diff --git a/scripts/run_1.14G_dp128_tp4_pp1_acc1_mbs4_seq8192_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp128_tp4_pp1_acc1_mbs4_seq8192_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..8d3b6f5bb4cd8e1a0242d640202a430125f1c429 --- /dev/null +++ b/scripts/run_1.14G_dp128_tp4_pp1_acc1_mbs4_seq8192_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp128_tp4_pp1_acc1_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=64 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.14G_dp128_tp4_pp1_acc1_mbs4_seq8192_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp16_tp16_pp1_acc32_mbs4_seq2048_zero1_tpmodeALL_vocab32k.sh b/scripts/run_1.14G_dp16_tp16_pp1_acc32_mbs4_seq2048_zero1_tpmodeALL_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..efee6f247d01ab2a8ff59dd862624edcc1a3c130 --- /dev/null +++ b/scripts/run_1.14G_dp16_tp16_pp1_acc32_mbs4_seq2048_zero1_tpmodeALL_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp16_tp16_pp1_acc32_mbs4_seq2048_zero1_tpmodeALL_vocab32k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=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_acc32_mbs4_seq2048_zero1_tpmodeALL_vocab32k.yaml diff --git a/scripts/run_1.14G_dp2_tp128_pp1_acc32_mbs32_seq2048_zero1_tpmodeALL_vocab32k.sh b/scripts/run_1.14G_dp2_tp128_pp1_acc32_mbs32_seq2048_zero1_tpmodeALL_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..72e863d6635a01b90ce066d11a9e1d2dfd43596b --- /dev/null +++ b/scripts/run_1.14G_dp2_tp128_pp1_acc32_mbs32_seq2048_zero1_tpmodeALL_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp2_tp128_pp1_acc32_mbs32_seq2048_zero1_tpmodeALL_vocab32k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=32 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.14G_dp2_tp128_pp1_acc32_mbs32_seq2048_zero1_tpmodeALL_vocab32k.yaml diff --git a/scripts/run_1.14G_dp2_tp128_pp1_acc4_mbs4_seq32768_zero1_tpmodeALL_vocab32k.sh b/scripts/run_1.14G_dp2_tp128_pp1_acc4_mbs4_seq32768_zero1_tpmodeALL_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..e3f8d403979e4e4c178b6f714b177c81c7ca4697 --- /dev/null +++ b/scripts/run_1.14G_dp2_tp128_pp1_acc4_mbs4_seq32768_zero1_tpmodeALL_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp2_tp128_pp1_acc4_mbs4_seq32768_zero1_tpmodeALL_vocab32k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=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_acc4_mbs4_seq32768_zero1_tpmodeALL_vocab32k.yaml diff --git a/scripts/run_1.14G_dp2_tp256_pp1_acc16_mbs4_seq32768_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp2_tp256_pp1_acc16_mbs4_seq32768_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..a2a088ac01f13909039009c1b8efeb9573ba4295 --- /dev/null +++ b/scripts/run_1.14G_dp2_tp256_pp1_acc16_mbs4_seq32768_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp2_tp256_pp1_acc16_mbs4_seq32768_zero1_tpmodeRED_vocab32k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=64 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.14G_dp2_tp256_pp1_acc16_mbs4_seq32768_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp2_tp64_pp1_acc128_mbs2_seq2048_zero1_tpmodeALL_vocab32k.sh b/scripts/run_1.14G_dp2_tp64_pp1_acc128_mbs2_seq2048_zero1_tpmodeALL_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..9a2cca82b37c4a4eeda3f7f79755c6c8a4b38f07 --- /dev/null +++ b/scripts/run_1.14G_dp2_tp64_pp1_acc128_mbs2_seq2048_zero1_tpmodeALL_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp2_tp64_pp1_acc128_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=16 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.14G_dp2_tp64_pp1_acc128_mbs2_seq2048_zero1_tpmodeALL_vocab32k.yaml diff --git a/scripts/run_1.14G_dp32_tp2_pp1_acc16_mbs1_seq2048_zero1_tpmodeALL_vocab32k.sh b/scripts/run_1.14G_dp32_tp2_pp1_acc16_mbs1_seq2048_zero1_tpmodeALL_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..cf67fce988ed76978a3cc56afba97cf6f8fa548f --- /dev/null +++ b/scripts/run_1.14G_dp32_tp2_pp1_acc16_mbs1_seq2048_zero1_tpmodeALL_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp32_tp2_pp1_acc16_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=8 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.14G_dp32_tp2_pp1_acc16_mbs1_seq2048_zero1_tpmodeALL_vocab32k.yaml diff --git a/scripts/run_1.14G_dp4_tp16_pp1_acc4_mbs32_seq8192_zero1_tpmodeALL_vocab32k.sh b/scripts/run_1.14G_dp4_tp16_pp1_acc4_mbs32_seq8192_zero1_tpmodeALL_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..52a1668a97fe4881bf3e28a75d3013f37222b47f --- /dev/null +++ b/scripts/run_1.14G_dp4_tp16_pp1_acc4_mbs32_seq8192_zero1_tpmodeALL_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp4_tp16_pp1_acc4_mbs32_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=8 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.14G_dp4_tp16_pp1_acc4_mbs32_seq8192_zero1_tpmodeALL_vocab32k.yaml diff --git a/scripts/run_1.14G_dp4_tp4_pp1_acc32_mbs1_seq8192_zero1_tpmodeALL_vocab32k.sh b/scripts/run_1.14G_dp4_tp4_pp1_acc32_mbs1_seq8192_zero1_tpmodeALL_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..c9d12a336ee5399103e6c03a67cd4f459df6d2bf --- /dev/null +++ b/scripts/run_1.14G_dp4_tp4_pp1_acc32_mbs1_seq8192_zero1_tpmodeALL_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp4_tp4_pp1_acc32_mbs1_seq8192_zero1_tpmodeALL_vocab32k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=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_acc32_mbs1_seq8192_zero1_tpmodeALL_vocab32k.yaml diff --git a/scripts/run_1.14G_dp4_tp4_pp1_acc8_mbs64_seq2048_zero1_tpmodeALL_vocab32k.sh b/scripts/run_1.14G_dp4_tp4_pp1_acc8_mbs64_seq2048_zero1_tpmodeALL_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..40c5a99b4e0eb4c24d269fbe743da018428813db --- /dev/null +++ b/scripts/run_1.14G_dp4_tp4_pp1_acc8_mbs64_seq2048_zero1_tpmodeALL_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp4_tp4_pp1_acc8_mbs64_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_dp4_tp4_pp1_acc8_mbs64_seq2048_zero1_tpmodeALL_vocab32k.yaml diff --git a/scripts/run_1.14G_dp64_tp2_pp1_acc2_mbs1_seq8192_zero1_tpmodeALL_vocab32k.sh b/scripts/run_1.14G_dp64_tp2_pp1_acc2_mbs1_seq8192_zero1_tpmodeALL_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..412a18edd9279b39506c72d8cdc52a321f3a2484 --- /dev/null +++ b/scripts/run_1.14G_dp64_tp2_pp1_acc2_mbs1_seq8192_zero1_tpmodeALL_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp64_tp2_pp1_acc2_mbs1_seq8192_zero1_tpmodeALL_vocab32k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=16 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.14G_dp64_tp2_pp1_acc2_mbs1_seq8192_zero1_tpmodeALL_vocab32k.yaml diff --git a/scripts/run_1.14G_dp8_tp1_pp2_acc64_mbs1_seq8192_zero1_tpmodeRED_vocab32k.sh b/scripts/run_1.14G_dp8_tp1_pp2_acc64_mbs1_seq8192_zero1_tpmodeRED_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..5344e1aa23e9cc22d7c4a77b827ba473fc8e1d9d --- /dev/null +++ b/scripts/run_1.14G_dp8_tp1_pp2_acc64_mbs1_seq8192_zero1_tpmodeRED_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp8_tp1_pp2_acc64_mbs1_seq8192_zero1_tpmodeRED_vocab32k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=2 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.14G_dp8_tp1_pp2_acc64_mbs1_seq8192_zero1_tpmodeRED_vocab32k.yaml diff --git a/scripts/run_1.14G_dp8_tp2_pp1_acc1_mbs16_seq32768_zero1_tpmodeALL_vocab32k.sh b/scripts/run_1.14G_dp8_tp2_pp1_acc1_mbs16_seq32768_zero1_tpmodeALL_vocab32k.sh new file mode 100644 index 0000000000000000000000000000000000000000..cf1e5851555b4b24cb9adef32eb3fee95c43dcb3 --- /dev/null +++ b/scripts/run_1.14G_dp8_tp2_pp1_acc1_mbs16_seq32768_zero1_tpmodeALL_vocab32k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.14G_dp8_tp2_pp1_acc1_mbs16_seq32768_zero1_tpmodeALL_vocab32k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=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_acc1_mbs16_seq32768_zero1_tpmodeALL_vocab32k.yaml diff --git a/scripts/run_1.34G_dp128_tp1_pp2_acc2_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp128_tp1_pp2_acc2_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..66524a1aa8abe505237f2aaf8454162227367793 --- /dev/null +++ b/scripts/run_1.34G_dp128_tp1_pp2_acc2_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_1.34G_dp128_tp1_pp2_acc2_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_1.34G_dp128_tp1_pp2_acc2_mbs1_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp128_tp1_pp2_acc2_mbs1_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp128_tp1_pp2_acc2_mbs1_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_1.34G_dp128_tp2_pp1_acc1_mbs4_seq2048_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp128_tp2_pp1_acc1_mbs4_seq2048_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..963c3d34305a246e63dd6654284545275d44bf8b --- /dev/null +++ b/scripts/run_1.34G_dp128_tp2_pp1_acc1_mbs4_seq2048_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp128_tp2_pp1_acc1_mbs4_seq2048_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=32 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp128_tp2_pp1_acc1_mbs4_seq2048_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp16_tp2_pp2_acc4_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp16_tp2_pp2_acc4_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..0479bba11b0ea9d3e814d1a10f99609c2dd96885 --- /dev/null +++ b/scripts/run_1.34G_dp16_tp2_pp2_acc4_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_1.34G_dp16_tp2_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=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_dp16_tp2_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_1.34G_dp16_tp2_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_1.34G_dp16_tp2_pp2_acc4_mbs4_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_1.34G_dp16_tp4_pp1_acc2_mbs4_seq32768_zero1_tpmodeALL_vocab131k.sh b/scripts/run_1.34G_dp16_tp4_pp1_acc2_mbs4_seq32768_zero1_tpmodeALL_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..626d8262f7afd37d9a7ac5559d06045e330ae070 --- /dev/null +++ b/scripts/run_1.34G_dp16_tp4_pp1_acc2_mbs4_seq32768_zero1_tpmodeALL_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp16_tp4_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=8 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp16_tp4_pp1_acc2_mbs4_seq32768_zero1_tpmodeALL_vocab131k.yaml diff --git a/scripts/run_1.34G_dp1_tp16_pp1_acc8_mbs32_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp1_tp16_pp1_acc8_mbs32_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..1917a9199de5aa7b9da5cbdef95629f272962c20 --- /dev/null +++ b/scripts/run_1.34G_dp1_tp16_pp1_acc8_mbs32_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,73 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp1_tp16_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=2 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +# Disable wandb +export WANDB_MODE=disabled + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + + +# Print GPU topology information +echo "=== GPU Topology ===" +nvidia-smi topo -m +echo "==================" + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp1_tp16_pp1_acc8_mbs32_seq4096_zero0_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp1_tp2_pp16_acc256_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp1_tp2_pp16_acc256_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..c93ed8228ba92d4523ab3f668575e377d1914a2e --- /dev/null +++ b/scripts/run_1.34G_dp1_tp2_pp16_acc256_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_1.34G_dp1_tp2_pp16_acc256_mbs1_seq4096_zero0_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:40:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=4 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_1.34G_dp1_tp2_pp16_acc256_mbs1_seq4096_zero0_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp1_tp2_pp16_acc256_mbs1_seq4096_zero0_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp1_tp2_pp16_acc256_mbs1_seq4096_zero0_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_1.34G_dp1_tp4_pp2_acc8_mbs32_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp1_tp4_pp2_acc8_mbs32_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..b039f2327f567754812b74d871433bf34b54f515 --- /dev/null +++ b/scripts/run_1.34G_dp1_tp4_pp2_acc8_mbs32_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,73 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp1_tp4_pp2_acc8_mbs32_seq4096_zero0_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high +#SBATCH --exclude=ip-26-0-160-192,ip-26-0-171-102 + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=1 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +# Disable wandb +export WANDB_MODE=disabled + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + + +# Print GPU topology information +echo "=== GPU Topology ===" +nvidia-smi topo -m +echo "==================" + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp1_tp4_pp2_acc8_mbs32_seq4096_zero0_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp2_tp1_pp4_acc128_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp2_tp1_pp4_acc128_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..6ed2aed62dc3320863a506d96ade40563e109560 --- /dev/null +++ b/scripts/run_1.34G_dp2_tp1_pp4_acc128_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_1.34G_dp2_tp1_pp4_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=1 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=INFO # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_1.34G_dp2_tp1_pp4_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_1.34G_dp2_tp1_pp4_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_1.34G_dp2_tp1_pp4_acc128_mbs1_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_1.34G_dp2_tp32_pp1_acc8_mbs32_seq2048_zero1_tpmodeALL_vocab131k.sh b/scripts/run_1.34G_dp2_tp32_pp1_acc8_mbs32_seq2048_zero1_tpmodeALL_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..6179ea2385908732ecc87562da9a4b113ac0789b --- /dev/null +++ b/scripts/run_1.34G_dp2_tp32_pp1_acc8_mbs32_seq2048_zero1_tpmodeALL_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp2_tp32_pp1_acc8_mbs32_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=8 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp2_tp32_pp1_acc8_mbs32_seq2048_zero1_tpmodeALL_vocab131k.yaml diff --git a/scripts/run_1.34G_dp2_tp4_pp1_acc32_mbs2_seq32768_zero1_tpmodeALL_vocab131k.sh b/scripts/run_1.34G_dp2_tp4_pp1_acc32_mbs2_seq32768_zero1_tpmodeALL_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..04562ed53d68166a4d071052ecf678e9733f688d --- /dev/null +++ b/scripts/run_1.34G_dp2_tp4_pp1_acc32_mbs2_seq32768_zero1_tpmodeALL_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp2_tp4_pp1_acc32_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=1 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp2_tp4_pp1_acc32_mbs2_seq32768_zero1_tpmodeALL_vocab131k.yaml diff --git a/scripts/run_1.34G_dp2_tp8_pp1_acc8_mbs16_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp2_tp8_pp1_acc8_mbs16_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..2de1543d20599c29990d4cc3aaf4ac037c02b595 --- /dev/null +++ b/scripts/run_1.34G_dp2_tp8_pp1_acc8_mbs16_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,124 @@ +#!/bin/bash +#SBATCH --job-name=bench_1.34G_dp2_tp8_pp1_acc8_mbs16_seq4096_zero0_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=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_acc8_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_1.34G_dp2_tp8_pp1_acc8_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_1.34G_dp2_tp8_pp1_acc8_mbs16_seq4096_zero0_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_1.34G_dp32_tp16_pp1_acc16_mbs1_seq2048_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp32_tp16_pp1_acc16_mbs1_seq2048_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..de586e1c9dee612d88b4285b3b5032c6914d1887 --- /dev/null +++ b/scripts/run_1.34G_dp32_tp16_pp1_acc16_mbs1_seq2048_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp32_tp16_pp1_acc16_mbs1_seq2048_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=64 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp32_tp16_pp1_acc16_mbs1_seq2048_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp32_tp16_pp1_acc1_mbs16_seq2048_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp32_tp16_pp1_acc1_mbs16_seq2048_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..ea9c18570c98246296a615bad37bb73aa8e5a9cb --- /dev/null +++ b/scripts/run_1.34G_dp32_tp16_pp1_acc1_mbs16_seq2048_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp32_tp16_pp1_acc1_mbs16_seq2048_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=64 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp32_tp16_pp1_acc1_mbs16_seq2048_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp32_tp16_pp1_acc1_mbs8_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp32_tp16_pp1_acc1_mbs8_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..8b66a666b0a7f8129b905ecee5470a801096e9f0 --- /dev/null +++ b/scripts/run_1.34G_dp32_tp16_pp1_acc1_mbs8_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,161 @@ +#!/bin/bash +#SBATCH --job-name=bench_1.34G_dp32_tp16_pp1_acc1_mbs8_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_1.34G_dp32_tp16_pp1_acc1_mbs8_seq4096_zero0_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp32_tp16_pp1_acc1_mbs8_seq4096_zero0_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp32_tp16_pp1_acc1_mbs8_seq4096_zero0_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_1.34G_dp32_tp16_pp1_acc4_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp32_tp16_pp1_acc4_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..69e1d540f085a510e8a6782306be913d875035b2 --- /dev/null +++ b/scripts/run_1.34G_dp32_tp16_pp1_acc4_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_1.34G_dp32_tp16_pp1_acc4_mbs2_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:40:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=64 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_1.34G_dp32_tp16_pp1_acc4_mbs2_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp32_tp16_pp1_acc4_mbs2_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp32_tp16_pp1_acc4_mbs2_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_1.34G_dp32_tp1_pp2_acc16_mbs1_seq2048_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp32_tp1_pp2_acc16_mbs1_seq2048_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..df14b1accdf28cf46c0a9915008d54a27339f886 --- /dev/null +++ b/scripts/run_1.34G_dp32_tp1_pp2_acc16_mbs1_seq2048_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp32_tp1_pp2_acc16_mbs1_seq2048_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=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_dp32_tp1_pp2_acc16_mbs1_seq2048_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp32_tp1_pp2_acc1_mbs16_seq2048_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp32_tp1_pp2_acc1_mbs16_seq2048_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..f0d7b3706ddf7c54eaff4536b82fccf9411c75d6 --- /dev/null +++ b/scripts/run_1.34G_dp32_tp1_pp2_acc1_mbs16_seq2048_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp32_tp1_pp2_acc1_mbs16_seq2048_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=8 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp32_tp1_pp2_acc1_mbs16_seq2048_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp32_tp2_pp1_acc2_mbs2_seq32768_zero1_tpmodeALL_vocab131k.sh b/scripts/run_1.34G_dp32_tp2_pp1_acc2_mbs2_seq32768_zero1_tpmodeALL_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..c3c6edfa3aed906ef1cad02ff21ca353317f9b18 --- /dev/null +++ b/scripts/run_1.34G_dp32_tp2_pp1_acc2_mbs2_seq32768_zero1_tpmodeALL_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp32_tp2_pp1_acc2_mbs2_seq32768_zero1_tpmodeALL_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=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_dp32_tp2_pp1_acc2_mbs2_seq32768_zero1_tpmodeALL_vocab131k.yaml diff --git a/scripts/run_1.34G_dp32_tp8_pp1_acc128_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp32_tp8_pp1_acc128_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..f9517ccbe962b98e846c14da38b4975c3f4059c4 --- /dev/null +++ b/scripts/run_1.34G_dp32_tp8_pp1_acc128_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,161 @@ +#!/bin/bash +#SBATCH --job-name=bench_1.34G_dp32_tp8_pp1_acc128_mbs1_seq4096_zero0_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:40:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=normal + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=32 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e +echo "Running script: $0" + + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_1.34G_dp32_tp8_pp1_acc128_mbs1_seq4096_zero0_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp32_tp8_pp1_acc128_mbs1_seq4096_zero0_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp32_tp8_pp1_acc128_mbs1_seq4096_zero0_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_1.34G_dp4_tp128_pp1_acc8_mbs4_seq32768_zero1_tpmodeALL_vocab131k.sh b/scripts/run_1.34G_dp4_tp128_pp1_acc8_mbs4_seq32768_zero1_tpmodeALL_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..2c475b1b210c6c3821ac71953a5cb130e1f20e3d --- /dev/null +++ b/scripts/run_1.34G_dp4_tp128_pp1_acc8_mbs4_seq32768_zero1_tpmodeALL_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp4_tp128_pp1_acc8_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=64 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp4_tp128_pp1_acc8_mbs4_seq32768_zero1_tpmodeALL_vocab131k.yaml diff --git a/scripts/run_1.34G_dp4_tp16_pp1_acc128_mbs1_seq2048_zero1_tpmodeALL_vocab131k.sh b/scripts/run_1.34G_dp4_tp16_pp1_acc128_mbs1_seq2048_zero1_tpmodeALL_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..43d0d21638bb6659f2436eaf3d2481163fdb1306 --- /dev/null +++ b/scripts/run_1.34G_dp4_tp16_pp1_acc128_mbs1_seq2048_zero1_tpmodeALL_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp4_tp16_pp1_acc128_mbs1_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=8 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp4_tp16_pp1_acc128_mbs1_seq2048_zero1_tpmodeALL_vocab131k.yaml diff --git a/scripts/run_1.34G_dp4_tp1_pp2_acc1_mbs32_seq8192_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp4_tp1_pp2_acc1_mbs32_seq8192_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..ae4bb1ef4d92930ee6b2d39934449ceca767ca4a --- /dev/null +++ b/scripts/run_1.34G_dp4_tp1_pp2_acc1_mbs32_seq8192_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp4_tp1_pp2_acc1_mbs32_seq8192_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=1 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp4_tp1_pp2_acc1_mbs32_seq8192_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp4_tp1_pp4_acc64_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp4_tp1_pp4_acc64_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..c80b9e2b7f17204ff038a828e677341b94838445 --- /dev/null +++ b/scripts/run_1.34G_dp4_tp1_pp4_acc64_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_1.34G_dp4_tp1_pp4_acc64_mbs1_seq4096_zero0_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:40:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=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 * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_1.34G_dp4_tp1_pp4_acc64_mbs1_seq4096_zero0_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp4_tp1_pp4_acc64_mbs1_seq4096_zero0_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp4_tp1_pp4_acc64_mbs1_seq4096_zero0_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_1.34G_dp4_tp32_pp1_acc2_mbs256_seq2048_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp4_tp32_pp1_acc2_mbs256_seq2048_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..d3ce09546cb796bead5407cfcb74e88d0c273eb8 --- /dev/null +++ b/scripts/run_1.34G_dp4_tp32_pp1_acc2_mbs256_seq2048_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp4_tp32_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=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_mbs256_seq2048_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp4_tp32_pp1_acc32_mbs1_seq32768_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp4_tp32_pp1_acc32_mbs1_seq32768_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..50d8f9d012f38d063992e6cedd85c808f589741f --- /dev/null +++ b/scripts/run_1.34G_dp4_tp32_pp1_acc32_mbs1_seq32768_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp4_tp32_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=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_seq32768_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp4_tp4_pp1_acc128_mbs4_seq2048_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp4_tp4_pp1_acc128_mbs4_seq2048_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..b2dfb4d009e714539753c29f63260888b3768466 --- /dev/null +++ b/scripts/run_1.34G_dp4_tp4_pp1_acc128_mbs4_seq2048_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp4_tp4_pp1_acc128_mbs4_seq2048_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=2 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp4_tp4_pp1_acc128_mbs4_seq2048_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp4_tp4_pp1_acc8_mbs8_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp4_tp4_pp1_acc8_mbs8_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..1650a34954484daa041eceaa9cd3619bb46ea63e --- /dev/null +++ b/scripts/run_1.34G_dp4_tp4_pp1_acc8_mbs8_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,124 @@ +#!/bin/bash +#SBATCH --job-name=bench_1.34G_dp4_tp4_pp1_acc8_mbs8_seq4096_zero0_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=2 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_1.34G_dp4_tp4_pp1_acc8_mbs8_seq4096_zero0_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp4_tp4_pp1_acc8_mbs8_seq4096_zero0_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp4_tp4_pp1_acc8_mbs8_seq4096_zero0_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_1.34G_dp64_tp2_pp1_acc1_mbs2_seq8192_zero1_tpmodeALL_vocab131k.sh b/scripts/run_1.34G_dp64_tp2_pp1_acc1_mbs2_seq8192_zero1_tpmodeALL_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..332abc10b4b6e68855628a98d9e05e9608eea2e7 --- /dev/null +++ b/scripts/run_1.34G_dp64_tp2_pp1_acc1_mbs2_seq8192_zero1_tpmodeALL_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp64_tp2_pp1_acc1_mbs2_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_dp64_tp2_pp1_acc1_mbs2_seq8192_zero1_tpmodeALL_vocab131k.yaml diff --git a/scripts/run_1.34G_dp64_tp2_pp1_acc32_mbs1_seq2048_zero1_tpmodeALL_vocab131k.sh b/scripts/run_1.34G_dp64_tp2_pp1_acc32_mbs1_seq2048_zero1_tpmodeALL_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..aeab60acc462efd4e8848bd1490c31b75ad9410b --- /dev/null +++ b/scripts/run_1.34G_dp64_tp2_pp1_acc32_mbs1_seq2048_zero1_tpmodeALL_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp64_tp2_pp1_acc32_mbs1_seq2048_zero1_tpmodeALL_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=16 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp64_tp2_pp1_acc32_mbs1_seq2048_zero1_tpmodeALL_vocab131k.yaml diff --git a/scripts/run_1.34G_dp64_tp4_pp1_acc4_mbs8_seq2048_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp64_tp4_pp1_acc4_mbs8_seq2048_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..0d0238a8083b9a53279e60eba26d2f625b208a21 --- /dev/null +++ b/scripts/run_1.34G_dp64_tp4_pp1_acc4_mbs8_seq2048_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp64_tp4_pp1_acc4_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=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_dp64_tp4_pp1_acc4_mbs8_seq2048_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_1.34G_dp8_tp2_pp1_acc2_mbs2_seq32768_zero1_tpmodeALL_vocab131k.sh b/scripts/run_1.34G_dp8_tp2_pp1_acc2_mbs2_seq32768_zero1_tpmodeALL_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..052739742d7308d8360025a291b979a26905fe14 --- /dev/null +++ b/scripts/run_1.34G_dp8_tp2_pp1_acc2_mbs2_seq32768_zero1_tpmodeALL_vocab131k.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_1.34G_dp8_tp2_pp1_acc2_mbs2_seq32768_zero1_tpmodeALL_vocab131k # Job name +#SBATCH --time=00:02:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=2 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_1.34G_dp8_tp2_pp1_acc2_mbs2_seq32768_zero1_tpmodeALL_vocab131k.yaml diff --git a/scripts/run_1.34G_dp8_tp4_pp4_acc16_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_1.34G_dp8_tp4_pp4_acc16_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..8c533ee18d0285f42bcc7eb9799a815d42012057 --- /dev/null +++ b/scripts/run_1.34G_dp8_tp4_pp4_acc16_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_1.34G_dp8_tp4_pp4_acc16_mbs2_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:40:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=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_1.34G_dp8_tp4_pp4_acc16_mbs2_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp8_tp4_pp4_acc16_mbs2_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_1.34G_dp8_tp4_pp4_acc16_mbs2_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_3.27G_dp2_tp16_pp16_acc1_mbs1_seq2048_zero0_tpmodeRED_l28_h3072_heads24.sh b/scripts/run_3.27G_dp2_tp16_pp16_acc1_mbs1_seq2048_zero0_tpmodeRED_l28_h3072_heads24.sh new file mode 100644 index 0000000000000000000000000000000000000000..93406b084a69b401e88d56c36d9aaed6079d7e47 --- /dev/null +++ b/scripts/run_3.27G_dp2_tp16_pp16_acc1_mbs1_seq2048_zero0_tpmodeRED_l28_h3072_heads24.sh @@ -0,0 +1,68 @@ +#!/bin/bash + +#SBATCH --job-name=bench_3.27G_dp2_tp16_pp16_acc1_mbs1_seq2048_zero0_tpmodeRED_l28_h3072_heads24 # Job name +#SBATCH --time=00:15:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=64 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# # Disable EFA by changing the provider to tcp +# export FI_PROVIDER=tcp + +# # Optionally, you can also unset these EFA-related variables +# unset FI_EFA_FORK_SAFE +# unset FI_EFA_ENABLE_SHM_TRANSFER + +# # If you want to ensure NCCL uses TCP +# export NCCL_IB_DISABLE=1 +# export NCCL_SOCKET_IFNAME=eth0 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_3.27G_dp2_tp16_pp16_acc1_mbs1_seq2048_zero0_tpmodeRED_l28_h3072_heads24.yaml diff --git a/scripts/run_3.56G_dp32_tp1_pp1_acc1_mbs2_seq2048_zero0_l28_h3072_heads24.sh b/scripts/run_3.56G_dp32_tp1_pp1_acc1_mbs2_seq2048_zero0_l28_h3072_heads24.sh new file mode 100644 index 0000000000000000000000000000000000000000..bc8d07ca4899ace67cd8e0ed340f2af5611f5ac3 --- /dev/null +++ b/scripts/run_3.56G_dp32_tp1_pp1_acc1_mbs2_seq2048_zero0_l28_h3072_heads24.sh @@ -0,0 +1,57 @@ +#!/bin/bash + +#SBATCH --job-name=bench_3.56G_dp32_tp1_pp1_acc1_mbs2_seq2048_zero0_l28_h3072_heads24 # Job name +#SBATCH --time=00:15:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=4 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +# export NCCL_DEBUG=INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + run_train.py \ + --config-file benchmark/configs/config_3.56G_dp32_tp1_pp1_acc1_mbs2_seq2048_zero0_l28_h3072_heads24.yaml diff --git a/scripts/run_3.57G_dp128_tp1_pp1_acc1_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_3.57G_dp128_tp1_pp1_acc1_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..9a18107db460baed31b59c177132fc80377fc1de --- /dev/null +++ b/scripts/run_3.57G_dp128_tp1_pp1_acc1_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,124 @@ +#!/bin/bash +#SBATCH --job-name=bench_3.57G_dp128_tp1_pp1_acc1_mbs2_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=16 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_3.57G_dp128_tp1_pp1_acc1_mbs2_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_3.57G_dp128_tp1_pp1_acc1_mbs2_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_3.57G_dp128_tp1_pp1_acc1_mbs2_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_3.57G_dp1_tp1_pp32_acc8_mbs32_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_3.57G_dp1_tp1_pp32_acc8_mbs32_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..67144b13c5683eedb9580be5e997fa17b5b323eb --- /dev/null +++ b/scripts/run_3.57G_dp1_tp1_pp32_acc8_mbs32_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,73 @@ +#!/bin/bash + +#SBATCH --job-name=bench_3.57G_dp1_tp1_pp32_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=4 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +# Disable wandb +export WANDB_MODE=disabled + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + + +# Print GPU topology information +echo "=== GPU Topology ===" +nvidia-smi topo -m +echo "==================" + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_3.57G_dp1_tp1_pp32_acc8_mbs32_seq4096_zero0_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_3.57G_dp1_tp1_pp8_acc4_mbs64_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_3.57G_dp1_tp1_pp8_acc4_mbs64_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..bbc4177253f5104af2f0fb93dbac986d34c01b65 --- /dev/null +++ b/scripts/run_3.57G_dp1_tp1_pp8_acc4_mbs64_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,73 @@ +#!/bin/bash + +#SBATCH --job-name=bench_3.57G_dp1_tp1_pp8_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_tp1_pp8_acc4_mbs64_seq4096_zero0_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_3.57G_dp1_tp1_pp8_acc7_mbs2_seq4096_zero0_tpmodeRED_vocab131k_cache.sh b/scripts/run_3.57G_dp1_tp1_pp8_acc7_mbs2_seq4096_zero0_tpmodeRED_vocab131k_cache.sh new file mode 100644 index 0000000000000000000000000000000000000000..c90d2629e0cb066d500f957c6418fb181c49d2bd --- /dev/null +++ b/scripts/run_3.57G_dp1_tp1_pp8_acc7_mbs2_seq4096_zero0_tpmodeRED_vocab131k_cache.sh @@ -0,0 +1,161 @@ +#!/bin/bash +#SBATCH --job-name=bench_3.57G_dp1_tp1_pp8_acc7_mbs2_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=1 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e +echo "Running script: $0" + + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_3.57G_dp1_tp1_pp8_acc7_mbs2_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_pp8_acc7_mbs2_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_pp8_acc7_mbs2_seq4096_zero0_tpmodeRED_vocab131k_cache.yaml +fi diff --git a/scripts/run_3.57G_dp1_tp4_pp8_acc8_mbs32_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_3.57G_dp1_tp4_pp8_acc8_mbs32_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..67f3fb281ee2bfda90d6adc7a1085902e11553a4 --- /dev/null +++ b/scripts/run_3.57G_dp1_tp4_pp8_acc8_mbs32_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,73 @@ +#!/bin/bash + +#SBATCH --job-name=bench_3.57G_dp1_tp4_pp8_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=4 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +# Disable wandb +export WANDB_MODE=disabled + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + + +# Print GPU topology information +echo "=== GPU Topology ===" +nvidia-smi topo -m +echo "==================" + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_3.57G_dp1_tp4_pp8_acc8_mbs32_seq4096_zero0_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_3.57G_dp1_tp8_pp1_acc1_mbs23_seq4096_zero0_tpmodeRED_vocab131k_cache.sh b/scripts/run_3.57G_dp1_tp8_pp1_acc1_mbs23_seq4096_zero0_tpmodeRED_vocab131k_cache.sh new file mode 100644 index 0000000000000000000000000000000000000000..dc9de1bc7af43988c662873d487aa271d04450cb --- /dev/null +++ b/scripts/run_3.57G_dp1_tp8_pp1_acc1_mbs23_seq4096_zero0_tpmodeRED_vocab131k_cache.sh @@ -0,0 +1,161 @@ +#!/bin/bash +#SBATCH --job-name=bench_3.57G_dp1_tp8_pp1_acc1_mbs23_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=1 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e +echo "Running script: $0" + + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_3.57G_dp1_tp8_pp1_acc1_mbs23_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_tp8_pp1_acc1_mbs23_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_tp8_pp1_acc1_mbs23_seq4096_zero0_tpmodeRED_vocab131k_cache.yaml +fi diff --git a/scripts/run_3.57G_dp2_tp8_pp1_acc1_mbs128_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_3.57G_dp2_tp8_pp1_acc1_mbs128_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..49bf6e79852a802cf77fd96712e5f0aff4f8b436 --- /dev/null +++ b/scripts/run_3.57G_dp2_tp8_pp1_acc1_mbs128_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,124 @@ +#!/bin/bash +#SBATCH --job-name=bench_3.57G_dp2_tp8_pp1_acc1_mbs128_seq4096_zero0_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=2 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_3.57G_dp2_tp8_pp1_acc1_mbs128_seq4096_zero0_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_3.57G_dp2_tp8_pp1_acc1_mbs128_seq4096_zero0_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_3.57G_dp2_tp8_pp1_acc1_mbs128_seq4096_zero0_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_3.57G_dp32_tp2_pp4_acc8_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_3.57G_dp32_tp2_pp4_acc8_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..b3c33d1d7273ffe461fef2e312c38211cf4dac77 --- /dev/null +++ b/scripts/run_3.57G_dp32_tp2_pp4_acc8_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,161 @@ +#!/bin/bash +#SBATCH --job-name=bench_3.57G_dp32_tp2_pp4_acc8_mbs1_seq4096_zero0_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:40:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=normal + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=32 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e +echo "Running script: $0" + + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_3.57G_dp32_tp2_pp4_acc8_mbs1_seq4096_zero0_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_3.57G_dp32_tp2_pp4_acc8_mbs1_seq4096_zero0_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_3.57G_dp32_tp2_pp4_acc8_mbs1_seq4096_zero0_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_3.8G_dp1_tp1_pp16_acc15_mbs1_seq4096_zero0_tpmodeRED_vocab131k_cache.sh b/scripts/run_3.8G_dp1_tp1_pp16_acc15_mbs1_seq4096_zero0_tpmodeRED_vocab131k_cache.sh new file mode 100644 index 0000000000000000000000000000000000000000..1663f1404eae2d45bac7be9142f22f2d09708b36 --- /dev/null +++ b/scripts/run_3.8G_dp1_tp1_pp16_acc15_mbs1_seq4096_zero0_tpmodeRED_vocab131k_cache.sh @@ -0,0 +1,161 @@ +#!/bin/bash +#SBATCH --job-name=bench_3.8G_dp1_tp1_pp16_acc15_mbs1_seq4096_zero0_tpmodeRED_vocab131k_cache # Job name +#SBATCH --time=00:40:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=2 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e +echo "Running script: $0" + + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_3.8G_dp1_tp1_pp16_acc15_mbs1_seq4096_zero0_tpmodeRED_vocab131k_cache" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_3.8G_dp1_tp1_pp16_acc15_mbs1_seq4096_zero0_tpmodeRED_vocab131k_cache.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_3.8G_dp1_tp1_pp16_acc15_mbs1_seq4096_zero0_tpmodeRED_vocab131k_cache.yaml +fi diff --git a/scripts/run_469G_dp1_tp16_pp4_acc16_mbs16_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_469G_dp1_tp16_pp4_acc16_mbs16_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..9d1e85f4c5bc2415117b5a450754363b2b9f0593 --- /dev/null +++ b/scripts/run_469G_dp1_tp16_pp4_acc16_mbs16_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,73 @@ +#!/bin/bash + +#SBATCH --job-name=bench_469G_dp1_tp16_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=8 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +# Disable wandb +export WANDB_MODE=disabled + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + + +# Print GPU topology information +echo "=== GPU Topology ===" +nvidia-smi topo -m +echo "==================" + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_469G_dp1_tp16_pp4_acc16_mbs16_seq4096_zero0_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_469G_dp2_tp1_pp16_acc128_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_469G_dp2_tp1_pp16_acc128_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..e1ea4bd5fc2482db19fcc6cb8bab619ace4035f6 --- /dev/null +++ b/scripts/run_469G_dp2_tp1_pp16_acc128_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,73 @@ +#!/bin/bash + +#SBATCH --job-name=bench_469G_dp2_tp1_pp16_acc128_mbs1_seq4096_zero0_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high +#SBATCH --exclude=ip-26-0-160-192,ip-26-0-171-102 + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=4 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +# Disable wandb +export WANDB_MODE=disabled + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + + +# Print GPU topology information +echo "=== GPU Topology ===" +nvidia-smi topo -m +echo "==================" + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_469G_dp2_tp1_pp16_acc128_mbs1_seq4096_zero0_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_469G_dp2_tp1_pp16_acc64_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_469G_dp2_tp1_pp16_acc64_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..41758c25be59eb3c68e0fb49c36c43883b7990bb --- /dev/null +++ b/scripts/run_469G_dp2_tp1_pp16_acc64_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_469G_dp2_tp1_pp16_acc64_mbs2_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=4 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=INFO # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_469G_dp2_tp1_pp16_acc64_mbs2_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_469G_dp2_tp1_pp16_acc64_mbs2_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_469G_dp2_tp1_pp16_acc64_mbs2_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_469G_dp2_tp32_pp2_acc1_mbs128_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_469G_dp2_tp32_pp2_acc1_mbs128_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..07e1e7b22b71529810310f93c8640a3340979790 --- /dev/null +++ b/scripts/run_469G_dp2_tp32_pp2_acc1_mbs128_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_469G_dp2_tp32_pp2_acc1_mbs128_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=16 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=INFO # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_469G_dp2_tp32_pp2_acc1_mbs128_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_469G_dp2_tp32_pp2_acc1_mbs128_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_469G_dp2_tp32_pp2_acc1_mbs128_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_469G_dp32_tp16_pp1_acc2_mbs4_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_469G_dp32_tp16_pp1_acc2_mbs4_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..7b69386b5efc6126865ae9dab09de55fffb59c40 --- /dev/null +++ b/scripts/run_469G_dp32_tp16_pp1_acc2_mbs4_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_469G_dp32_tp16_pp1_acc2_mbs4_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_469G_dp32_tp16_pp1_acc2_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_469G_dp32_tp16_pp1_acc2_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_469G_dp32_tp16_pp1_acc2_mbs4_seq4096_zero0_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_469G_dp4_tp32_pp4_acc32_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_469G_dp4_tp32_pp4_acc32_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..3a0ce1d8fa2ff4ce97ea2ec6f9506f69e6ab3819 --- /dev/null +++ b/scripts/run_469G_dp4_tp32_pp4_acc32_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_469G_dp4_tp32_pp4_acc32_mbs2_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:40:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=64 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_469G_dp4_tp32_pp4_acc32_mbs2_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_469G_dp4_tp32_pp4_acc32_mbs2_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_469G_dp4_tp32_pp4_acc32_mbs2_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_469G_dp8_tp1_pp2_acc1_mbs32_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_469G_dp8_tp1_pp2_acc1_mbs32_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..9c57a43c9718aa3c9e6a9df3f956637345d35167 --- /dev/null +++ b/scripts/run_469G_dp8_tp1_pp2_acc1_mbs32_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,73 @@ +#!/bin/bash + +#SBATCH --job-name=bench_469G_dp8_tp1_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=2 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +# Disable wandb +export WANDB_MODE=disabled + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + + +# Print GPU topology information +echo "=== GPU Topology ===" +nvidia-smi topo -m +echo "==================" + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_469G_dp8_tp1_pp2_acc1_mbs32_seq4096_zero0_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_469G_dp8_tp4_pp16_acc16_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_469G_dp8_tp4_pp16_acc16_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..f2df3b3a263179387605acffc0907a923f1dd60b --- /dev/null +++ b/scripts/run_469G_dp8_tp4_pp16_acc16_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_469G_dp8_tp4_pp16_acc16_mbs2_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:40:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=64 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_469G_dp8_tp4_pp16_acc16_mbs2_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_469G_dp8_tp4_pp16_acc16_mbs2_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_469G_dp8_tp4_pp16_acc16_mbs2_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_469G_dp8_tp4_pp2_acc32_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_469G_dp8_tp4_pp2_acc32_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..e463ba0fafcd0d48a890a78ddc4fda8c1ef95000 --- /dev/null +++ b/scripts/run_469G_dp8_tp4_pp2_acc32_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_469G_dp8_tp4_pp2_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=8 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_469G_dp8_tp4_pp2_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_469G_dp8_tp4_pp2_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_469G_dp8_tp4_pp2_acc32_mbs1_seq4096_zero0_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_469G_dp8_tp8_pp2_acc4_mbs8_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_469G_dp8_tp8_pp2_acc4_mbs8_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..4d7ca8cdc769cbd161d4ac4f08f7119bf031f145 --- /dev/null +++ b/scripts/run_469G_dp8_tp8_pp2_acc4_mbs8_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,73 @@ +#!/bin/bash + +#SBATCH --job-name=bench_469G_dp8_tp8_pp2_acc4_mbs8_seq4096_zero0_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high +#SBATCH --exclude=ip-26-0-160-192,ip-26-0-171-102 + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=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_469G_dp8_tp8_pp2_acc4_mbs8_seq4096_zero0_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_8.86G_dp16_tp4_pp2_acc8_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_8.86G_dp16_tp4_pp2_acc8_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..0bf36a5c4d68f57ca0e36c979909449e8ae0bb3f --- /dev/null +++ b/scripts/run_8.86G_dp16_tp4_pp2_acc8_mbs2_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_8.86G_dp16_tp4_pp2_acc8_mbs2_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=16 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=INFO # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_8.86G_dp16_tp4_pp2_acc8_mbs2_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_8.86G_dp16_tp4_pp2_acc8_mbs2_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_8.86G_dp16_tp4_pp2_acc8_mbs2_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_8.86G_dp2_tp2_pp4_acc8_mbs16_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_8.86G_dp2_tp2_pp4_acc8_mbs16_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..9cf6a56bb30a4008cc3d78185de9d30004750f11 --- /dev/null +++ b/scripts/run_8.86G_dp2_tp2_pp4_acc8_mbs16_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_8.86G_dp2_tp2_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=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_tp2_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_8.86G_dp2_tp2_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_8.86G_dp2_tp2_pp4_acc8_mbs16_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_8.86G_dp2_tp32_pp8_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_8.86G_dp2_tp32_pp8_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..913cfaa638ca0ac018ebfc0466efee84eb23b6a8 --- /dev/null +++ b/scripts/run_8.86G_dp2_tp32_pp8_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_8.86G_dp2_tp32_pp8_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:40:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=64 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_8.86G_dp2_tp32_pp8_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_8.86G_dp2_tp32_pp8_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_8.86G_dp2_tp32_pp8_acc16_mbs8_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_8.86G_dp2_tp4_pp1_acc128_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_8.86G_dp2_tp4_pp1_acc128_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..cc275a2ee502b9fd99dce179fb08c2fe2c1eaed7 --- /dev/null +++ b/scripts/run_8.86G_dp2_tp4_pp1_acc128_mbs1_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,73 @@ +#!/bin/bash + +#SBATCH --job-name=bench_8.86G_dp2_tp4_pp1_acc128_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=1 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +# Disable wandb +export WANDB_MODE=disabled + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + + +# Print GPU topology information +echo "=== GPU Topology ===" +nvidia-smi topo -m +echo "==================" + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_8.86G_dp2_tp4_pp1_acc128_mbs1_seq4096_zero1_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_8.86G_dp32_tp8_pp1_acc2_mbs64_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_8.86G_dp32_tp8_pp1_acc2_mbs64_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..3381a576be9bf9c90879a8e90d5b5814be5ba397 --- /dev/null +++ b/scripts/run_8.86G_dp32_tp8_pp1_acc2_mbs64_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,161 @@ +#!/bin/bash +#SBATCH --job-name=bench_8.86G_dp32_tp8_pp1_acc2_mbs64_seq4096_zero0_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:40:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=normal + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=32 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e +echo "Running script: $0" + + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_8.86G_dp32_tp8_pp1_acc2_mbs64_seq4096_zero0_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_8.86G_dp32_tp8_pp1_acc2_mbs64_seq4096_zero0_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_8.86G_dp32_tp8_pp1_acc2_mbs64_seq4096_zero0_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_8.86G_dp8_tp1_pp1_acc32_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh b/scripts/run_8.86G_dp8_tp1_pp1_acc32_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..54f9ae608c6767821d7229c6f0ad9f38816e1dc8 --- /dev/null +++ b/scripts/run_8.86G_dp8_tp1_pp1_acc32_mbs1_seq4096_zero0_tpmodeRED_vocab131k.sh @@ -0,0 +1,73 @@ +#!/bin/bash + +#SBATCH --job-name=bench_8.86G_dp8_tp1_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=1 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +set -x -e + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +# Disable wandb +export WANDB_MODE=disabled + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + + +# Print GPU topology information +echo "=== GPU Topology ===" +nvidia-smi topo -m +echo "==================" + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun +srun --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_8.86G_dp8_tp1_pp1_acc32_mbs1_seq4096_zero0_tpmodeRED_vocab131k.yaml diff --git a/scripts/run_8.86G_dp8_tp8_pp8_acc8_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_8.86G_dp8_tp8_pp8_acc8_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..7c3908e96d309ce00e69dcb6f436f9f719e56d77 --- /dev/null +++ b/scripts/run_8.86G_dp8_tp8_pp8_acc8_mbs4_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_8.86G_dp8_tp8_pp8_acc8_mbs4_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=00:40:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=64 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=WARN # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_8.86G_dp8_tp8_pp8_acc8_mbs4_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_8.86G_dp8_tp8_pp8_acc8_mbs4_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_8.86G_dp8_tp8_pp8_acc8_mbs4_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_80G_dp8_tp1_pp2_acc1_mbs32_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_80G_dp8_tp1_pp2_acc1_mbs32_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..61daf3c3882665a542170a88e2d7138d6cd288c1 --- /dev/null +++ b/scripts/run_80G_dp8_tp1_pp2_acc1_mbs32_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_80G_dp8_tp1_pp2_acc1_mbs32_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=2 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=INFO # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_80G_dp8_tp1_pp2_acc1_mbs32_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp8_tp1_pp2_acc1_mbs32_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp8_tp1_pp2_acc1_mbs32_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi diff --git a/scripts/run_80G_dp8_tp8_pp4_acc4_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh b/scripts/run_80G_dp8_tp8_pp4_acc4_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh new file mode 100644 index 0000000000000000000000000000000000000000..0d13fa474d49c6a667aa979859c1348a45549eb2 --- /dev/null +++ b/scripts/run_80G_dp8_tp8_pp4_acc4_mbs8_seq4096_zero1_tpmodeRED_vocab131k.sh @@ -0,0 +1,159 @@ +#!/bin/bash +#SBATCH --job-name=bench_80G_dp8_tp8_pp4_acc4_mbs8_seq4096_zero1_tpmodeRED_vocab131k # Job name +#SBATCH --time=01:10:00 +#SBATCH --partition=hopper-prod +#SBATCH --qos=high + +#SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out + +#SBATCH --nodes=32 # Number of nodes (modify as needed) +#SBATCH --ntasks-per-node=1 # Number of tasks per node +#SBATCH --cpus-per-task=60 # CPU cores per task +#SBATCH --gres=gpu:8 # Number of GPUs per node +#SBATCH --exclusive # Exclusive use of nodes +#SBATCH --wait-all-nodes=1 # fail if any node is not ready + +# run using +# sbatch --nodes=1 run_multinode.sh +# or +# SALLOC_JOBID=13482276 NNODES=1 bash run_multinode.sh + +set -x -e + +# If not running under SLURM, set default SLURM environment variables +if [ -z "${SLURM_JOB_ID}" ]; then + if [ -z "${SALLOC_JOBID}" ]; then + echo "Error: SALLOC_JOBID environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + if [ -z "${NNODES}" ]; then + echo "Error: NNODES environment variable is required but not set. Please run this script within an salloc session." + exit 1 + fi + export SALLOC_MODE=1 + export SLURM_JOB_ID=$SALLOC_JOBID + export SLURM_NNODES=$NNODES + export SLURM_JOB_NODELIST=$(squeue -j $SALLOC_JOBID -h -o "%N") +fi + +# Load any necessary modules for your system +source /etc/profile.d/modules.sh # for some reason module isn't loaded +module load cuda/12.1 +# Unset FI_PROVIDER to avoid potential libfabric provider issues +# unset FI_PROVIDER + + +# Activate your conda environment if needed +source /fsx/nouamane/miniconda/bin/activate +conda activate 2-1-cu121 +export PATH=/fsx/nouamane/miniconda/envs/2-1-cu121/bin:$PATH + +# Get the node names from SLURM +if [ -z "${SALLOC_MODE}" ]; then # sbatch mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST` + +else # srun mode + export NODELIST=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n$SLURM_NNODES` +fi +export MASTER_NODE=`scontrol show hostnames $SLURM_JOB_NODELIST | head -n1` +export MASTER_PORT=12356 + +# Calculate total number of processes +export NNODES=$SLURM_NNODES +export GPUS_PER_NODE=8 +export WORLD_SIZE=$(($NNODES * $GPUS_PER_NODE)) + +# Set some environment variables for better distributed training +export CUDA_DEVICE_MAX_CONNECTIONS=1 +export NCCL_DEBUG=INFO # INFO, WARN +# export NCCL_DEBUG_SUBSYS=ALL +# export CUDA_LAUNCH_BLOCKING=1 + +# Nanotron specific +export NANOTRON_BENCHMARK=1 +export WANDB_MODE=disabled + +# export TORCH_NCCL_USE_COMM_NONBLOCKING=1 + +# Trying to avoid hangs +export TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + +# debug +export TORCH_DISTRIBUTED_DEBUG=DETAIL + +# export NCCL_P2P_LEVEL=NVL +# export CUDA_LAUNCH_BLOCKING=1 +# export NCCL_IB_CUDA_SUPPORT=0 # Disable RDMA +# export NCCL_NET_GDR_LEVEL=LOC +# Test Script - save as test_comm.sh + +# Test 1 - Force TCP +# echo "Running with TCP only..." +# export NCCL_P2P_LEVEL=LOC + +# # Match bandwidth patterns +# export NCCL_MAX_NCHANNELS=2 +# export NCCL_MIN_NCHANNELS=2 + + +# export NCCL_NET_GDR_LEVEL=LOC # Disable RDMA +# export NCCL_SHM_DISABLE=0 # disables the Shared Memory (SHM) transport +# export NCCL_IB_DISABLE=0 # disables the InfiniBand (IB) transport +# export NCCL_IB_TIMEOUT=60 # 20 = ~4 seconds , 21 = ~8 seconds , 22 = ~16 seconds +# export NCCL_IB_RETRY_CNT=7 # Increase retry count as well + +# Force SHM +# export NCCL_NET_PLUGIN=none # fixes hang but doesnt work multinode +# export NCCL_SOCKET_NTHREADS=1 +# export FI_PROVIDER="tcp" + +# Print GPU topology information +if [ -z "${SALLOC_MODE}" ]; then + echo "=== GPU Topology ===" + nvidia-smi topo -m + echo "==================" + export SRUN_ALLOC_ARGS="" +else + export JOBNAME="bench_80G_dp8_tp8_pp4_acc4_mbs8_seq4096_zero1_tpmodeRED_vocab131k" + export OUTPUT_FILE="/fsx/nouamane/projects/nanotron/logs/$SLURM_JOB_ID-$(date +%Y-%m-%d-%H-%M-%S)-$JOBNAME.out" + export SRUN_ALLOC_ARGS="--jobid=$SLURM_JOB_ID --nodes=$NNODES --gres=gpu:$GPUS_PER_NODE --time=01:02:00 --job-name=$JOBNAME" +fi + + +# Print some debugging information +echo "Master node: $MASTER_NODE" +echo "All nodes: $NODELIST" +echo "World size: $WORLD_SIZE" + +# Launch the training script using srun in background +if [ -n "${SALLOC_MODE}" ]; then # srun mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp8_tp8_pp4_acc4_mbs8_seq4096_zero1_tpmodeRED_vocab131k.yaml > $OUTPUT_FILE 2>&1 & + # Store the process ID + SRUN_PID=$! + echo "Job started in background with PID: $SRUN_PID" | tee -a $OUTPUT_FILE + + # Optionally, you can add: + echo "To check job status: ps -p $SRUN_PID" | tee -a $OUTPUT_FILE + echo "To kill the job: kill $SRUN_PID" | tee -a $OUTPUT_FILE + +else # sbatch mode + srun $SRUN_ALLOC_ARGS --wait=0 --kill-on-bad-exit=1 torchrun \ + --nnodes=$NNODES \ + --nproc_per_node=$GPUS_PER_NODE \ + --rdzv_id=$SLURM_JOB_ID \ + --rdzv_backend=c10d \ + --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \ + --max_restarts 0 \ + --rdzv_conf timeout=60 \ + /fsx/nouamane/projects/nanotron/run_train.py \ + --config-file benchmark/configs/config_80G_dp8_tp8_pp4_acc4_mbs8_seq4096_zero1_tpmodeRED_vocab131k.yaml +fi