b/scripts/run_80G_dp32_tp16_pp1_acc1_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,124 @@ +#!/bin/bash +#SBATCH --job-name=bench_80G_dp32_tp16_pp1_acc1_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=64 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_80G_dp32_tp16_pp1_acc1_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_dp32_tp16_pp1_acc1_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_dp32_tp16_pp1_acc1_mbs8_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_80G_dp4_tp16_pp2_acc64_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_80G_dp4_tp16_pp2_acc64_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..e9165311d64dee48fdf376411a0a1cf68ec2ac4a --- /dev/null +++ b/scripts/run_80G_dp4_tp16_pp2_acc64_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_80G_dp4_tp16_pp2_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=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_dp4_tp16_pp2_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_80G_dp4_tp16_pp2_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_80G_dp4_tp16_pp2_acc64_mbs1_seq4096_zero0_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_80G_dp8_tp1_pp4_acc4_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_80G_dp8_tp1_pp4_acc4_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..115429406e04d815f6f4c558202fc756efe9bc08 --- /dev/null +++ b/scripts/run_80G_dp8_tp1_pp4_acc4_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_80G_dp8_tp1_pp4_acc4_mbs8_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=4 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=INFO # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_80G_dp8_tp1_pp4_acc4_mbs8_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp8_tp1_pp4_acc4_mbs8_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp8_tp1_pp4_acc4_mbs8_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_80G_dp8_tp8_pp2_acc16_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_80G_dp8_tp8_pp2_acc16_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..314409c271967b138608400a576edf54a8f07c4e --- /dev/null +++ b/scripts/run_80G_dp8_tp8_pp2_acc16_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_80G_dp8_tp8_pp2_acc16_mbs2_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=16 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=INFO # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_80G_dp8_tp8_pp2_acc16_mbs2_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp8_tp8_pp2_acc16_mbs2_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp8_tp8_pp2_acc16_mbs2_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_dp1_tp1_pp1_acc1_mbs1_seq2048_zero0.sh b/scripts/run_dp1_tp1_pp1_acc1_mbs1_seq2048_zero0.sh new file mode 100644 index 0000000000000000000000000000000000000000..b3d70dda0be72e520e6c108bbb258370ee2dacbd --- /dev/null +++ b/scripts/run_dp1_tp1_pp1_acc1_mbs1_seq2048_zero0.sh @@ -0,0 +1,57 @@ +#!/bin/bash + +#SBATCH --job-name=bench_dp1_tp1_pp1_acc1_mbs1_seq2048_zero0 # 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=1 # 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_dp1_tp1_pp1_acc1_mbs1_seq2048_zero0.yaml