diff --git a/scripts/run_1.14G_dp128_tp2_pp1_acc8_mbs2_seq2048_zero1_tpmodeALL_vocab32k.sh b/scripts/run_1.14G_dp128_tp2_pp1_acc8_mbs2_seq2048_zero1_tpmodeALL_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..06b3255bc4f6f08ce8045ea5cfa7c310e8cac060 --- /dev/null +++ b/scripts/run_1.14G_dp128_tp2_pp1_acc8_mbs2_seq2048_zero1_tpmodeALL_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp128_tp2_pp1_acc8_mbs2_seq2048_zero1_tpmodeALL_vocab32k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=32 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.14G_dp128_tp2_pp1_acc8_mbs2_seq2048_zero1_tpmodeALL_vocab32k.yaml diff --git a/scripts/run_1.14G_dp16_tp16_pp1_acc4_mbs8_seq2048_zero1_tpmodeALL_vocab32k.sh b/scripts/run_1.14G_dp16_tp16_pp1_acc4_mbs8_seq2048_zero1_tpmodeALL_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..b4d5e5e21a4c501f74083ef83f3ba5d799217b8f --- /dev/null +++ b/scripts/run_1.14G_dp16_tp16_pp1_acc4_mbs8_seq2048_zero1_tpmodeALL_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp16_tp16_pp1_acc4_mbs8_seq2048_zero1_tpmodeALL_vocab32k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=32 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.14G_dp16_tp16_pp1_acc4_mbs8_seq2048_zero1_tpmodeALL_vocab32k.yaml diff --git a/scripts/run_1.14G_dp16_tp32_pp1_acc8_mbs16_seq2048_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp16_tp32_pp1_acc8_mbs16_seq2048_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..4fe1d3b57d65c4d1487bc9cfa36eda808adcabff --- /dev/null +++ b/scripts/run_1.14G_dp16_tp32_pp1_acc8_mbs16_seq2048_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp16_tp32_pp1_acc8_mbs16_seq2048_zero1_tpmodeRED_vocab32k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=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_dp16_tp32_pp1_acc8_mbs16_seq2048_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp16_tp8_pp1_acc2_mbs4_seq8192_zero1_tpmodeALL_vocab32k.sh b/scripts/run_1.14G_dp16_tp8_pp1_acc2_mbs4_seq8192_zero1_tpmodeALL_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..01c1cd0ae338ab10905e3d67a6dd77c261869cdd --- /dev/null +++ b/scripts/run_1.14G_dp16_tp8_pp1_acc2_mbs4_seq8192_zero1_tpmodeALL_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp16_tp8_pp1_acc2_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=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_dp16_tp8_pp1_acc2_mbs4_seq8192_zero1_tpmodeALL_vocab32k.yaml diff --git a/scripts/run_1.14G_dp256_tp2_pp1_acc1_mbs2_seq8192_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp256_tp2_pp1_acc1_mbs2_seq8192_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..b272c32e656366e17aec000d0a754076bc0ddf70 --- /dev/null +++ b/scripts/run_1.14G_dp256_tp2_pp1_acc1_mbs2_seq8192_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp256_tp2_pp1_acc1_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=64 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.14G_dp256_tp2_pp1_acc1_mbs2_seq8192_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp2_tp128_pp1_acc16_mbs16_seq8192_zero1_tpmodeALL_vocab32k.sh b/scripts/run_1.14G_dp2_tp128_pp1_acc16_mbs16_seq8192_zero1_tpmodeALL_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..0612535eeb8c997c94ad8274d934b6f85ac65572 --- /dev/null +++ b/scripts/run_1.14G_dp2_tp128_pp1_acc16_mbs16_seq8192_zero1_tpmodeALL_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp2_tp128_pp1_acc16_mbs16_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_acc16_mbs16_seq8192_zero1_tpmodeALL_vocab32k.yaml diff --git a/scripts/run_1.14G_dp2_tp32_pp1_acc16_mbs1_seq32768_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp2_tp32_pp1_acc16_mbs1_seq32768_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..fc99434629eab0c4ebe1df8589855920a039038f --- /dev/null +++ b/scripts/run_1.14G_dp2_tp32_pp1_acc16_mbs1_seq32768_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp2_tp32_pp1_acc16_mbs1_seq32768_zero1_tpmodeRED_vocab32k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=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_dp2_tp32_pp1_acc16_mbs1_seq32768_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp2_tp64_pp1_acc64_mbs16_seq2048_zero1_tpmodeALL_vocab32k.sh b/scripts/run_1.14G_dp2_tp64_pp1_acc64_mbs16_seq2048_zero1_tpmodeALL_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..29d77fc02bc0f88339bbc9af7fc7f5ebeda0b2ef --- /dev/null +++ b/scripts/run_1.14G_dp2_tp64_pp1_acc64_mbs16_seq2048_zero1_tpmodeALL_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp2_tp64_pp1_acc64_mbs16_seq2048_zero1_tpmodeALL_vocab32k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=16 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.14G_dp2_tp64_pp1_acc64_mbs16_seq2048_zero1_tpmodeALL_vocab32k.yaml diff --git a/scripts/run_1.14G_dp2_tp8_pp1_acc1_mbs64_seq8192_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp2_tp8_pp1_acc1_mbs64_seq8192_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..c26ab22ae342b4174b37bf3ed4e893797c892e64 --- /dev/null +++ b/scripts/run_1.14G_dp2_tp8_pp1_acc1_mbs64_seq8192_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp2_tp8_pp1_acc1_mbs64_seq8192_zero1_tpmodeRED_vocab32k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=2 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.14G_dp2_tp8_pp1_acc1_mbs64_seq8192_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp32_tp1_pp2_acc4_mbs4_seq8192_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp32_tp1_pp2_acc4_mbs4_seq8192_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..0edd26410eab7201523f67e47b703ebf7f6caf2c --- /dev/null +++ b/scripts/run_1.14G_dp32_tp1_pp2_acc4_mbs4_seq8192_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp32_tp1_pp2_acc4_mbs4_seq8192_zero1_tpmodeRED_vocab32k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=8 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.14G_dp32_tp1_pp2_acc4_mbs4_seq8192_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp4_tp128_pp1_acc32_mbs4_seq2048_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp4_tp128_pp1_acc32_mbs4_seq2048_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..efa07d8723afcec090bd4e2704aac5bd846046e6 --- /dev/null +++ b/scripts/run_1.14G_dp4_tp128_pp1_acc32_mbs4_seq2048_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp4_tp128_pp1_acc32_mbs4_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=64 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.14G_dp4_tp128_pp1_acc32_mbs4_seq2048_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp4_tp16_pp1_acc1_mbs32_seq8192_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp4_tp16_pp1_acc1_mbs32_seq8192_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..e8a1a06c7c59628347af29445e9ba9a9e6fc50f5 --- /dev/null +++ b/scripts/run_1.14G_dp4_tp16_pp1_acc1_mbs32_seq8192_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp4_tp16_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=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_acc1_mbs32_seq8192_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp4_tp1_pp2_acc4_mbs8_seq8192_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp4_tp1_pp2_acc4_mbs8_seq8192_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..01e3aec6b5714bebace7e24d21a7532875bf7dab --- /dev/null +++ b/scripts/run_1.14G_dp4_tp1_pp2_acc4_mbs8_seq8192_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp4_tp1_pp2_acc4_mbs8_seq8192_zero1_tpmodeRED_vocab32k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=1 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.14G_dp4_tp1_pp2_acc4_mbs8_seq8192_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp4_tp2_pp1_acc256_mbs2_seq2048_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp4_tp2_pp1_acc256_mbs2_seq2048_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..cf800e24ea6e5bbee4ac5e013e5deb051ee88675 --- /dev/null +++ b/scripts/run_1.14G_dp4_tp2_pp1_acc256_mbs2_seq2048_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp4_tp2_pp1_acc256_mbs2_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_dp4_tp2_pp1_acc256_mbs2_seq2048_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp4_tp2_pp1_acc4_mbs32_seq2048_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp4_tp2_pp1_acc4_mbs32_seq2048_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..f1a6430d6963a7b31603da7d810c2fe01d9fa37c --- /dev/null +++ b/scripts/run_1.14G_dp4_tp2_pp1_acc4_mbs32_seq2048_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp4_tp2_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=1 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.14G_dp4_tp2_pp1_acc4_mbs32_seq2048_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp4_tp4_pp1_acc2_mbs16_seq8192_zero1_tpmodeALL_vocab32k.sh b/scripts/run_1.14G_dp4_tp4_pp1_acc2_mbs16_seq8192_zero1_tpmodeALL_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..cbfb58294d5fe431af3070e55e16c350f529b1ce --- /dev/null +++ b/scripts/run_1.14G_dp4_tp4_pp1_acc2_mbs16_seq8192_zero1_tpmodeALL_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp4_tp4_pp1_acc2_mbs16_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=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_dp4_tp4_pp1_acc2_mbs16_seq8192_zero1_tpmodeALL_vocab32k.yaml diff --git a/scripts/run_1.14G_dp4_tp4_pp1_acc64_mbs2_seq8192_zero1_tpmodeALL_vocab32k.sh b/scripts/run_1.14G_dp4_tp4_pp1_acc64_mbs2_seq8192_zero1_tpmodeALL_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..be864914c93352cc5059ba55ff1a44840765edbb --- /dev/null +++ b/scripts/run_1.14G_dp4_tp4_pp1_acc64_mbs2_seq8192_zero1_tpmodeALL_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp4_tp4_pp1_acc64_mbs2_seq8192_zero1_tpmodeALL_vocab32k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=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_dp4_tp4_pp1_acc64_mbs2_seq8192_zero1_tpmodeALL_vocab32k.yaml diff --git a/scripts/run_1.14G_dp4_tp8_pp1_acc1_mbs8_seq32768_zero1_tpmodeALL_vocab32k.sh b/scripts/run_1.14G_dp4_tp8_pp1_acc1_mbs8_seq32768_zero1_tpmodeALL_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..99667b868ec0d97852ff6f71e7f4957ffc7a2d8e --- /dev/null +++ b/scripts/run_1.14G_dp4_tp8_pp1_acc1_mbs8_seq32768_zero1_tpmodeALL_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp4_tp8_pp1_acc1_mbs8_seq32768_zero1_tpmodeALL_vocab32k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=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_acc1_mbs8_seq32768_zero1_tpmodeALL_vocab32k.yaml diff --git a/scripts/run_1.14G_dp8_tp2_pp1_acc4_mbs4_seq8192_zero1_tpmodeALL_vocab32k.sh b/scripts/run_1.14G_dp8_tp2_pp1_acc4_mbs4_seq8192_zero1_tpmodeALL_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..70bfc6d0904be8df5bfeea245b584b2404296d06 --- /dev/null +++ b/scripts/run_1.14G_dp8_tp2_pp1_acc4_mbs4_seq8192_zero1_tpmodeALL_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp8_tp2_pp1_acc4_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=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_acc4_mbs4_seq8192_zero1_tpmodeALL_vocab32k.yaml diff --git a/scripts/run_1.14G_dp8_tp2_pp1_acc64_mbs1_seq2048_zero1_tpmodeALL_vocab32k.sh b/scripts/run_1.14G_dp8_tp2_pp1_acc64_mbs1_seq2048_zero1_tpmodeALL_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..641bb93f48330d3e2ec9c301317ed442a612ec06 --- /dev/null +++ b/scripts/run_1.14G_dp8_tp2_pp1_acc64_mbs1_seq2048_zero1_tpmodeALL_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp8_tp2_pp1_acc64_mbs1_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=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_acc64_mbs1_seq2048_zero1_tpmodeALL_vocab32k.yaml diff --git a/scripts/run_1.14G_dp8_tp4_pp1_acc16_mbs4_seq2048_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp8_tp4_pp1_acc16_mbs4_seq2048_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..6d9c0112380c1bbf1bd18969cb4c8baf4b895f10 --- /dev/null +++ b/scripts/run_1.14G_dp8_tp4_pp1_acc16_mbs4_seq2048_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp8_tp4_pp1_acc16_mbs4_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_dp8_tp4_pp1_acc16_mbs4_seq2048_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.34G_dp16_tp8_pp1_acc4_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp16_tp8_pp1_acc4_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..ce2591460f387a88c4bfd2c395bfd366658f6b1b --- /dev/null +++ b/scripts/run_1.34G_dp16_tp8_pp1_acc4_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,124 @@ +#!/bin/bash +#SBATCH --job-name=bench_1.34G_dp16_tp8_pp1_acc4_mbs4_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=16 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_1.34G_dp16_tp8_pp1_acc4_mbs4_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp16_tp8_pp1_acc4_mbs4_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp16_tp8_pp1_acc4_mbs4_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_1.34G_dp1_tp8_pp1_acc8_mbs32_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp1_tp8_pp1_acc8_mbs32_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..0f85939c2ab9069707b1296d8de9f1667fa6bef2 --- /dev/null +++ b/scripts/run_1.34G_dp1_tp8_pp1_acc8_mbs32_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,73 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp1_tp8_pp1_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_1.34G_dp1_tp8_pp1_acc8_mbs32_seq4096_zero0_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp2_tp128_pp1_acc2_mbs128_seq2048_zero1_tpmodeALL_vocab131k.sh b/scripts/run_1.34G_dp2_tp128_pp1_acc2_mbs128_seq2048_zero1_tpmodeALL_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..13e31140adb9a147fd3b8bcdb3ce747ba8028a82 --- /dev/null +++ b/scripts/run_1.34G_dp2_tp128_pp1_acc2_mbs128_seq2048_zero1_tpmodeALL_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp2_tp128_pp1_acc2_mbs128_seq2048_zero1_tpmodeALL_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=32 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp2_tp128_pp1_acc2_mbs128_seq2048_zero1_tpmodeALL_vocab131k.yaml diff --git a/scripts/run_1.34G_dp2_tp16_pp1_acc16_mbs4_seq8192_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp2_tp16_pp1_acc16_mbs4_seq8192_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..1b60dce8e1eb8c7b2d75cac5b7db37a6204acce6 --- /dev/null +++ b/scripts/run_1.34G_dp2_tp16_pp1_acc16_mbs4_seq8192_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp2_tp16_pp1_acc16_mbs4_seq8192_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=4 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp2_tp16_pp1_acc16_mbs4_seq8192_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp2_tp16_pp1_acc2_mbs32_seq32768_zero1_tpmodeALL_vocab131k.sh b/scripts/run_1.34G_dp2_tp16_pp1_acc2_mbs32_seq32768_zero1_tpmodeALL_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..f1b2ce0d6e39437cc9a965c21bed9cd3f33c1507 --- /dev/null +++ b/scripts/run_1.34G_dp2_tp16_pp1_acc2_mbs32_seq32768_zero1_tpmodeALL_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp2_tp16_pp1_acc2_mbs32_seq32768_zero1_tpmodeALL_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=4 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp2_tp16_pp1_acc2_mbs32_seq32768_zero1_tpmodeALL_vocab131k.yaml diff --git a/scripts/run_1.34G_dp2_tp4_pp1_acc4_mbs64_seq8192_zero1_tpmodeALL_vocab131k.sh b/scripts/run_1.34G_dp2_tp4_pp1_acc4_mbs64_seq8192_zero1_tpmodeALL_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..fd47b46e9f1a3dd8cbaac4f44ce03dbf30ff243d --- /dev/null +++ b/scripts/run_1.34G_dp2_tp4_pp1_acc4_mbs64_seq8192_zero1_tpmodeALL_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp2_tp4_pp1_acc4_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=1 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp2_tp4_pp1_acc4_mbs64_seq8192_zero1_tpmodeALL_vocab131k.yaml diff --git a/scripts/run_1.34G_dp2_tp8_pp1_acc2_mbs128_seq8192_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp2_tp8_pp1_acc2_mbs128_seq8192_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..22e43e329c92ee08d97e98b60c2429466c218748 --- /dev/null +++ b/scripts/run_1.34G_dp2_tp8_pp1_acc2_mbs128_seq8192_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp2_tp8_pp1_acc2_mbs128_seq8192_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=2 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp2_tp8_pp1_acc2_mbs128_seq8192_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp2_tp8_pp1_acc4_mbs32_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp2_tp8_pp1_acc4_mbs32_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..68bfd870bd00634af8fccce83926cec46cf189c1 --- /dev/null +++ b/scripts/run_1.34G_dp2_tp8_pp1_acc4_mbs32_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,124 @@ +#!/bin/bash +#SBATCH --job-name=bench_1.34G_dp2_tp8_pp1_acc4_mbs32_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=2 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_1.34G_dp2_tp8_pp1_acc4_mbs32_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp2_tp8_pp1_acc4_mbs32_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp2_tp8_pp1_acc4_mbs32_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_1.34G_dp2_tp8_pp1_acc64_mbs1_seq32768_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp2_tp8_pp1_acc64_mbs1_seq32768_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..1762730a919e5f587c1d31d535d61c685cb33dd1 --- /dev/null +++ b/scripts/run_1.34G_dp2_tp8_pp1_acc64_mbs1_seq32768_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp2_tp8_pp1_acc64_mbs1_seq32768_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=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_acc64_mbs1_seq32768_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp32_tp4_pp1_acc8_mbs8_seq2048_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp32_tp4_pp1_acc8_mbs8_seq2048_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..daa4d59702a772d9bcbd9557aaa9f898bd028fca --- /dev/null +++ b/scripts/run_1.34G_dp32_tp4_pp1_acc8_mbs8_seq2048_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp32_tp4_pp1_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=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_dp32_tp4_pp1_acc8_mbs8_seq2048_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp32_tp8_pp1_acc1_mbs16_seq8192_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp32_tp8_pp1_acc1_mbs16_seq8192_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..c8e26edb8bf249b823bea37beef009f26e45f191 --- /dev/null +++ b/scripts/run_1.34G_dp32_tp8_pp1_acc1_mbs16_seq8192_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp32_tp8_pp1_acc1_mbs16_seq8192_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=32 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp32_tp8_pp1_acc1_mbs16_seq8192_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp4_tp16_pp4_acc16_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp4_tp16_pp4_acc16_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..45106694c4ee106054fd73ab611e55623f69f04b --- /dev/null +++ b/scripts/run_1.34G_dp4_tp16_pp4_acc16_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_1.34G_dp4_tp16_pp4_acc16_mbs4_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=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_1.34G_dp4_tp16_pp4_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_1.34G_dp4_tp16_pp4_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_1.34G_dp4_tp16_pp4_acc16_mbs4_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_1.34G_dp4_tp2_pp1_acc16_mbs32_seq2048_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp4_tp2_pp1_acc16_mbs32_seq2048_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..e52bf5e89bff7e5f06fb76116a33702f9341137d --- /dev/null +++ b/scripts/run_1.34G_dp4_tp2_pp1_acc16_mbs32_seq2048_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp4_tp2_pp1_acc16_mbs32_seq2048_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=1 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp4_tp2_pp1_acc16_mbs32_seq2048_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp4_tp32_pp1_acc2_mbs4_seq32768_zero1_tpmodeALL_vocab131k.sh b/scripts/run_1.34G_dp4_tp32_pp1_acc2_mbs4_seq32768_zero1_tpmodeALL_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..7b582f89977b8fc37456e0543027c1e0d135a96d --- /dev/null +++ b/scripts/run_1.34G_dp4_tp32_pp1_acc2_mbs4_seq32768_zero1_tpmodeALL_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp4_tp32_pp1_acc2_mbs4_seq32768_zero1_tpmodeALL_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=16 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp4_tp32_pp1_acc2_mbs4_seq32768_zero1_tpmodeALL_vocab131k.yaml diff --git a/scripts/run_1.34G_dp4_tp32_pp1_acc32_mbs1_seq8192_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp4_tp32_pp1_acc32_mbs1_seq8192_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..02bf9ccb930088389d3d1fc95c968aacbbf3abf5 --- /dev/null +++ b/scripts/run_1.34G_dp4_tp32_pp1_acc32_mbs1_seq8192_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp4_tp32_pp1_acc32_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_dp4_tp32_pp1_acc32_mbs1_seq8192_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp4_tp32_pp1_acc4_mbs2_seq32768_zero1_tpmodeALL_vocab131k.sh b/scripts/run_1.34G_dp4_tp32_pp1_acc4_mbs2_seq32768_zero1_tpmodeALL_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..60caf839fb557e1d0a03e3a885a1b95d3e37ac3d --- /dev/null +++ b/scripts/run_1.34G_dp4_tp32_pp1_acc4_mbs2_seq32768_zero1_tpmodeALL_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp4_tp32_pp1_acc4_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=16 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp4_tp32_pp1_acc4_mbs2_seq32768_zero1_tpmodeALL_vocab131k.yaml diff --git a/scripts/run_1.34G_dp4_tp64_pp1_acc2_mbs256_seq2048_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp4_tp64_pp1_acc2_mbs256_seq2048_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..9d7c72af3ce97951c7655c68500bcb8de07f57ab --- /dev/null +++ b/scripts/run_1.34G_dp4_tp64_pp1_acc2_mbs256_seq2048_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp4_tp64_pp1_acc2_mbs256_seq2048_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=32 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp4_tp64_pp1_acc2_mbs256_seq2048_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp4_tp64_pp1_acc32_mbs1_seq32768_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp4_tp64_pp1_acc32_mbs1_seq32768_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..40deff668c5e22811f2484e6de837bb305881b5c --- /dev/null +++ b/scripts/run_1.34G_dp4_tp64_pp1_acc32_mbs1_seq32768_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp4_tp64_pp1_acc32_mbs1_seq32768_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=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_acc32_mbs1_seq32768_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp4_tp8_pp1_acc2_mbs4_seq32768_zero1_tpmodeALL_vocab131k.sh b/scripts/run_1.34G_dp4_tp8_pp1_acc2_mbs4_seq32768_zero1_tpmodeALL_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..aea246e723b9014ecf7314893d2af7a230831230 --- /dev/null +++ b/scripts/run_1.34G_dp4_tp8_pp1_acc2_mbs4_seq32768_zero1_tpmodeALL_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp4_tp8_pp1_acc2_mbs4_seq32768_zero1_tpmodeALL_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=4 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp4_tp8_pp1_acc2_mbs4_seq32768_zero1_tpmodeALL_vocab131k.yaml diff --git a/scripts/run_1.34G_dp4_tp8_pp1_acc4_mbs2_seq32768_zero1_tpmodeALL_vocab131k.sh b/scripts/run_1.34G_dp4_tp8_pp1_acc4_mbs2_seq32768_zero1_tpmodeALL_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..33c90a88c574415e0eaf1cd4def0e7d7a144701b --- /dev/null +++ b/scripts/run_1.34G_dp4_tp8_pp1_acc4_mbs2_seq32768_zero1_tpmodeALL_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp4_tp8_pp1_acc4_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=4 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp4_tp8_pp1_acc4_mbs2_seq32768_zero1_tpmodeALL_vocab131k.yaml diff --git a/scripts/run_1.34G_dp8_tp16_pp1_acc2_mbs32_seq8192_zero1_tpmodeALL_vocab131k.sh b/scripts/run_1.34G_dp8_tp16_pp1_acc2_mbs32_seq8192_zero1_tpmodeALL_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..ebfa2b1712563ef4a5024856047f6394a4f14d2c --- /dev/null +++ b/scripts/run_1.34G_dp8_tp16_pp1_acc2_mbs32_seq8192_zero1_tpmodeALL_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp8_tp16_pp1_acc2_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=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_dp8_tp16_pp1_acc2_mbs32_seq8192_zero1_tpmodeALL_vocab131k.yaml diff --git a/scripts/run_1.34G_dp8_tp4_pp2_acc2_mbs16_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp8_tp4_pp2_acc2_mbs16_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..fe04929e6148e655e99421cb06c0fa249fe60075 --- /dev/null +++ b/scripts/run_1.34G_dp8_tp4_pp2_acc2_mbs16_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_1.34G_dp8_tp4_pp2_acc2_mbs16_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=8 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=INFO # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_1.34G_dp8_tp4_pp2_acc2_mbs16_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp8_tp4_pp2_acc2_mbs16_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp8_tp4_pp2_acc2_mbs16_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_1.34G_dp8_tp8_pp1_acc16_mbs1_seq8192_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp8_tp8_pp1_acc16_mbs1_seq8192_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..7c768b78e92c10b5fdb7443265996dc11cf24f34 --- /dev/null +++ b/scripts/run_1.34G_dp8_tp8_pp1_acc16_mbs1_seq8192_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp8_tp8_pp1_acc16_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_dp8_tp8_pp1_acc16_mbs1_seq8192_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_235M_dp1_tp8_pp1_acc1_mbs8_seq4096_zero0_tpmodeALL_l12_h1024_heads16.sh b/scripts/run_235M_dp1_tp8_pp1_acc1_mbs8_seq4096_zero0_tpmodeALL_l12_h1024_heads16.sh new file mode 100644 index 0000000000000000000000000000000000000000..f73ebbcb28192fbf95a8e5274e4d5b6d60d336d0 --- /dev/null +++ b/scripts/run_235M_dp1_tp8_pp1_acc1_mbs8_seq4096_zero0_tpmodeALL_l12_h1024_heads16.sh @@ -0,0 +1,57 @@ +#!/bin/bash + +#SBATCH --job-name=bench_235M_dp1_tp8_pp1_acc1_mbs8_seq4096_zero0_tpmodeALL_l12_h1024_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 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_235M_dp1_tp8_pp1_acc1_mbs8_seq4096_zero0_tpmodeALL_l12_h1024_heads16.yaml diff --git a/scripts/run_3.57G_dp16_tp1_pp16_acc16_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_3.57G_dp16_tp1_pp16_acc16_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..c4113f6c42cc37bb5ad9804c59a5d99a60b68823 --- /dev/null +++ b/scripts/run_3.57G_dp16_tp1_pp16_acc16_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_3.57G_dp16_tp1_pp16_acc16_mbs1_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=32 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=INFO # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_3.57G_dp16_tp1_pp16_acc16_mbs1_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_3.57G_dp16_tp1_pp16_acc16_mbs1_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_3.57G_dp16_tp1_pp16_acc16_mbs1_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_3.57G_dp16_tp1_pp2_acc4_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_3.57G_dp16_tp1_pp2_acc4_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..9bbbe28e8d63c3c1ffbf09862422f2b15b710a51 --- /dev/null +++ b/scripts/run_3.57G_dp16_tp1_pp2_acc4_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_3.57G_dp16_tp1_pp2_acc4_mbs4_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=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_dp16_tp1_pp2_acc4_mbs4_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_3.57G_dp16_tp1_pp2_acc4_mbs4_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_3.57G_dp16_tp1_pp2_acc4_mbs4_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_3.57G_dp1_tp1_pp16_acc15_mbs3_seq4096_zero0_tpmodeRED_vocab131k_cache.sh b/scripts/run_3.57G_dp1_tp1_pp16_acc15_mbs3_seq4096_zero0_tpmodeRED_vocab131k_cache.sh new file mode 100644 index 0000000000000000000000000000000000000000..6f8fc26fb5c9833fc9f29438010637dbd5ab655a --- /dev/null +++ b/scripts/run_3.57G_dp1_tp1_pp16_acc15_mbs3_seq4096_zero0_tpmodeRED_vocab131k_cache.sh @@ -0,0 +1,161 @@ +#!/bin/bash +#SBATCH --job-name=bench_3.57G_dp1_tp1_pp16_acc15_mbs3_seq4096_zero0_tpmodeRED_vocab131k_cache # Job name +#SBATCH --time=00:40:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=2 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e +echo "Running script: $0" + + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_3.57G_dp1_tp1_pp16_acc15_mbs3_seq4096_zero0_tpmodeRED_vocab131k_cache" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_3.57G_dp1_tp1_pp16_acc15_mbs3_seq4096_zero0_tpmodeRED_vocab131k_cache.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_3.57G_dp1_tp1_pp16_acc15_mbs3_seq4096_zero0_tpmodeRED_vocab131k_cache.yaml +fi diff --git a/scripts/run_3.57G_dp1_tp2_pp4_acc4_mbs64_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_3.57G_dp1_tp2_pp4_acc4_mbs64_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..6c47d8bb5e5841e4af0b3b6dc4f71a728d049cbd --- /dev/null +++ b/scripts/run_3.57G_dp1_tp2_pp4_acc4_mbs64_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,73 @@ +#!/bin/bash + +#SBATCH --job-name=bench_3.57G_dp1_tp2_pp4_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=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_3.57G_dp1_tp2_pp4_acc4_mbs64_seq4096_zero0_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_3.57G_dp2_tp1_pp4_acc8_mbs16_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_3.57G_dp2_tp1_pp4_acc8_mbs16_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..6ab3ab5874b4349021a8122d3a5e22ec9a944cb0 --- /dev/null +++ b/scripts/run_3.57G_dp2_tp1_pp4_acc8_mbs16_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_3.57G_dp2_tp1_pp4_acc8_mbs16_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=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_acc8_mbs16_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_3.57G_dp2_tp1_pp4_acc8_mbs16_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_3.57G_dp2_tp1_pp4_acc8_mbs16_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_3.57G_dp4_tp8_pp1_acc64_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_3.57G_dp4_tp8_pp1_acc64_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..8e3e77bf63bb32b69aa1a1e4db130bc087cb4035 --- /dev/null +++ b/scripts/run_3.57G_dp4_tp8_pp1_acc64_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,124 @@ +#!/bin/bash +#SBATCH --job-name=bench_3.57G_dp4_tp8_pp1_acc64_mbs1_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=4 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$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_dp4_tp8_pp1_acc64_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_3.57G_dp4_tp8_pp1_acc64_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_3.57G_dp4_tp8_pp1_acc64_mbs1_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_3.57G_dp64_tp2_pp4_acc4_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_3.57G_dp64_tp2_pp4_acc4_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..c49f7b9bf6c8658abcb939dd5a41a86283e509ee --- /dev/null +++ b/scripts/run_3.57G_dp64_tp2_pp4_acc4_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,161 @@ +#!/bin/bash +#SBATCH --job-name=bench_3.57G_dp64_tp2_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=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_3.57G_dp64_tp2_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_3.57G_dp64_tp2_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_3.57G_dp64_tp2_pp4_acc4_mbs1_seq4096_zero0_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_3.57G_dp8_tp1_pp16_acc32_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_3.57G_dp8_tp1_pp16_acc32_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..76620c21dfa837c17087483fed6c6e8b00c97227 --- /dev/null +++ b/scripts/run_3.57G_dp8_tp1_pp16_acc32_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_3.57G_dp8_tp1_pp16_acc32_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_3.57G_dp8_tp1_pp16_acc32_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_3.57G_dp8_tp1_pp16_acc32_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_3.57G_dp8_tp1_pp16_acc32_mbs1_seq4096_zero0_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_3.57G_dp8_tp2_pp1_acc32_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_3.57G_dp8_tp2_pp1_acc32_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..57ed5b06fb84a360fce685738339d00d629b5e71 --- /dev/null +++ b/scripts/run_3.57G_dp8_tp2_pp1_acc32_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,124 @@ +#!/bin/bash +#SBATCH --job-name=bench_3.57G_dp8_tp2_pp1_acc32_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=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_dp8_tp2_pp1_acc32_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_3.57G_dp8_tp2_pp1_acc32_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_3.57G_dp8_tp2_pp1_acc32_mbs1_seq4096_zero0_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_469G_dp32_tp4_pp4_acc4_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_469G_dp32_tp4_pp4_acc4_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..b655a9456a87799c8ad59c8f9c671a8132cb4d55 --- /dev/null +++ b/scripts/run_469G_dp32_tp4_pp4_acc4_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_469G_dp32_tp4_pp4_acc4_mbs2_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=64 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=INFO # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_469G_dp32_tp4_pp4_acc4_mbs2_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_469G_dp32_tp4_pp4_acc4_mbs2_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_469G_dp32_tp4_pp4_acc4_mbs2_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_469G_dp4_tp16_pp1_acc64_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_469G_dp4_tp16_pp1_acc64_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..5e0b506a0cf6f6ec36e754b88dbe0aead4727df3 --- /dev/null +++ b/scripts/run_469G_dp4_tp16_pp1_acc64_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,73 @@ +#!/bin/bash + +#SBATCH --job-name=bench_469G_dp4_tp16_pp1_acc64_mbs1_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high +#SBATCH --exclude=ip-26-0-160-192,ip-26-0-171-102 + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=8 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +# Disable wandb +export WANDB_MODE=disabled + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + + +# Print GPU topology information +echo "=== GPU Topology ===" +nvidia-smi topo -m +echo "==================" + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_469G_dp4_tp16_pp1_acc64_mbs1_seq4096_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_469G_dp8_tp4_pp2_acc32_mbs1_seq4096_zero0_tpmodeRED_vocab131k_gqa8.sh b/scripts/run_469G_dp8_tp4_pp2_acc32_mbs1_seq4096_zero0_tpmodeRED_vocab131k_gqa8.sh new file mode 100644 index 0000000000000000000000000000000000000000..02bb57918e91eef92be4a5b07c9de184a1d30d67 --- /dev/null +++ b/scripts/run_469G_dp8_tp4_pp2_acc32_mbs1_seq4096_zero0_tpmodeRED_vocab131k_gqa8.sh @@ -0,0 +1,161 @@ +#!/bin/bash +#SBATCH --job-name=bench_469G_dp8_tp4_pp2_acc32_mbs1_seq4096_zero0_tpmodeRED_vocab131k_gqa8 # Job name +#SBATCH --time=00:40:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=8 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e +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_469G_dp8_tp4_pp2_acc32_mbs1_seq4096_zero0_tpmodeRED_vocab131k_gqa8" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_469G_dp8_tp4_pp2_acc32_mbs1_seq4096_zero0_tpmodeRED_vocab131k_gqa8.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_469G_dp8_tp4_pp2_acc32_mbs1_seq4096_zero0_tpmodeRED_vocab131k_gqa8.yaml +fi diff --git a/scripts/run_469G_dp8_tp8_pp1_acc32_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_469G_dp8_tp8_pp1_acc32_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..acdb4aa7484610f17012fbb6b5aceb3ff56a8e3b --- /dev/null +++ b/scripts/run_469G_dp8_tp8_pp1_acc32_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,73 @@ +#!/bin/bash + +#SBATCH --job-name=bench_469G_dp8_tp8_pp1_acc32_mbs1_seq4096_zero0_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high +#SBATCH --exclude=ip-26-0-160-192,ip-26-0-171-102 + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=8 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +# Disable wandb +export WANDB_MODE=disabled + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + + +# Print GPU topology information +echo "=== GPU Topology ===" +nvidia-smi topo -m +echo "==================" + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_469G_dp8_tp8_pp1_acc32_mbs1_seq4096_zero0_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_8.86G_dp16_tp1_pp8_acc16_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_8.86G_dp16_tp1_pp8_acc16_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..1b8291bfc86c170fd83a01d4185e7906e40349a6 --- /dev/null +++ b/scripts/run_8.86G_dp16_tp1_pp8_acc16_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_8.86G_dp16_tp1_pp8_acc16_mbs1_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=16 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=INFO # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_8.86G_dp16_tp1_pp8_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_tp1_pp8_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_tp1_pp8_acc16_mbs1_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_8.86G_dp2_tp2_pp2_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_8.86G_dp2_tp2_pp2_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..8f7c3d2b3732edcdcd8d8c357511e18abddc2eec --- /dev/null +++ b/scripts/run_8.86G_dp2_tp2_pp2_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_8.86G_dp2_tp2_pp2_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=1 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=INFO # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_8.86G_dp2_tp2_pp2_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_8.86G_dp2_tp2_pp2_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_8.86G_dp2_tp2_pp2_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_8.86G_dp2_tp4_pp2_acc128_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_8.86G_dp2_tp4_pp2_acc128_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..479fa1ab39ed29b9d3e3e4aca63cab5d11584e77 --- /dev/null +++ b/scripts/run_8.86G_dp2_tp4_pp2_acc128_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_8.86G_dp2_tp4_pp2_acc128_mbs1_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=2 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=INFO # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_8.86G_dp2_tp4_pp2_acc128_mbs1_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_8.86G_dp2_tp4_pp2_acc128_mbs1_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_8.86G_dp2_tp4_pp2_acc128_mbs1_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_8.86G_dp8_tp8_pp2_acc1_mbs32_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_8.86G_dp8_tp8_pp2_acc1_mbs32_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..8c55c3fdec37353c6989c4a970cffa81748663b2 --- /dev/null +++ b/scripts/run_8.86G_dp8_tp8_pp2_acc1_mbs32_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,73 @@ +#!/bin/bash + +#SBATCH --job-name=bench_8.86G_dp8_tp8_pp2_acc1_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=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_8.86G_dp8_tp8_pp2_acc1_mbs32_seq4096_zero0_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_80G_dp128_tp2_pp1_acc2_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_80G_dp128_tp2_pp1_acc2_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..59fff17ae11767caaa43f06c845307b8f1c0b143 --- /dev/null +++ b/scripts/run_80G_dp128_tp2_pp1_acc2_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,124 @@ +#!/bin/bash +#SBATCH --job-name=bench_80G_dp128_tp2_pp1_acc2_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_80G_dp128_tp2_pp1_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_80G_dp128_tp2_pp1_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_80G_dp128_tp2_pp1_acc2_mbs1_seq4096_zero0_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_80G_dp16_tp32_pp1_acc1_mbs16_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_80G_dp16_tp32_pp1_acc1_mbs16_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..7efd008a64de10a60937e30222c973ad93b3972e --- /dev/null +++ b/scripts/run_80G_dp16_tp32_pp1_acc1_mbs16_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_80G_dp16_tp32_pp1_acc1_mbs16_seq4096_zero0_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:40:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=64 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_80G_dp16_tp32_pp1_acc1_mbs16_seq4096_zero0_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp16_tp32_pp1_acc1_mbs16_seq4096_zero0_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp16_tp32_pp1_acc1_mbs16_seq4096_zero0_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_80G_dp1_tp1_pp8_acc256_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_80G_dp1_tp1_pp8_acc256_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..69adaafd824bc8a466c644cff5d0e28f90ab7d57 --- /dev/null +++ b/scripts/run_80G_dp1_tp1_pp8_acc256_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,73 @@ +#!/bin/bash + +#SBATCH --job-name=bench_80G_dp1_tp1_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=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_dp1_tp1_pp8_acc256_mbs1_seq4096_zero0_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_80G_dp1_tp2_pp4_acc16_mbs16_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_80G_dp1_tp2_pp4_acc16_mbs16_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..998eeb8382c4ed70c7b66150285d90c582da1655 --- /dev/null +++ b/scripts/run_80G_dp1_tp2_pp4_acc16_mbs16_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,73 @@ +#!/bin/bash + +#SBATCH --job-name=bench_80G_dp1_tp2_pp4_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=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_dp1_tp2_pp4_acc16_mbs16_seq4096_zero0_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_80G_dp1_tp8_pp32_acc256_mbs1_seq4096_zero0_tpmodeRED_vocab131k_gqa8.sh b/scripts/run_80G_dp1_tp8_pp32_acc256_mbs1_seq4096_zero0_tpmodeRED_vocab131k_gqa8.sh new file mode 100644 index 0000000000000000000000000000000000000000..4f829cf47dc4ade655410681f9997f131df29eb0 --- /dev/null +++ b/scripts/run_80G_dp1_tp8_pp32_acc256_mbs1_seq4096_zero0_tpmodeRED_vocab131k_gqa8.sh @@ -0,0 +1,161 @@ +#!/bin/bash +#SBATCH --job-name=bench_80G_dp1_tp8_pp32_acc256_mbs1_seq4096_zero0_tpmodeRED_vocab131k_gqa8 # Job name +#SBATCH --time=00:40:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=32 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e +echo "Running script: $0" + + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_80G_dp1_tp8_pp32_acc256_mbs1_seq4096_zero0_tpmodeRED_vocab131k_gqa8" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp1_tp8_pp32_acc256_mbs1_seq4096_zero0_tpmodeRED_vocab131k_gqa8.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp1_tp8_pp32_acc256_mbs1_seq4096_zero0_tpmodeRED_vocab131k_gqa8.yaml +fi diff --git a/scripts/run_80G_dp256_tp1_pp1_acc1_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_80G_dp256_tp1_pp1_acc1_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..eda9e556175bcdc014e4a0a8cf2cc3920a685a2b --- /dev/null +++ b/scripts/run_80G_dp256_tp1_pp1_acc1_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,124 @@ +#!/bin/bash +#SBATCH --job-name=bench_80G_dp256_tp1_pp1_acc1_mbs1_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=32 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$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_dp256_tp1_pp1_acc1_mbs1_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp256_tp1_pp1_acc1_mbs1_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp256_tp1_pp1_acc1_mbs1_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_80G_dp2_tp4_pp16_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_80G_dp2_tp4_pp16_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..d78917f75f0559a2050538c7ff0165154c83f3cd --- /dev/null +++ b/scripts/run_80G_dp2_tp4_pp16_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_80G_dp2_tp4_pp16_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=16 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=INFO # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_80G_dp2_tp4_pp16_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp2_tp4_pp16_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp2_tp4_pp16_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_80G_dp32_tp4_pp2_acc8_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_80G_dp32_tp4_pp2_acc8_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..3f9a7be9e6691a0e3caa90a673313811373e7adc --- /dev/null +++ b/scripts/run_80G_dp32_tp4_pp2_acc8_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_80G_dp32_tp4_pp2_acc8_mbs1_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:40:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=32 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_80G_dp32_tp4_pp2_acc8_mbs1_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp32_tp4_pp2_acc8_mbs1_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp32_tp4_pp2_acc8_mbs1_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_80G_dp4_tp16_pp2_acc2_mbs32_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_80G_dp4_tp16_pp2_acc2_mbs32_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..2fffed7dbcaf8997dee8d052c905749615e06b6c --- /dev/null +++ b/scripts/run_80G_dp4_tp16_pp2_acc2_mbs32_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_80G_dp4_tp16_pp2_acc2_mbs32_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=16 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=INFO # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_80G_dp4_tp16_pp2_acc2_mbs32_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp4_tp16_pp2_acc2_mbs32_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp4_tp16_pp2_acc2_mbs32_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_80G_dp4_tp16_pp2_acc8_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_80G_dp4_tp16_pp2_acc8_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..461fdd8a2a2a1e2a8aa990c048869e0c400c08f8 --- /dev/null +++ b/scripts/run_80G_dp4_tp16_pp2_acc8_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_80G_dp4_tp16_pp2_acc8_mbs8_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=16 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=INFO # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_80G_dp4_tp16_pp2_acc8_mbs8_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp4_tp16_pp2_acc8_mbs8_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp4_tp16_pp2_acc8_mbs8_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_80G_dp4_tp2_pp1_acc64_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_80G_dp4_tp2_pp1_acc64_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..d39a97dde137a4acd1cb8100447ec35a4412fb84 --- /dev/null +++ b/scripts/run_80G_dp4_tp2_pp1_acc64_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,73 @@ +#!/bin/bash + +#SBATCH --job-name=bench_80G_dp4_tp2_pp1_acc64_mbs1_seq4096_zero0_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high +#SBATCH --exclude=ip-26-0-160-192,ip-26-0-171-102 + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=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_dp4_tp2_pp1_acc64_mbs1_seq4096_zero0_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_80G_dp64_tp1_pp1_acc4_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_80G_dp64_tp1_pp1_acc4_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..cda8fe42c71ad897cc670f2f6e81c9f7ca9b4efb --- /dev/null +++ b/scripts/run_80G_dp64_tp1_pp1_acc4_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,73 @@ +#!/bin/bash + +#SBATCH --job-name=bench_80G_dp64_tp1_pp1_acc4_mbs1_seq4096_zero0_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high +#SBATCH --exclude=ip-26-0-160-192,ip-26-0-171-102 + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=8 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +# Disable wandb +export WANDB_MODE=disabled + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + + +# Print GPU topology information +echo "=== GPU Topology ===" +nvidia-smi topo -m +echo "==================" + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp64_tp1_pp1_acc4_mbs1_seq4096_zero0_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_80G_dp64_tp2_pp1_acc1_mbs4_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_80G_dp64_tp2_pp1_acc1_mbs4_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..49a682a612e3831908f7fb4be5fddc572f7521d8 --- /dev/null +++ b/scripts/run_80G_dp64_tp2_pp1_acc1_mbs4_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,124 @@ +#!/bin/bash +#SBATCH --job-name=bench_80G_dp64_tp2_pp1_acc1_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_80G_dp64_tp2_pp1_acc1_mbs4_seq4096_zero0_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp64_tp2_pp1_acc1_mbs4_seq4096_zero0_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp64_tp2_pp1_acc1_mbs4_seq4096_zero0_tpmodeRED_vocab131k.yaml +fi