add results bench picotron

result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp2_pp8_mbs1_ga64_sl4096/config.json

@@ -0,0 +1,51 @@
+{
+    "distributed": {
+        "tp_size": 2,
+        "cp_size": 1,
+        "pp_size": 8,
+        "dp_size": 16,
+        "pp_engine": "1f1b",
+        "backend": "nccl",
+        "use_cpu": false
+    },
+    "model": {
+        "name": "meta-llama/Llama-2-70b-hf",
+        "num_hidden_layers": 80,
+        "num_attention_heads": 64,
+        "num_key_value_heads": 8,
+        "dtype": "bfloat16",
+        "use_flash_attention": true,
+        "use_fused_adam": false
+    },
+    "training": {
+        "seed": 42,
+        "learning_rate": 0.0003,
+        "total_train_steps": 200,
+        "seq_length": 4096,
+        "micro_batch_size": 1,
+        "gradient_accumulation_steps": 64,
+        "num_samples": 400000,
+        "max_tokens": null
+    },
+    "dataset": {
+        "name": "roneneldan/TinyStories",
+        "num_workers": 0,
+        "num_proc": 4
+    },
+    "checkpoint": {
+        "save_dir": "result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp2_pp8_mbs1_ga64_sl4096",
+        "save_frequency": 300,
+        "load_path": "",
+        "hf_hub_safetensors_path": "/fsx/ferdinandmom/hf_model_ckpt/Llama-2-70b-chat-hf"
+    },
+    "logging": {
+        "use_wandb": true,
+        "project_name": "picotron",
+        "run_name": "70b_32_node_with_dp_tp_pp/dp16_tp2_pp8_mbs1_ga64_sl4096"
+    },
+    "environment": {
+        "OMP_NUM_THREADS": "1",
+        "TOKENIZERS_PARALLELISM": "false",
+        "FLASH_ATTEN": "1"
+    }
+}

result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp2_pp8_mbs1_ga64_sl4096/job.slurm

@@ -0,0 +1,94 @@
+#!/bin/bash
+
+#SBATCH --job-name=job-picotron
+#SBATCH --time=00:30:00
+#SBATCH --partition=hopper-prod
+#SBATCH --nodes=32
+#SBATCH --gres=gpu:8
+#SBATCH --qos=normal
+#SBATCH --ntasks-per-node=1
+#SBATCH --cpus-per-task=96
+#SBATCH --exclusive
+#SBATCH --output=/fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp2_pp8_mbs1_ga64_sl4096/log_%j.out
+#SBATCH --error=/fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp2_pp8_mbs1_ga64_sl4096/log_%j.out
+
+# Function to update status based on squeue output
+update_status() {
+    job_id=$1
+    status_file=$2
+    # For unknown reasons, it doenst update status for pending. It only works for running 
+    while true; do
+        job_status=$(squeue --job $job_id --noheader --format=%T)
+        echo "Job status: $job_status"
+        if [ -z "$job_status" ]; then
+            # Job has finished or is not found
+            break
+        elif [ "$job_status" = "RUNNING" ]; then
+            printf "running" > $status_file
+            break
+        fi
+        sleep 10
+    done
+}
+
+# Misc initializations.
+echo "========================"
+echo "START TIME: $(date)"
+source /etc/profile.d/modules.sh
+source /fsx/ferdinandmom/miniforge3/etc/profile.d/conda.sh
+conda activate /fsx/ferdinandmom/miniforge3/envs/env-picotron
+echo python3 version = $(python3 --version)
+echo "========================"
+
+# Slurm stuff
+export HOSTNAMES=$(scontrol show hostnames "$SLURM_JOB_NODELIST")
+export MASTER_ADDR=$(scontrol show hostnames "$SLURM_JOB_NODELIST" | head -n 1)
+export MASTER_PORT=$((1024 + RANDOM % 64511))
+
+export TMPDIR=/scratch
+export TORCH_HOME=/fsx/$USER/.cache/torch
+export HF_HOME=/fsx/$USER/.cache/huggingface
+export WANDB_DIR=/fsx/$USER/.cache/wandb
+export CUBLAS_WORKSPACE_CONFIG=":4096:8"
+export CUDA_DEVICE_MAX_CONNECTIONS="1"
+export FI_PROVIDER="efa"
+
+module load cuda/12.1
+
+GIT_REPO="/fsx/ferdinandmom/ferdinand-hf/picotron/"
+CMD="$GIT_REPO/train.py --config /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp2_pp8_mbs1_ga64_sl4096/config.json"
+
+git checkout loading_big_model
+# huggingface-cli login --token $HUGGINGFACE_TOKEN
+
+LAUNCHER="torchrun --nproc_per_node=8 --nnode=32 --node_rank=$SLURM_NODEID --rdzv_endpoint ${MASTER_ADDR}:${MASTER_PORT} --rdzv_backend c10d --max_restarts 0 --tee 3"
+
+# Checkout the bench_cluster branch
+cd $GIT_REPO
+# Get the current job ID
+job_id=${SLURM_JOB_ID}
+
+# Update status to "pending" or "running" in the background
+update_status $job_id /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp2_pp8_mbs1_ga64_sl4096/status.txt &
+
+# Run the main command
+echo "Running command: $CMD"
+srun -u $LAUNCHER $CMD
+exit_status=$?
+
+job_id=$SLURM_JOB_ID
+
+# Update status based on the exit status of `srun`
+if [ $exit_status -eq 0 ]; then
+    printf "completed" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp2_pp8_mbs1_ga64_sl4096/status.txt
+else
+    if grep -q "OutOfMemoryError" /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp2_pp8_mbs1_ga64_sl4096/log_${job_id}.out; then
+        printf "oom" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp2_pp8_mbs1_ga64_sl4096/status.txt
+    elif grep -q " CUDA error: an illegal memory access" /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp2_pp8_mbs1_ga64_sl4096/log_${job_id}.out; then
+        printf "oom" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp2_pp8_mbs1_ga64_sl4096/status.txt
+    elif grep -q "Timeout" /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp2_pp8_mbs1_ga64_sl4096/log_${job_id}.out; then
+        printf "timeout" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp2_pp8_mbs1_ga64_sl4096/status.txt
+    else
+        printf "fail" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp2_pp8_mbs1_ga64_sl4096/status.txt
+    fi 
+fi

result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp2_pp8_mbs1_ga64_sl4096/metrics.csv

@@ -0,0 +1,2 @@
+run_name,status,dp,tp,pp,micro_batch_size,grad_acc,seq_len,avg_tokens_s_gpu,avg_mfu
+dp16_tp2_pp8_mbs1_ga64_sl4096,oom,16,2,8,1,64,4096,,

result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp2_pp8_mbs1_ga64_sl4096/status.txt

@@ -0,0 +1 @@
+oom

result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp4_pp4_mbs1_ga64_sl4096/config.json

@@ -0,0 +1,51 @@
+{
+    "distributed": {
+        "tp_size": 4,
+        "cp_size": 1,
+        "pp_size": 4,
+        "dp_size": 16,
+        "pp_engine": "1f1b",
+        "backend": "nccl",
+        "use_cpu": false
+    },
+    "model": {
+        "name": "meta-llama/Llama-2-70b-hf",
+        "num_hidden_layers": 80,
+        "num_attention_heads": 64,
+        "num_key_value_heads": 8,
+        "dtype": "bfloat16",
+        "use_flash_attention": true,
+        "use_fused_adam": false
+    },
+    "training": {
+        "seed": 42,
+        "learning_rate": 0.0003,
+        "total_train_steps": 200,
+        "seq_length": 4096,
+        "micro_batch_size": 1,
+        "gradient_accumulation_steps": 64,
+        "num_samples": 400000,
+        "max_tokens": null
+    },
+    "dataset": {
+        "name": "roneneldan/TinyStories",
+        "num_workers": 0,
+        "num_proc": 4
+    },
+    "checkpoint": {
+        "save_dir": "result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp4_pp4_mbs1_ga64_sl4096",
+        "save_frequency": 300,
+        "load_path": "",
+        "hf_hub_safetensors_path": "/fsx/ferdinandmom/hf_model_ckpt/Llama-2-70b-chat-hf"
+    },
+    "logging": {
+        "use_wandb": true,
+        "project_name": "picotron",
+        "run_name": "70b_32_node_with_dp_tp_pp/dp16_tp4_pp4_mbs1_ga64_sl4096"
+    },
+    "environment": {
+        "OMP_NUM_THREADS": "1",
+        "TOKENIZERS_PARALLELISM": "false",
+        "FLASH_ATTEN": "1"
+    }
+}

result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp4_pp4_mbs1_ga64_sl4096/job.slurm

@@ -0,0 +1,94 @@
+#!/bin/bash
+
+#SBATCH --job-name=job-picotron
+#SBATCH --time=00:30:00
+#SBATCH --partition=hopper-prod
+#SBATCH --nodes=32
+#SBATCH --gres=gpu:8
+#SBATCH --qos=normal
+#SBATCH --ntasks-per-node=1
+#SBATCH --cpus-per-task=96
+#SBATCH --exclusive
+#SBATCH --output=/fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp4_pp4_mbs1_ga64_sl4096/log_%j.out
+#SBATCH --error=/fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp4_pp4_mbs1_ga64_sl4096/log_%j.out
+
+# Function to update status based on squeue output
+update_status() {
+    job_id=$1
+    status_file=$2
+    # For unknown reasons, it doenst update status for pending. It only works for running 
+    while true; do
+        job_status=$(squeue --job $job_id --noheader --format=%T)
+        echo "Job status: $job_status"
+        if [ -z "$job_status" ]; then
+            # Job has finished or is not found
+            break
+        elif [ "$job_status" = "RUNNING" ]; then
+            printf "running" > $status_file
+            break
+        fi
+        sleep 10
+    done
+}
+
+# Misc initializations.
+echo "========================"
+echo "START TIME: $(date)"
+source /etc/profile.d/modules.sh
+source /fsx/ferdinandmom/miniforge3/etc/profile.d/conda.sh
+conda activate /fsx/ferdinandmom/miniforge3/envs/env-picotron
+echo python3 version = $(python3 --version)
+echo "========================"
+
+# Slurm stuff
+export HOSTNAMES=$(scontrol show hostnames "$SLURM_JOB_NODELIST")
+export MASTER_ADDR=$(scontrol show hostnames "$SLURM_JOB_NODELIST" | head -n 1)
+export MASTER_PORT=$((1024 + RANDOM % 64511))
+
+export TMPDIR=/scratch
+export TORCH_HOME=/fsx/$USER/.cache/torch
+export HF_HOME=/fsx/$USER/.cache/huggingface
+export WANDB_DIR=/fsx/$USER/.cache/wandb
+export CUBLAS_WORKSPACE_CONFIG=":4096:8"
+export CUDA_DEVICE_MAX_CONNECTIONS="1"
+export FI_PROVIDER="efa"
+
+module load cuda/12.1
+
+GIT_REPO="/fsx/ferdinandmom/ferdinand-hf/picotron/"
+CMD="$GIT_REPO/train.py --config /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp4_pp4_mbs1_ga64_sl4096/config.json"
+
+git checkout loading_big_model
+# huggingface-cli login --token $HUGGINGFACE_TOKEN
+
+LAUNCHER="torchrun --nproc_per_node=8 --nnode=32 --node_rank=$SLURM_NODEID --rdzv_endpoint ${MASTER_ADDR}:${MASTER_PORT} --rdzv_backend c10d --max_restarts 0 --tee 3"
+
+# Checkout the bench_cluster branch
+cd $GIT_REPO
+# Get the current job ID
+job_id=${SLURM_JOB_ID}
+
+# Update status to "pending" or "running" in the background
+update_status $job_id /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp4_pp4_mbs1_ga64_sl4096/status.txt &
+
+# Run the main command
+echo "Running command: $CMD"
+srun -u $LAUNCHER $CMD
+exit_status=$?
+
+job_id=$SLURM_JOB_ID
+
+# Update status based on the exit status of `srun`
+if [ $exit_status -eq 0 ]; then
+    printf "completed" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp4_pp4_mbs1_ga64_sl4096/status.txt
+else
+    if grep -q "OutOfMemoryError" /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp4_pp4_mbs1_ga64_sl4096/log_${job_id}.out; then
+        printf "oom" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp4_pp4_mbs1_ga64_sl4096/status.txt
+    elif grep -q " CUDA error: an illegal memory access" /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp4_pp4_mbs1_ga64_sl4096/log_${job_id}.out; then
+        printf "oom" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp4_pp4_mbs1_ga64_sl4096/status.txt
+    elif grep -q "Timeout" /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp4_pp4_mbs1_ga64_sl4096/log_${job_id}.out; then
+        printf "timeout" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp4_pp4_mbs1_ga64_sl4096/status.txt
+    else
+        printf "fail" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp4_pp4_mbs1_ga64_sl4096/status.txt
+    fi 
+fi

result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp4_pp4_mbs1_ga64_sl4096/metrics.csv

@@ -0,0 +1,2 @@
+run_name,status,dp,tp,pp,micro_batch_size,grad_acc,seq_len,avg_tokens_s_gpu,avg_mfu
+dp16_tp4_pp4_mbs1_ga64_sl4096,oom,16,4,4,1,64,4096,,

result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp4_pp4_mbs1_ga64_sl4096/status.txt

@@ -0,0 +1 @@
+oom

result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp8_pp2_mbs1_ga64_sl4096/config.json

@@ -0,0 +1,51 @@
+{
+    "distributed": {
+        "tp_size": 8,
+        "cp_size": 1,
+        "pp_size": 2,
+        "dp_size": 16,
+        "pp_engine": "1f1b",
+        "backend": "nccl",
+        "use_cpu": false
+    },
+    "model": {
+        "name": "meta-llama/Llama-2-70b-hf",
+        "num_hidden_layers": 80,
+        "num_attention_heads": 64,
+        "num_key_value_heads": 8,
+        "dtype": "bfloat16",
+        "use_flash_attention": true,
+        "use_fused_adam": false
+    },
+    "training": {
+        "seed": 42,
+        "learning_rate": 0.0003,
+        "total_train_steps": 200,
+        "seq_length": 4096,
+        "micro_batch_size": 1,
+        "gradient_accumulation_steps": 64,
+        "num_samples": 400000,
+        "max_tokens": null
+    },
+    "dataset": {
+        "name": "roneneldan/TinyStories",
+        "num_workers": 0,
+        "num_proc": 4
+    },
+    "checkpoint": {
+        "save_dir": "result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp8_pp2_mbs1_ga64_sl4096",
+        "save_frequency": 300,
+        "load_path": "",
+        "hf_hub_safetensors_path": "/fsx/ferdinandmom/hf_model_ckpt/Llama-2-70b-chat-hf"
+    },
+    "logging": {
+        "use_wandb": true,
+        "project_name": "picotron",
+        "run_name": "70b_32_node_with_dp_tp_pp/dp16_tp8_pp2_mbs1_ga64_sl4096"
+    },
+    "environment": {
+        "OMP_NUM_THREADS": "1",
+        "TOKENIZERS_PARALLELISM": "false",
+        "FLASH_ATTEN": "1"
+    }
+}

result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp8_pp2_mbs1_ga64_sl4096/job.slurm

@@ -0,0 +1,94 @@
+#!/bin/bash
+
+#SBATCH --job-name=job-picotron
+#SBATCH --time=00:30:00
+#SBATCH --partition=hopper-prod
+#SBATCH --nodes=32
+#SBATCH --gres=gpu:8
+#SBATCH --qos=normal
+#SBATCH --ntasks-per-node=1
+#SBATCH --cpus-per-task=96
+#SBATCH --exclusive
+#SBATCH --output=/fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp8_pp2_mbs1_ga64_sl4096/log_%j.out
+#SBATCH --error=/fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp8_pp2_mbs1_ga64_sl4096/log_%j.out
+
+# Function to update status based on squeue output
+update_status() {
+    job_id=$1
+    status_file=$2
+    # For unknown reasons, it doenst update status for pending. It only works for running 
+    while true; do
+        job_status=$(squeue --job $job_id --noheader --format=%T)
+        echo "Job status: $job_status"
+        if [ -z "$job_status" ]; then
+            # Job has finished or is not found
+            break
+        elif [ "$job_status" = "RUNNING" ]; then
+            printf "running" > $status_file
+            break
+        fi
+        sleep 10
+    done
+}
+
+# Misc initializations.
+echo "========================"
+echo "START TIME: $(date)"
+source /etc/profile.d/modules.sh
+source /fsx/ferdinandmom/miniforge3/etc/profile.d/conda.sh
+conda activate /fsx/ferdinandmom/miniforge3/envs/env-picotron
+echo python3 version = $(python3 --version)
+echo "========================"
+
+# Slurm stuff
+export HOSTNAMES=$(scontrol show hostnames "$SLURM_JOB_NODELIST")
+export MASTER_ADDR=$(scontrol show hostnames "$SLURM_JOB_NODELIST" | head -n 1)
+export MASTER_PORT=$((1024 + RANDOM % 64511))
+
+export TMPDIR=/scratch
+export TORCH_HOME=/fsx/$USER/.cache/torch
+export HF_HOME=/fsx/$USER/.cache/huggingface
+export WANDB_DIR=/fsx/$USER/.cache/wandb
+export CUBLAS_WORKSPACE_CONFIG=":4096:8"
+export CUDA_DEVICE_MAX_CONNECTIONS="1"
+export FI_PROVIDER="efa"
+
+module load cuda/12.1
+
+GIT_REPO="/fsx/ferdinandmom/ferdinand-hf/picotron/"
+CMD="$GIT_REPO/train.py --config /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp8_pp2_mbs1_ga64_sl4096/config.json"
+
+git checkout loading_big_model
+# huggingface-cli login --token $HUGGINGFACE_TOKEN
+
+LAUNCHER="torchrun --nproc_per_node=8 --nnode=32 --node_rank=$SLURM_NODEID --rdzv_endpoint ${MASTER_ADDR}:${MASTER_PORT} --rdzv_backend c10d --max_restarts 0 --tee 3"
+
+# Checkout the bench_cluster branch
+cd $GIT_REPO
+# Get the current job ID
+job_id=${SLURM_JOB_ID}
+
+# Update status to "pending" or "running" in the background
+update_status $job_id /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp8_pp2_mbs1_ga64_sl4096/status.txt &
+
+# Run the main command
+echo "Running command: $CMD"
+srun -u $LAUNCHER $CMD
+exit_status=$?
+
+job_id=$SLURM_JOB_ID
+
+# Update status based on the exit status of `srun`
+if [ $exit_status -eq 0 ]; then
+    printf "completed" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp8_pp2_mbs1_ga64_sl4096/status.txt
+else
+    if grep -q "OutOfMemoryError" /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp8_pp2_mbs1_ga64_sl4096/log_${job_id}.out; then
+        printf "oom" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp8_pp2_mbs1_ga64_sl4096/status.txt
+    elif grep -q " CUDA error: an illegal memory access" /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp8_pp2_mbs1_ga64_sl4096/log_${job_id}.out; then
+        printf "oom" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp8_pp2_mbs1_ga64_sl4096/status.txt
+    elif grep -q "Timeout" /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp8_pp2_mbs1_ga64_sl4096/log_${job_id}.out; then
+        printf "timeout" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp8_pp2_mbs1_ga64_sl4096/status.txt
+    else
+        printf "fail" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp8_pp2_mbs1_ga64_sl4096/status.txt
+    fi 
+fi

result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp8_pp2_mbs1_ga64_sl4096/metrics.csv

@@ -0,0 +1,2 @@
+run_name,status,dp,tp,pp,micro_batch_size,grad_acc,seq_len,avg_tokens_s_gpu,avg_mfu
+dp16_tp8_pp2_mbs1_ga64_sl4096,oom,16,8,2,1,64,4096,,

result_blog_posts/70b_32_node_with_dp_tp_pp/dp16_tp8_pp2_mbs1_ga64_sl4096/status.txt

@@ -0,0 +1 @@
+oom

result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp2_pp128_mbs1_ga1024_sl4096/config.json

@@ -0,0 +1,51 @@
+{
+    "distributed": {
+        "tp_size": 2,
+        "cp_size": 1,
+        "pp_size": 128,
+        "dp_size": 1,
+        "pp_engine": "1f1b",
+        "backend": "nccl",
+        "use_cpu": false
+    },
+    "model": {
+        "name": "meta-llama/Llama-2-70b-hf",
+        "num_hidden_layers": 80,
+        "num_attention_heads": 64,
+        "num_key_value_heads": 8,
+        "dtype": "bfloat16",
+        "use_flash_attention": true,
+        "use_fused_adam": false
+    },
+    "training": {
+        "seed": 42,
+        "learning_rate": 0.0003,
+        "total_train_steps": 200,
+        "seq_length": 4096,
+        "micro_batch_size": 1,
+        "gradient_accumulation_steps": 1024,
+        "num_samples": 400000,
+        "max_tokens": null
+    },
+    "dataset": {
+        "name": "roneneldan/TinyStories",
+        "num_workers": 0,
+        "num_proc": 4
+    },
+    "checkpoint": {
+        "save_dir": "result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp2_pp128_mbs1_ga1024_sl4096",
+        "save_frequency": 300,
+        "load_path": "",
+        "hf_hub_safetensors_path": "/fsx/ferdinandmom/hf_model_ckpt/Llama-2-70b-chat-hf"
+    },
+    "logging": {
+        "use_wandb": true,
+        "project_name": "picotron",
+        "run_name": "70b_32_node_with_dp_tp_pp/dp1_tp2_pp128_mbs1_ga1024_sl4096"
+    },
+    "environment": {
+        "OMP_NUM_THREADS": "1",
+        "TOKENIZERS_PARALLELISM": "false",
+        "FLASH_ATTEN": "1"
+    }
+}

result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp2_pp128_mbs1_ga1024_sl4096/job.slurm

@@ -0,0 +1,94 @@
+#!/bin/bash
+
+#SBATCH --job-name=job-picotron
+#SBATCH --time=00:30:00
+#SBATCH --partition=hopper-prod
+#SBATCH --nodes=32
+#SBATCH --gres=gpu:8
+#SBATCH --qos=normal
+#SBATCH --ntasks-per-node=1
+#SBATCH --cpus-per-task=96
+#SBATCH --exclusive
+#SBATCH --output=/fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp2_pp128_mbs1_ga1024_sl4096/log_%j.out
+#SBATCH --error=/fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp2_pp128_mbs1_ga1024_sl4096/log_%j.out
+
+# Function to update status based on squeue output
+update_status() {
+    job_id=$1
+    status_file=$2
+    # For unknown reasons, it doenst update status for pending. It only works for running 
+    while true; do
+        job_status=$(squeue --job $job_id --noheader --format=%T)
+        echo "Job status: $job_status"
+        if [ -z "$job_status" ]; then
+            # Job has finished or is not found
+            break
+        elif [ "$job_status" = "RUNNING" ]; then
+            printf "running" > $status_file
+            break
+        fi
+        sleep 10
+    done
+}
+
+# Misc initializations.
+echo "========================"
+echo "START TIME: $(date)"
+source /etc/profile.d/modules.sh
+source /fsx/ferdinandmom/miniforge3/etc/profile.d/conda.sh
+conda activate /fsx/ferdinandmom/miniforge3/envs/env-picotron
+echo python3 version = $(python3 --version)
+echo "========================"
+
+# Slurm stuff
+export HOSTNAMES=$(scontrol show hostnames "$SLURM_JOB_NODELIST")
+export MASTER_ADDR=$(scontrol show hostnames "$SLURM_JOB_NODELIST" | head -n 1)
+export MASTER_PORT=$((1024 + RANDOM % 64511))
+
+export TMPDIR=/scratch
+export TORCH_HOME=/fsx/$USER/.cache/torch
+export HF_HOME=/fsx/$USER/.cache/huggingface
+export WANDB_DIR=/fsx/$USER/.cache/wandb
+export CUBLAS_WORKSPACE_CONFIG=":4096:8"
+export CUDA_DEVICE_MAX_CONNECTIONS="1"
+export FI_PROVIDER="efa"
+
+module load cuda/12.1
+
+GIT_REPO="/fsx/ferdinandmom/ferdinand-hf/picotron/"
+CMD="$GIT_REPO/train.py --config /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp2_pp128_mbs1_ga1024_sl4096/config.json"
+
+git checkout loading_big_model
+# huggingface-cli login --token $HUGGINGFACE_TOKEN
+
+LAUNCHER="torchrun --nproc_per_node=8 --nnode=32 --node_rank=$SLURM_NODEID --rdzv_endpoint ${MASTER_ADDR}:${MASTER_PORT} --rdzv_backend c10d --max_restarts 0 --tee 3"
+
+# Checkout the bench_cluster branch
+cd $GIT_REPO
+# Get the current job ID
+job_id=${SLURM_JOB_ID}
+
+# Update status to "pending" or "running" in the background
+update_status $job_id /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp2_pp128_mbs1_ga1024_sl4096/status.txt &
+
+# Run the main command
+echo "Running command: $CMD"
+srun -u $LAUNCHER $CMD
+exit_status=$?
+
+job_id=$SLURM_JOB_ID
+
+# Update status based on the exit status of `srun`
+if [ $exit_status -eq 0 ]; then
+    printf "completed" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp2_pp128_mbs1_ga1024_sl4096/status.txt
+else
+    if grep -q "OutOfMemoryError" /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp2_pp128_mbs1_ga1024_sl4096/log_${job_id}.out; then
+        printf "oom" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp2_pp128_mbs1_ga1024_sl4096/status.txt
+    elif grep -q " CUDA error: an illegal memory access" /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp2_pp128_mbs1_ga1024_sl4096/log_${job_id}.out; then
+        printf "oom" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp2_pp128_mbs1_ga1024_sl4096/status.txt
+    elif grep -q "Timeout" /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp2_pp128_mbs1_ga1024_sl4096/log_${job_id}.out; then
+        printf "timeout" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp2_pp128_mbs1_ga1024_sl4096/status.txt
+    else
+        printf "fail" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp2_pp128_mbs1_ga1024_sl4096/status.txt
+    fi 
+fi

result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp2_pp128_mbs1_ga1024_sl4096/metrics.csv

@@ -0,0 +1,2 @@
+run_name,status,dp,tp,pp,micro_batch_size,grad_acc,seq_len,avg_tokens_s_gpu,avg_mfu
+dp1_tp2_pp128_mbs1_ga1024_sl4096,timeout,1,2,128,1,1024,4096,,

result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp2_pp128_mbs1_ga1024_sl4096/status.txt

@@ -0,0 +1 @@
+timeout

result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp4_pp64_mbs1_ga1024_sl4096/config.json

@@ -0,0 +1,51 @@
+{
+    "distributed": {
+        "tp_size": 4,
+        "cp_size": 1,
+        "pp_size": 64,
+        "dp_size": 1,
+        "pp_engine": "1f1b",
+        "backend": "nccl",
+        "use_cpu": false
+    },
+    "model": {
+        "name": "meta-llama/Llama-2-70b-hf",
+        "num_hidden_layers": 80,
+        "num_attention_heads": 64,
+        "num_key_value_heads": 8,
+        "dtype": "bfloat16",
+        "use_flash_attention": true,
+        "use_fused_adam": false
+    },
+    "training": {
+        "seed": 42,
+        "learning_rate": 0.0003,
+        "total_train_steps": 200,
+        "seq_length": 4096,
+        "micro_batch_size": 1,
+        "gradient_accumulation_steps": 1024,
+        "num_samples": 400000,
+        "max_tokens": null
+    },
+    "dataset": {
+        "name": "roneneldan/TinyStories",
+        "num_workers": 0,
+        "num_proc": 4
+    },
+    "checkpoint": {
+        "save_dir": "result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp4_pp64_mbs1_ga1024_sl4096",
+        "save_frequency": 300,
+        "load_path": "",
+        "hf_hub_safetensors_path": "/fsx/ferdinandmom/hf_model_ckpt/Llama-2-70b-chat-hf"
+    },
+    "logging": {
+        "use_wandb": true,
+        "project_name": "picotron",
+        "run_name": "70b_32_node_with_dp_tp_pp/dp1_tp4_pp64_mbs1_ga1024_sl4096"
+    },
+    "environment": {
+        "OMP_NUM_THREADS": "1",
+        "TOKENIZERS_PARALLELISM": "false",
+        "FLASH_ATTEN": "1"
+    }
+}

result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp4_pp64_mbs1_ga1024_sl4096/job.slurm

@@ -0,0 +1,94 @@
+#!/bin/bash
+
+#SBATCH --job-name=job-picotron
+#SBATCH --time=00:30:00
+#SBATCH --partition=hopper-prod
+#SBATCH --nodes=32
+#SBATCH --gres=gpu:8
+#SBATCH --qos=normal
+#SBATCH --ntasks-per-node=1
+#SBATCH --cpus-per-task=96
+#SBATCH --exclusive
+#SBATCH --output=/fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp4_pp64_mbs1_ga1024_sl4096/log_%j.out
+#SBATCH --error=/fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp4_pp64_mbs1_ga1024_sl4096/log_%j.out
+
+# Function to update status based on squeue output
+update_status() {
+    job_id=$1
+    status_file=$2
+    # For unknown reasons, it doenst update status for pending. It only works for running 
+    while true; do
+        job_status=$(squeue --job $job_id --noheader --format=%T)
+        echo "Job status: $job_status"
+        if [ -z "$job_status" ]; then
+            # Job has finished or is not found
+            break
+        elif [ "$job_status" = "RUNNING" ]; then
+            printf "running" > $status_file
+            break
+        fi
+        sleep 10
+    done
+}
+
+# Misc initializations.
+echo "========================"
+echo "START TIME: $(date)"
+source /etc/profile.d/modules.sh
+source /fsx/ferdinandmom/miniforge3/etc/profile.d/conda.sh
+conda activate /fsx/ferdinandmom/miniforge3/envs/env-picotron
+echo python3 version = $(python3 --version)
+echo "========================"
+
+# Slurm stuff
+export HOSTNAMES=$(scontrol show hostnames "$SLURM_JOB_NODELIST")
+export MASTER_ADDR=$(scontrol show hostnames "$SLURM_JOB_NODELIST" | head -n 1)
+export MASTER_PORT=$((1024 + RANDOM % 64511))
+
+export TMPDIR=/scratch
+export TORCH_HOME=/fsx/$USER/.cache/torch
+export HF_HOME=/fsx/$USER/.cache/huggingface
+export WANDB_DIR=/fsx/$USER/.cache/wandb
+export CUBLAS_WORKSPACE_CONFIG=":4096:8"
+export CUDA_DEVICE_MAX_CONNECTIONS="1"
+export FI_PROVIDER="efa"
+
+module load cuda/12.1
+
+GIT_REPO="/fsx/ferdinandmom/ferdinand-hf/picotron/"
+CMD="$GIT_REPO/train.py --config /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp4_pp64_mbs1_ga1024_sl4096/config.json"
+
+git checkout loading_big_model
+# huggingface-cli login --token $HUGGINGFACE_TOKEN
+
+LAUNCHER="torchrun --nproc_per_node=8 --nnode=32 --node_rank=$SLURM_NODEID --rdzv_endpoint ${MASTER_ADDR}:${MASTER_PORT} --rdzv_backend c10d --max_restarts 0 --tee 3"
+
+# Checkout the bench_cluster branch
+cd $GIT_REPO
+# Get the current job ID
+job_id=${SLURM_JOB_ID}
+
+# Update status to "pending" or "running" in the background
+update_status $job_id /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp4_pp64_mbs1_ga1024_sl4096/status.txt &
+
+# Run the main command
+echo "Running command: $CMD"
+srun -u $LAUNCHER $CMD
+exit_status=$?
+
+job_id=$SLURM_JOB_ID
+
+# Update status based on the exit status of `srun`
+if [ $exit_status -eq 0 ]; then
+    printf "completed" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp4_pp64_mbs1_ga1024_sl4096/status.txt
+else
+    if grep -q "OutOfMemoryError" /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp4_pp64_mbs1_ga1024_sl4096/log_${job_id}.out; then
+        printf "oom" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp4_pp64_mbs1_ga1024_sl4096/status.txt
+    elif grep -q " CUDA error: an illegal memory access" /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp4_pp64_mbs1_ga1024_sl4096/log_${job_id}.out; then
+        printf "oom" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp4_pp64_mbs1_ga1024_sl4096/status.txt
+    elif grep -q "Timeout" /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp4_pp64_mbs1_ga1024_sl4096/log_${job_id}.out; then
+        printf "timeout" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp4_pp64_mbs1_ga1024_sl4096/status.txt
+    else
+        printf "fail" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp4_pp64_mbs1_ga1024_sl4096/status.txt
+    fi 
+fi

result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp4_pp64_mbs1_ga1024_sl4096/metrics.csv

@@ -0,0 +1,2 @@
+run_name,status,dp,tp,pp,micro_batch_size,grad_acc,seq_len,avg_tokens_s_gpu,avg_mfu
+dp1_tp4_pp64_mbs1_ga1024_sl4096,fail,1,4,64,1,1024,4096,,

result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp4_pp64_mbs1_ga1024_sl4096/status.txt

@@ -0,0 +1 @@
+fail

result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp8_pp32_mbs1_ga1024_sl4096/config.json

@@ -0,0 +1,51 @@
+{
+    "distributed": {
+        "tp_size": 8,
+        "cp_size": 1,
+        "pp_size": 32,
+        "dp_size": 1,
+        "pp_engine": "1f1b",
+        "backend": "nccl",
+        "use_cpu": false
+    },
+    "model": {
+        "name": "meta-llama/Llama-2-70b-hf",
+        "num_hidden_layers": 80,
+        "num_attention_heads": 64,
+        "num_key_value_heads": 8,
+        "dtype": "bfloat16",
+        "use_flash_attention": true,
+        "use_fused_adam": false
+    },
+    "training": {
+        "seed": 42,
+        "learning_rate": 0.0003,
+        "total_train_steps": 200,
+        "seq_length": 4096,
+        "micro_batch_size": 1,
+        "gradient_accumulation_steps": 1024,
+        "num_samples": 400000,
+        "max_tokens": null
+    },
+    "dataset": {
+        "name": "roneneldan/TinyStories",
+        "num_workers": 0,
+        "num_proc": 4
+    },
+    "checkpoint": {
+        "save_dir": "result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp8_pp32_mbs1_ga1024_sl4096",
+        "save_frequency": 300,
+        "load_path": "",
+        "hf_hub_safetensors_path": "/fsx/ferdinandmom/hf_model_ckpt/Llama-2-70b-chat-hf"
+    },
+    "logging": {
+        "use_wandb": true,
+        "project_name": "picotron",
+        "run_name": "70b_32_node_with_dp_tp_pp/dp1_tp8_pp32_mbs1_ga1024_sl4096"
+    },
+    "environment": {
+        "OMP_NUM_THREADS": "1",
+        "TOKENIZERS_PARALLELISM": "false",
+        "FLASH_ATTEN": "1"
+    }
+}

result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp8_pp32_mbs1_ga1024_sl4096/job.slurm

@@ -0,0 +1,94 @@
+#!/bin/bash
+
+#SBATCH --job-name=job-picotron
+#SBATCH --time=00:30:00
+#SBATCH --partition=hopper-prod
+#SBATCH --nodes=32
+#SBATCH --gres=gpu:8
+#SBATCH --qos=normal
+#SBATCH --ntasks-per-node=1
+#SBATCH --cpus-per-task=96
+#SBATCH --exclusive
+#SBATCH --output=/fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp8_pp32_mbs1_ga1024_sl4096/log_%j.out
+#SBATCH --error=/fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp8_pp32_mbs1_ga1024_sl4096/log_%j.out
+
+# Function to update status based on squeue output
+update_status() {
+    job_id=$1
+    status_file=$2
+    # For unknown reasons, it doenst update status for pending. It only works for running 
+    while true; do
+        job_status=$(squeue --job $job_id --noheader --format=%T)
+        echo "Job status: $job_status"
+        if [ -z "$job_status" ]; then
+            # Job has finished or is not found
+            break
+        elif [ "$job_status" = "RUNNING" ]; then
+            printf "running" > $status_file
+            break
+        fi
+        sleep 10
+    done
+}
+
+# Misc initializations.
+echo "========================"
+echo "START TIME: $(date)"
+source /etc/profile.d/modules.sh
+source /fsx/ferdinandmom/miniforge3/etc/profile.d/conda.sh
+conda activate /fsx/ferdinandmom/miniforge3/envs/env-picotron
+echo python3 version = $(python3 --version)
+echo "========================"
+
+# Slurm stuff
+export HOSTNAMES=$(scontrol show hostnames "$SLURM_JOB_NODELIST")
+export MASTER_ADDR=$(scontrol show hostnames "$SLURM_JOB_NODELIST" | head -n 1)
+export MASTER_PORT=$((1024 + RANDOM % 64511))
+
+export TMPDIR=/scratch
+export TORCH_HOME=/fsx/$USER/.cache/torch
+export HF_HOME=/fsx/$USER/.cache/huggingface
+export WANDB_DIR=/fsx/$USER/.cache/wandb
+export CUBLAS_WORKSPACE_CONFIG=":4096:8"
+export CUDA_DEVICE_MAX_CONNECTIONS="1"
+export FI_PROVIDER="efa"
+
+module load cuda/12.1
+
+GIT_REPO="/fsx/ferdinandmom/ferdinand-hf/picotron/"
+CMD="$GIT_REPO/train.py --config /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp8_pp32_mbs1_ga1024_sl4096/config.json"
+
+git checkout loading_big_model
+# huggingface-cli login --token $HUGGINGFACE_TOKEN
+
+LAUNCHER="torchrun --nproc_per_node=8 --nnode=32 --node_rank=$SLURM_NODEID --rdzv_endpoint ${MASTER_ADDR}:${MASTER_PORT} --rdzv_backend c10d --max_restarts 0 --tee 3"
+
+# Checkout the bench_cluster branch
+cd $GIT_REPO
+# Get the current job ID
+job_id=${SLURM_JOB_ID}
+
+# Update status to "pending" or "running" in the background
+update_status $job_id /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp8_pp32_mbs1_ga1024_sl4096/status.txt &
+
+# Run the main command
+echo "Running command: $CMD"
+srun -u $LAUNCHER $CMD
+exit_status=$?
+
+job_id=$SLURM_JOB_ID
+
+# Update status based on the exit status of `srun`
+if [ $exit_status -eq 0 ]; then
+    printf "completed" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp8_pp32_mbs1_ga1024_sl4096/status.txt
+else
+    if grep -q "OutOfMemoryError" /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp8_pp32_mbs1_ga1024_sl4096/log_${job_id}.out; then
+        printf "oom" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp8_pp32_mbs1_ga1024_sl4096/status.txt
+    elif grep -q " CUDA error: an illegal memory access" /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp8_pp32_mbs1_ga1024_sl4096/log_${job_id}.out; then
+        printf "oom" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp8_pp32_mbs1_ga1024_sl4096/status.txt
+    elif grep -q "Timeout" /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp8_pp32_mbs1_ga1024_sl4096/log_${job_id}.out; then
+        printf "timeout" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp8_pp32_mbs1_ga1024_sl4096/status.txt
+    else
+        printf "fail" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp8_pp32_mbs1_ga1024_sl4096/status.txt
+    fi 
+fi

result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp8_pp32_mbs1_ga1024_sl4096/metrics.csv

@@ -0,0 +1,2 @@
+run_name,status,dp,tp,pp,micro_batch_size,grad_acc,seq_len,avg_tokens_s_gpu,avg_mfu
+dp1_tp8_pp32_mbs1_ga1024_sl4096,running,1,8,32,1,1024,4096,127,6

result_blog_posts/70b_32_node_with_dp_tp_pp/dp1_tp8_pp32_mbs1_ga1024_sl4096/status.txt

@@ -0,0 +1 @@
+running

result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp2_pp64_mbs1_ga512_sl4096/config.json

@@ -0,0 +1,51 @@
+{
+    "distributed": {
+        "tp_size": 2,
+        "cp_size": 1,
+        "pp_size": 64,
+        "dp_size": 2,
+        "pp_engine": "1f1b",
+        "backend": "nccl",
+        "use_cpu": false
+    },
+    "model": {
+        "name": "meta-llama/Llama-2-70b-hf",
+        "num_hidden_layers": 80,
+        "num_attention_heads": 64,
+        "num_key_value_heads": 8,
+        "dtype": "bfloat16",
+        "use_flash_attention": true,
+        "use_fused_adam": false
+    },
+    "training": {
+        "seed": 42,
+        "learning_rate": 0.0003,
+        "total_train_steps": 200,
+        "seq_length": 4096,
+        "micro_batch_size": 1,
+        "gradient_accumulation_steps": 512,
+        "num_samples": 400000,
+        "max_tokens": null
+    },
+    "dataset": {
+        "name": "roneneldan/TinyStories",
+        "num_workers": 0,
+        "num_proc": 4
+    },
+    "checkpoint": {
+        "save_dir": "result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp2_pp64_mbs1_ga512_sl4096",
+        "save_frequency": 300,
+        "load_path": "",
+        "hf_hub_safetensors_path": "/fsx/ferdinandmom/hf_model_ckpt/Llama-2-70b-chat-hf"
+    },
+    "logging": {
+        "use_wandb": true,
+        "project_name": "picotron",
+        "run_name": "70b_32_node_with_dp_tp_pp/dp2_tp2_pp64_mbs1_ga512_sl4096"
+    },
+    "environment": {
+        "OMP_NUM_THREADS": "1",
+        "TOKENIZERS_PARALLELISM": "false",
+        "FLASH_ATTEN": "1"
+    }
+}

result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp2_pp64_mbs1_ga512_sl4096/job.slurm

@@ -0,0 +1,94 @@
+#!/bin/bash
+
+#SBATCH --job-name=job-picotron
+#SBATCH --time=00:30:00
+#SBATCH --partition=hopper-prod
+#SBATCH --nodes=32
+#SBATCH --gres=gpu:8
+#SBATCH --qos=normal
+#SBATCH --ntasks-per-node=1
+#SBATCH --cpus-per-task=96
+#SBATCH --exclusive
+#SBATCH --output=/fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp2_pp64_mbs1_ga512_sl4096/log_%j.out
+#SBATCH --error=/fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp2_pp64_mbs1_ga512_sl4096/log_%j.out
+
+# Function to update status based on squeue output
+update_status() {
+    job_id=$1
+    status_file=$2
+    # For unknown reasons, it doenst update status for pending. It only works for running 
+    while true; do
+        job_status=$(squeue --job $job_id --noheader --format=%T)
+        echo "Job status: $job_status"
+        if [ -z "$job_status" ]; then
+            # Job has finished or is not found
+            break
+        elif [ "$job_status" = "RUNNING" ]; then
+            printf "running" > $status_file
+            break
+        fi
+        sleep 10
+    done
+}
+
+# Misc initializations.
+echo "========================"
+echo "START TIME: $(date)"
+source /etc/profile.d/modules.sh
+source /fsx/ferdinandmom/miniforge3/etc/profile.d/conda.sh
+conda activate /fsx/ferdinandmom/miniforge3/envs/env-picotron
+echo python3 version = $(python3 --version)
+echo "========================"
+
+# Slurm stuff
+export HOSTNAMES=$(scontrol show hostnames "$SLURM_JOB_NODELIST")
+export MASTER_ADDR=$(scontrol show hostnames "$SLURM_JOB_NODELIST" | head -n 1)
+export MASTER_PORT=$((1024 + RANDOM % 64511))
+
+export TMPDIR=/scratch
+export TORCH_HOME=/fsx/$USER/.cache/torch
+export HF_HOME=/fsx/$USER/.cache/huggingface
+export WANDB_DIR=/fsx/$USER/.cache/wandb
+export CUBLAS_WORKSPACE_CONFIG=":4096:8"
+export CUDA_DEVICE_MAX_CONNECTIONS="1"
+export FI_PROVIDER="efa"
+
+module load cuda/12.1
+
+GIT_REPO="/fsx/ferdinandmom/ferdinand-hf/picotron/"
+CMD="$GIT_REPO/train.py --config /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp2_pp64_mbs1_ga512_sl4096/config.json"
+
+git checkout loading_big_model
+# huggingface-cli login --token $HUGGINGFACE_TOKEN
+
+LAUNCHER="torchrun --nproc_per_node=8 --nnode=32 --node_rank=$SLURM_NODEID --rdzv_endpoint ${MASTER_ADDR}:${MASTER_PORT} --rdzv_backend c10d --max_restarts 0 --tee 3"
+
+# Checkout the bench_cluster branch
+cd $GIT_REPO
+# Get the current job ID
+job_id=${SLURM_JOB_ID}
+
+# Update status to "pending" or "running" in the background
+update_status $job_id /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp2_pp64_mbs1_ga512_sl4096/status.txt &
+
+# Run the main command
+echo "Running command: $CMD"
+srun -u $LAUNCHER $CMD
+exit_status=$?
+
+job_id=$SLURM_JOB_ID
+
+# Update status based on the exit status of `srun`
+if [ $exit_status -eq 0 ]; then
+    printf "completed" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp2_pp64_mbs1_ga512_sl4096/status.txt
+else
+    if grep -q "OutOfMemoryError" /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp2_pp64_mbs1_ga512_sl4096/log_${job_id}.out; then
+        printf "oom" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp2_pp64_mbs1_ga512_sl4096/status.txt
+    elif grep -q " CUDA error: an illegal memory access" /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp2_pp64_mbs1_ga512_sl4096/log_${job_id}.out; then
+        printf "oom" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp2_pp64_mbs1_ga512_sl4096/status.txt
+    elif grep -q "Timeout" /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp2_pp64_mbs1_ga512_sl4096/log_${job_id}.out; then
+        printf "timeout" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp2_pp64_mbs1_ga512_sl4096/status.txt
+    else
+        printf "fail" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp2_pp64_mbs1_ga512_sl4096/status.txt
+    fi 
+fi

result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp2_pp64_mbs1_ga512_sl4096/metrics.csv

@@ -0,0 +1,2 @@
+run_name,status,dp,tp,pp,micro_batch_size,grad_acc,seq_len,avg_tokens_s_gpu,avg_mfu
+dp2_tp2_pp64_mbs1_ga512_sl4096,oom,2,2,64,1,512,4096,,

result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp2_pp64_mbs1_ga512_sl4096/status.txt

@@ -0,0 +1 @@
+oom

result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp4_pp32_mbs1_ga512_sl4096/config.json

@@ -0,0 +1,51 @@
+{
+    "distributed": {
+        "tp_size": 4,
+        "cp_size": 1,
+        "pp_size": 32,
+        "dp_size": 2,
+        "pp_engine": "1f1b",
+        "backend": "nccl",
+        "use_cpu": false
+    },
+    "model": {
+        "name": "meta-llama/Llama-2-70b-hf",
+        "num_hidden_layers": 80,
+        "num_attention_heads": 64,
+        "num_key_value_heads": 8,
+        "dtype": "bfloat16",
+        "use_flash_attention": true,
+        "use_fused_adam": false
+    },
+    "training": {
+        "seed": 42,
+        "learning_rate": 0.0003,
+        "total_train_steps": 200,
+        "seq_length": 4096,
+        "micro_batch_size": 1,
+        "gradient_accumulation_steps": 512,
+        "num_samples": 400000,
+        "max_tokens": null
+    },
+    "dataset": {
+        "name": "roneneldan/TinyStories",
+        "num_workers": 0,
+        "num_proc": 4
+    },
+    "checkpoint": {
+        "save_dir": "result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp4_pp32_mbs1_ga512_sl4096",
+        "save_frequency": 300,
+        "load_path": "",
+        "hf_hub_safetensors_path": "/fsx/ferdinandmom/hf_model_ckpt/Llama-2-70b-chat-hf"
+    },
+    "logging": {
+        "use_wandb": true,
+        "project_name": "picotron",
+        "run_name": "70b_32_node_with_dp_tp_pp/dp2_tp4_pp32_mbs1_ga512_sl4096"
+    },
+    "environment": {
+        "OMP_NUM_THREADS": "1",
+        "TOKENIZERS_PARALLELISM": "false",
+        "FLASH_ATTEN": "1"
+    }
+}

result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp4_pp32_mbs1_ga512_sl4096/job.slurm

@@ -0,0 +1,94 @@
+#!/bin/bash
+
+#SBATCH --job-name=job-picotron
+#SBATCH --time=00:30:00
+#SBATCH --partition=hopper-prod
+#SBATCH --nodes=32
+#SBATCH --gres=gpu:8
+#SBATCH --qos=normal
+#SBATCH --ntasks-per-node=1
+#SBATCH --cpus-per-task=96
+#SBATCH --exclusive
+#SBATCH --output=/fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp4_pp32_mbs1_ga512_sl4096/log_%j.out
+#SBATCH --error=/fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp4_pp32_mbs1_ga512_sl4096/log_%j.out
+
+# Function to update status based on squeue output
+update_status() {
+    job_id=$1
+    status_file=$2
+    # For unknown reasons, it doenst update status for pending. It only works for running 
+    while true; do
+        job_status=$(squeue --job $job_id --noheader --format=%T)
+        echo "Job status: $job_status"
+        if [ -z "$job_status" ]; then
+            # Job has finished or is not found
+            break
+        elif [ "$job_status" = "RUNNING" ]; then
+            printf "running" > $status_file
+            break
+        fi
+        sleep 10
+    done
+}
+
+# Misc initializations.
+echo "========================"
+echo "START TIME: $(date)"
+source /etc/profile.d/modules.sh
+source /fsx/ferdinandmom/miniforge3/etc/profile.d/conda.sh
+conda activate /fsx/ferdinandmom/miniforge3/envs/env-picotron
+echo python3 version = $(python3 --version)
+echo "========================"
+
+# Slurm stuff
+export HOSTNAMES=$(scontrol show hostnames "$SLURM_JOB_NODELIST")
+export MASTER_ADDR=$(scontrol show hostnames "$SLURM_JOB_NODELIST" | head -n 1)
+export MASTER_PORT=$((1024 + RANDOM % 64511))
+
+export TMPDIR=/scratch
+export TORCH_HOME=/fsx/$USER/.cache/torch
+export HF_HOME=/fsx/$USER/.cache/huggingface
+export WANDB_DIR=/fsx/$USER/.cache/wandb
+export CUBLAS_WORKSPACE_CONFIG=":4096:8"
+export CUDA_DEVICE_MAX_CONNECTIONS="1"
+export FI_PROVIDER="efa"
+
+module load cuda/12.1
+
+GIT_REPO="/fsx/ferdinandmom/ferdinand-hf/picotron/"
+CMD="$GIT_REPO/train.py --config /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp4_pp32_mbs1_ga512_sl4096/config.json"
+
+git checkout loading_big_model
+# huggingface-cli login --token $HUGGINGFACE_TOKEN
+
+LAUNCHER="torchrun --nproc_per_node=8 --nnode=32 --node_rank=$SLURM_NODEID --rdzv_endpoint ${MASTER_ADDR}:${MASTER_PORT} --rdzv_backend c10d --max_restarts 0 --tee 3"
+
+# Checkout the bench_cluster branch
+cd $GIT_REPO
+# Get the current job ID
+job_id=${SLURM_JOB_ID}
+
+# Update status to "pending" or "running" in the background
+update_status $job_id /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp4_pp32_mbs1_ga512_sl4096/status.txt &
+
+# Run the main command
+echo "Running command: $CMD"
+srun -u $LAUNCHER $CMD
+exit_status=$?
+
+job_id=$SLURM_JOB_ID
+
+# Update status based on the exit status of `srun`
+if [ $exit_status -eq 0 ]; then
+    printf "completed" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp4_pp32_mbs1_ga512_sl4096/status.txt
+else
+    if grep -q "OutOfMemoryError" /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp4_pp32_mbs1_ga512_sl4096/log_${job_id}.out; then
+        printf "oom" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp4_pp32_mbs1_ga512_sl4096/status.txt
+    elif grep -q " CUDA error: an illegal memory access" /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp4_pp32_mbs1_ga512_sl4096/log_${job_id}.out; then
+        printf "oom" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp4_pp32_mbs1_ga512_sl4096/status.txt
+    elif grep -q "Timeout" /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp4_pp32_mbs1_ga512_sl4096/log_${job_id}.out; then
+        printf "timeout" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp4_pp32_mbs1_ga512_sl4096/status.txt
+    else
+        printf "fail" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp4_pp32_mbs1_ga512_sl4096/status.txt
+    fi 
+fi

result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp4_pp32_mbs1_ga512_sl4096/metrics.csv

@@ -0,0 +1,2 @@
+run_name,status,dp,tp,pp,micro_batch_size,grad_acc,seq_len,avg_tokens_s_gpu,avg_mfu
+dp2_tp4_pp32_mbs1_ga512_sl4096,timeout,2,4,32,1,512,4096,,

result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp4_pp32_mbs1_ga512_sl4096/status.txt

@@ -0,0 +1 @@
+timeout

result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp8_pp16_mbs1_ga512_sl4096/config.json

@@ -0,0 +1,51 @@
+{
+    "distributed": {
+        "tp_size": 8,
+        "cp_size": 1,
+        "pp_size": 16,
+        "dp_size": 2,
+        "pp_engine": "1f1b",
+        "backend": "nccl",
+        "use_cpu": false
+    },
+    "model": {
+        "name": "meta-llama/Llama-2-70b-hf",
+        "num_hidden_layers": 80,
+        "num_attention_heads": 64,
+        "num_key_value_heads": 8,
+        "dtype": "bfloat16",
+        "use_flash_attention": true,
+        "use_fused_adam": false
+    },
+    "training": {
+        "seed": 42,
+        "learning_rate": 0.0003,
+        "total_train_steps": 200,
+        "seq_length": 4096,
+        "micro_batch_size": 1,
+        "gradient_accumulation_steps": 512,
+        "num_samples": 400000,
+        "max_tokens": null
+    },
+    "dataset": {
+        "name": "roneneldan/TinyStories",
+        "num_workers": 0,
+        "num_proc": 4
+    },
+    "checkpoint": {
+        "save_dir": "result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp8_pp16_mbs1_ga512_sl4096",
+        "save_frequency": 300,
+        "load_path": "",
+        "hf_hub_safetensors_path": "/fsx/ferdinandmom/hf_model_ckpt/Llama-2-70b-chat-hf"
+    },
+    "logging": {
+        "use_wandb": true,
+        "project_name": "picotron",
+        "run_name": "70b_32_node_with_dp_tp_pp/dp2_tp8_pp16_mbs1_ga512_sl4096"
+    },
+    "environment": {
+        "OMP_NUM_THREADS": "1",
+        "TOKENIZERS_PARALLELISM": "false",
+        "FLASH_ATTEN": "1"
+    }
+}

result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp8_pp16_mbs1_ga512_sl4096/job.slurm

@@ -0,0 +1,94 @@
+#!/bin/bash
+
+#SBATCH --job-name=job-picotron
+#SBATCH --time=00:30:00
+#SBATCH --partition=hopper-prod
+#SBATCH --nodes=32
+#SBATCH --gres=gpu:8
+#SBATCH --qos=normal
+#SBATCH --ntasks-per-node=1
+#SBATCH --cpus-per-task=96
+#SBATCH --exclusive
+#SBATCH --output=/fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp8_pp16_mbs1_ga512_sl4096/log_%j.out
+#SBATCH --error=/fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp8_pp16_mbs1_ga512_sl4096/log_%j.out
+
+# Function to update status based on squeue output
+update_status() {
+    job_id=$1
+    status_file=$2
+    # For unknown reasons, it doenst update status for pending. It only works for running 
+    while true; do
+        job_status=$(squeue --job $job_id --noheader --format=%T)
+        echo "Job status: $job_status"
+        if [ -z "$job_status" ]; then
+            # Job has finished or is not found
+            break
+        elif [ "$job_status" = "RUNNING" ]; then
+            printf "running" > $status_file
+            break
+        fi
+        sleep 10
+    done
+}
+
+# Misc initializations.
+echo "========================"
+echo "START TIME: $(date)"
+source /etc/profile.d/modules.sh
+source /fsx/ferdinandmom/miniforge3/etc/profile.d/conda.sh
+conda activate /fsx/ferdinandmom/miniforge3/envs/env-picotron
+echo python3 version = $(python3 --version)
+echo "========================"
+
+# Slurm stuff
+export HOSTNAMES=$(scontrol show hostnames "$SLURM_JOB_NODELIST")
+export MASTER_ADDR=$(scontrol show hostnames "$SLURM_JOB_NODELIST" | head -n 1)
+export MASTER_PORT=$((1024 + RANDOM % 64511))
+
+export TMPDIR=/scratch
+export TORCH_HOME=/fsx/$USER/.cache/torch
+export HF_HOME=/fsx/$USER/.cache/huggingface
+export WANDB_DIR=/fsx/$USER/.cache/wandb
+export CUBLAS_WORKSPACE_CONFIG=":4096:8"
+export CUDA_DEVICE_MAX_CONNECTIONS="1"
+export FI_PROVIDER="efa"
+
+module load cuda/12.1
+
+GIT_REPO="/fsx/ferdinandmom/ferdinand-hf/picotron/"
+CMD="$GIT_REPO/train.py --config /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp8_pp16_mbs1_ga512_sl4096/config.json"
+
+git checkout loading_big_model
+# huggingface-cli login --token $HUGGINGFACE_TOKEN
+
+LAUNCHER="torchrun --nproc_per_node=8 --nnode=32 --node_rank=$SLURM_NODEID --rdzv_endpoint ${MASTER_ADDR}:${MASTER_PORT} --rdzv_backend c10d --max_restarts 0 --tee 3"
+
+# Checkout the bench_cluster branch
+cd $GIT_REPO
+# Get the current job ID
+job_id=${SLURM_JOB_ID}
+
+# Update status to "pending" or "running" in the background
+update_status $job_id /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp8_pp16_mbs1_ga512_sl4096/status.txt &
+
+# Run the main command
+echo "Running command: $CMD"
+srun -u $LAUNCHER $CMD
+exit_status=$?
+
+job_id=$SLURM_JOB_ID
+
+# Update status based on the exit status of `srun`
+if [ $exit_status -eq 0 ]; then
+    printf "completed" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp8_pp16_mbs1_ga512_sl4096/status.txt
+else
+    if grep -q "OutOfMemoryError" /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp8_pp16_mbs1_ga512_sl4096/log_${job_id}.out; then
+        printf "oom" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp8_pp16_mbs1_ga512_sl4096/status.txt
+    elif grep -q " CUDA error: an illegal memory access" /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp8_pp16_mbs1_ga512_sl4096/log_${job_id}.out; then
+        printf "oom" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp8_pp16_mbs1_ga512_sl4096/status.txt
+    elif grep -q "Timeout" /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp8_pp16_mbs1_ga512_sl4096/log_${job_id}.out; then
+        printf "timeout" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp8_pp16_mbs1_ga512_sl4096/status.txt
+    else
+        printf "fail" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp8_pp16_mbs1_ga512_sl4096/status.txt
+    fi 
+fi

result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp8_pp16_mbs1_ga512_sl4096/metrics.csv

@@ -0,0 +1,2 @@
+run_name,status,dp,tp,pp,micro_batch_size,grad_acc,seq_len,avg_tokens_s_gpu,avg_mfu
+dp2_tp8_pp16_mbs1_ga512_sl4096,running,2,8,16,1,512,4096,208,9

result_blog_posts/70b_32_node_with_dp_tp_pp/dp2_tp8_pp16_mbs1_ga512_sl4096/status.txt

@@ -0,0 +1 @@
+running

result_blog_posts/70b_32_node_with_dp_tp_pp/dp32_tp2_pp4_mbs1_ga32_sl4096/config.json

@@ -0,0 +1,51 @@
+{
+    "distributed": {
+        "tp_size": 2,
+        "cp_size": 1,
+        "pp_size": 4,
+        "dp_size": 32,
+        "pp_engine": "1f1b",
+        "backend": "nccl",
+        "use_cpu": false
+    },
+    "model": {
+        "name": "meta-llama/Llama-2-70b-hf",
+        "num_hidden_layers": 80,
+        "num_attention_heads": 64,
+        "num_key_value_heads": 8,
+        "dtype": "bfloat16",
+        "use_flash_attention": true,
+        "use_fused_adam": false
+    },
+    "training": {
+        "seed": 42,
+        "learning_rate": 0.0003,
+        "total_train_steps": 200,
+        "seq_length": 4096,
+        "micro_batch_size": 1,
+        "gradient_accumulation_steps": 32,
+        "num_samples": 400000,
+        "max_tokens": null
+    },
+    "dataset": {
+        "name": "roneneldan/TinyStories",
+        "num_workers": 0,
+        "num_proc": 4
+    },
+    "checkpoint": {
+        "save_dir": "result_blog_posts/70b_32_node_with_dp_tp_pp/dp32_tp2_pp4_mbs1_ga32_sl4096",
+        "save_frequency": 300,
+        "load_path": "",
+        "hf_hub_safetensors_path": "/fsx/ferdinandmom/hf_model_ckpt/Llama-2-70b-chat-hf"
+    },
+    "logging": {
+        "use_wandb": true,
+        "project_name": "picotron",
+        "run_name": "70b_32_node_with_dp_tp_pp/dp32_tp2_pp4_mbs1_ga32_sl4096"
+    },
+    "environment": {
+        "OMP_NUM_THREADS": "1",
+        "TOKENIZERS_PARALLELISM": "false",
+        "FLASH_ATTEN": "1"
+    }
+}

result_blog_posts/70b_32_node_with_dp_tp_pp/dp32_tp2_pp4_mbs1_ga32_sl4096/job.slurm

@@ -0,0 +1,94 @@
+#!/bin/bash
+
+#SBATCH --job-name=job-picotron
+#SBATCH --time=00:30:00
+#SBATCH --partition=hopper-prod
+#SBATCH --nodes=32
+#SBATCH --gres=gpu:8
+#SBATCH --qos=normal
+#SBATCH --ntasks-per-node=1
+#SBATCH --cpus-per-task=96
+#SBATCH --exclusive
+#SBATCH --output=/fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp32_tp2_pp4_mbs1_ga32_sl4096/log_%j.out
+#SBATCH --error=/fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp32_tp2_pp4_mbs1_ga32_sl4096/log_%j.out
+
+# Function to update status based on squeue output
+update_status() {
+    job_id=$1
+    status_file=$2
+    # For unknown reasons, it doenst update status for pending. It only works for running 
+    while true; do
+        job_status=$(squeue --job $job_id --noheader --format=%T)
+        echo "Job status: $job_status"
+        if [ -z "$job_status" ]; then
+            # Job has finished or is not found
+            break
+        elif [ "$job_status" = "RUNNING" ]; then
+            printf "running" > $status_file
+            break
+        fi
+        sleep 10
+    done
+}
+
+# Misc initializations.
+echo "========================"
+echo "START TIME: $(date)"
+source /etc/profile.d/modules.sh
+source /fsx/ferdinandmom/miniforge3/etc/profile.d/conda.sh
+conda activate /fsx/ferdinandmom/miniforge3/envs/env-picotron
+echo python3 version = $(python3 --version)
+echo "========================"
+
+# Slurm stuff
+export HOSTNAMES=$(scontrol show hostnames "$SLURM_JOB_NODELIST")
+export MASTER_ADDR=$(scontrol show hostnames "$SLURM_JOB_NODELIST" | head -n 1)
+export MASTER_PORT=$((1024 + RANDOM % 64511))
+
+export TMPDIR=/scratch
+export TORCH_HOME=/fsx/$USER/.cache/torch
+export HF_HOME=/fsx/$USER/.cache/huggingface
+export WANDB_DIR=/fsx/$USER/.cache/wandb
+export CUBLAS_WORKSPACE_CONFIG=":4096:8"
+export CUDA_DEVICE_MAX_CONNECTIONS="1"
+export FI_PROVIDER="efa"
+
+module load cuda/12.1
+
+GIT_REPO="/fsx/ferdinandmom/ferdinand-hf/picotron/"
+CMD="$GIT_REPO/train.py --config /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp32_tp2_pp4_mbs1_ga32_sl4096/config.json"
+
+git checkout loading_big_model
+# huggingface-cli login --token $HUGGINGFACE_TOKEN
+
+LAUNCHER="torchrun --nproc_per_node=8 --nnode=32 --node_rank=$SLURM_NODEID --rdzv_endpoint ${MASTER_ADDR}:${MASTER_PORT} --rdzv_backend c10d --max_restarts 0 --tee 3"
+
+# Checkout the bench_cluster branch
+cd $GIT_REPO
+# Get the current job ID
+job_id=${SLURM_JOB_ID}
+
+# Update status to "pending" or "running" in the background
+update_status $job_id /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp32_tp2_pp4_mbs1_ga32_sl4096/status.txt &
+
+# Run the main command
+echo "Running command: $CMD"
+srun -u $LAUNCHER $CMD
+exit_status=$?
+
+job_id=$SLURM_JOB_ID
+
+# Update status based on the exit status of `srun`
+if [ $exit_status -eq 0 ]; then
+    printf "completed" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp32_tp2_pp4_mbs1_ga32_sl4096/status.txt
+else
+    if grep -q "OutOfMemoryError" /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp32_tp2_pp4_mbs1_ga32_sl4096/log_${job_id}.out; then
+        printf "oom" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp32_tp2_pp4_mbs1_ga32_sl4096/status.txt
+    elif grep -q " CUDA error: an illegal memory access" /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp32_tp2_pp4_mbs1_ga32_sl4096/log_${job_id}.out; then
+        printf "oom" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp32_tp2_pp4_mbs1_ga32_sl4096/status.txt
+    elif grep -q "Timeout" /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp32_tp2_pp4_mbs1_ga32_sl4096/log_${job_id}.out; then
+        printf "timeout" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp32_tp2_pp4_mbs1_ga32_sl4096/status.txt
+    else
+        printf "fail" > /fsx/ferdinandmom/ferdinand-hf/picotron/result_blog_posts/70b_32_node_with_dp_tp_pp/dp32_tp2_pp4_mbs1_ga32_sl4096/status.txt
+    fi 
+fi

result_blog_posts/70b_32_node_with_dp_tp_pp/dp32_tp2_pp4_mbs1_ga32_sl4096/metrics.csv

@@ -0,0 +1,2 @@
+run_name,status,dp,tp,pp,micro_batch_size,grad_acc,seq_len,avg_tokens_s_gpu,avg_mfu
+dp32_tp2_pp4_mbs1_ga32_sl4096,oom,32,2,4,1,32,4096,,

result_blog_posts/70b_32_node_with_dp_tp_pp/dp32_tp2_pp4_mbs1_ga32_sl4096/status.txt

@@ -0,0 +1 @@
+oom