add scripts

.gitignore

@@ -8,7 +8,7 @@ venv/
 workdir/
 makefile
 *.out
-*.sh
+# *.sh
 *.swp
 */data/
 *events.out.tfevents*

script/jax/a100/auto_run_array.sh

@@ -0,0 +1,104 @@
+#!/bin/bash
+
+module load cuda/12.1
+
+source ~/miniforge3/etc/profile.d/conda.sh
+conda activate alpe
+
+# forcefully enable TF32
+export NVIDIA_TF32_OVERRIDE=1
+
+# Env vars
+export OMP_NUM_THREADS=2  # TODO: check
+export HOME=/home/najroldi
+export CODE_DIR=/home/najroldi/algorithmic-efficiency
+export EXP_DIR=/fast/najroldi/exp/algoperf
+export DATA_DIR=/fast/najroldi/data
+
+# Job specific vars
+workload_or_id=$1
+framework=$2
+submission=$3
+search_space=$4
+name=$5
+study=$6
+num_tuning_trials=$7
+rng_seed=$8
+
+workload_list=(
+  criteo1tb
+  fastmri
+  imagenet_resnet
+  imagenet_vit
+  librispeech_conformer
+  librispeech_deepspeech
+  ogbg
+  wmt
+)
+
+# in bash need to explicitly declare an associative array
+declare -A workload_to_dataset=(
+  [criteo1tb]="criteo1tb"
+  [fastmri]="fastmri"
+  [imagenet_resnet]="imagenet"
+  [imagenet_vit]="imagenet"
+  [librispeech_conformer]="librispeech"
+  [librispeech_deepspeech]="librispeech"
+  [ogbg]="ogbg"
+  [wmt]="wmt"
+)
+
+# Determine workload based on workload_or_id
+if [[ "$workload_or_id" =~ ^[0-7]+$ ]]; then
+  # workload_or_id is a valid index (0-7), treat it as job ID and get workload from list
+  job_id=$workload_or_id
+  workload=${workload_list[$job_id]}
+else
+  # treat workload_or_id as an explicit workload name
+  workload=$workload_or_id
+fi
+
+dataset=${workload_to_dataset[$workload]}
+
+# Experiment name
+# experiment_name="${name}_${workload}_${framework}"
+experiment_name=${name}
+
+# Librispeech tokenizer path
+tokenizer_path=''
+if [ "$dataset" = "librispeech" ]; then
+    tokenizer_path="${DATA_DIR}/librispeech/spm_model.vocab"
+fi
+
+# Increase num_workers on imagenet
+eval_num_workers=0
+if [ "$dataset" == "imagenet" ] && [ "$framework" == "pytorch" ]; then
+  eval_num_workers=4
+fi
+
+# Execute python script
+python \
+  $CODE_DIR/submission_runner.py \
+  --workload=$workload \
+  --framework=$framework \
+  --tuning_ruleset=external \
+  --data_dir=$DATA_DIR/$dataset \
+  --imagenet_v2_data_dir=$DATA_DIR/$dataset \
+  --librispeech_tokenizer_vocab_path=$tokenizer_path \
+  --submission_path=$submission \
+  --tuning_search_space=$search_space \
+  --num_tuning_trials=$num_tuning_trials \
+  --hparam_start_index=0 \
+  --hparam_end_index=1 \
+  --max_pct_of_global_steps=0.1 \
+  --experiment_dir=$EXP_DIR  \
+  --experiment_name=$experiment_name \
+  --save_intermediate_checkpoints=False \
+  --save_checkpoints=False \
+  --use_wandb \
+  --rng_seed=$rng_seed \
+  --torch_compile=True \
+  --allow_tf32=False \
+  --halve_CUDA_mem=False \
+  --pytorch_eval_num_workers=$eval_num_workers \
+  --overwrite

script/jax/v100/auto_run_array.sh

@@ -0,0 +1,100 @@
+#!/bin/bash
+
+module load cuda/12.1
+
+source ~/miniforge3/etc/profile.d/conda.sh
+conda activate alpe_jax
+
+# Env vars
+export OMP_NUM_THREADS=2 # TODO: check
+export HOME=/home/najroldi
+export CODE_DIR=/home/najroldi/algorithmic-efficiency
+export EXP_DIR=/fast/najroldi/exp/algoperf
+export DATA_DIR=/fast/najroldi/data
+
+# Job specific vars
+workload_or_id=$1
+framework=$2
+submission=$3
+search_space=$4
+name=$5
+study=$6
+num_tuning_trials=$7
+rng_seed=$8
+
+workload_list=(
+  criteo1tb
+  fastmri
+  imagenet_resnet
+  imagenet_vit
+  librispeech_conformer
+  librispeech_deepspeech
+  ogbg
+  wmt
+)
+
+# in bash need to explicitly declare an associative array
+declare -A workload_to_dataset=(
+  [criteo1tb]="criteo1tb"
+  [fastmri]="fastmri"
+  [imagenet_resnet]="imagenet"
+  [imagenet_vit]="imagenet"
+  [librispeech_conformer]="librispeech"
+  [librispeech_deepspeech]="librispeech"
+  [ogbg]="ogbg"
+  [wmt]="wmt"
+)
+
+# Determine workload based on workload_or_id
+if [[ "$workload_or_id" =~ ^[0-7]+$ ]]; then
+  # workload_or_id is a valid index (0-7), treat it as job ID and get workload from list
+  job_id=$workload_or_id
+  workload=${workload_list[$job_id]}
+else
+  # treat workload_or_id as an explicit workload name
+  workload=$workload_or_id
+fi
+
+dataset=${workload_to_dataset[$workload]}
+
+# Experiment name
+# experiment_name="${name}_${workload}_${framework}"
+experiment_name=${name}
+
+# Librispeech tokenizer path
+tokenizer_path=''
+if [ "$dataset" = "librispeech" ]; then
+    tokenizer_path="${DATA_DIR}/librispeech/spm_model.vocab"
+fi
+
+# Increase num_workers on imagenet
+eval_num_workers=0
+if [ "$dataset" == "imagenet" ] && [ "$framework" == "pytorch" ]; then
+  eval_num_workers=4
+fi
+
+# Execute python script
+python \
+  $CODE_DIR/submission_runner.py \
+  --workload=$workload \
+  --framework=$framework \
+  --tuning_ruleset=external \
+  --data_dir=$DATA_DIR/$dataset \
+  --imagenet_v2_data_dir=$DATA_DIR/$dataset \
+  --librispeech_tokenizer_vocab_path=$tokenizer_path \
+  --submission_path=$submission \
+  --tuning_search_space=$search_space \
+  --num_tuning_trials=$num_tuning_trials \
+  --hparam_start_index=0 \
+  --hparam_end_index=1 \
+  --max_pct_of_global_steps=0.1 \
+  --experiment_dir=$EXP_DIR  \
+  --experiment_name=$experiment_name \
+  --save_intermediate_checkpoints=False \
+  --save_checkpoints=False \
+  --use_wandb \
+  --rng_seed=$rng_seed \
+  --allow_tf32=False \
+  --halve_CUDA_mem=False \
+  --pytorch_eval_num_workers=$eval_num_workers \
+  --overwrite

script/pytorch/2xa100/auto_run_array.sh

@@ -0,0 +1,117 @@
+#!/bin/bash
+
+source ~/miniforge3/etc/profile.d/conda.sh
+conda activate alpe
+
+# Env vars
+export OMP_NUM_THREADS=2 # TODO: check
+export HOME=/home/najroldi
+export CODE_DIR=/home/najroldi/algorithmic-efficiency
+export EXP_DIR=/fast/najroldi/exp/algoperf
+export DATA_DIR=/fast/najroldi/data
+
+# Job specific vars
+workload_or_id=$1
+framework=$2
+submission=$3
+search_space=$4
+name=$5
+study=$6
+num_tuning_trials=$7
+rng_seed=$8
+allow_tf_32=$9
+halve_cuda_mem=$10
+
+workload_list=(
+  criteo1tb
+  fastmri
+  imagenet_resnet
+  imagenet_vit
+  librispeech_conformer
+  librispeech_deepspeech
+  ogbg
+  wmt
+)
+
+# in bash need to explicitly declare an associative array
+declare -A workload_to_dataset=(
+  [criteo1tb]="criteo1tb"
+  [fastmri]="fastmri"
+  [imagenet_resnet]="imagenet"
+  [imagenet_vit]="imagenet"
+  [librispeech_conformer]="librispeech"
+  [librispeech_deepspeech]="librispeech"
+  [ogbg]="ogbg"
+  [wmt]="wmt"
+)
+
+# Determine workload based on workload_or_id
+if [[ "$workload_or_id" =~ ^[0-7]+$ ]]; then
+  # workload_or_id is a valid index (0-7), treat it as job ID and get workload from list
+  job_id=$workload_or_id
+  workload=${workload_list[$job_id]}
+else
+  # treat workload_or_id as an explicit workload name
+  workload=$workload_or_id
+fi
+
+dataset=${workload_to_dataset[$workload]}
+
+# Experiment name
+# experiment_name="${name}_${workload}_${framework}"
+experiment_name=${name}
+
+# Librispeech tokenizer path
+tokenizer_path=''
+if [ "$dataset" = "librispeech" ]; then
+    tokenizer_path="${DATA_DIR}/librispeech/spm_model.vocab"
+fi
+
+# Increase num_workers on imagenet
+eval_num_workers=0
+if [ "$dataset" == "imagenet" ] && [ "$framework" == "pytorch" ]; then
+  eval_num_workers=4
+fi
+
+# allow_tf_32
+allow_tf_32_flag=False
+if [ "$allow_tf_32" == "1" ]; then
+  allow_tf_32_flag=True
+fi
+
+# allow_tf_32
+halve_cuda_mem_flag=False
+if [ "$halve_cuda_mem" == "1" ]; then
+  halve_cuda_mem_flag=True
+fi
+
+# Execute python script
+torchrun \
+  --redirects 1:0 \
+  --standalone \
+  --nnodes=1 \
+  --nproc_per_node=2 \
+  $CODE_DIR/submission_runner.py \
+  --workload=$workload \
+  --framework=$framework \
+  --tuning_ruleset=external \
+  --data_dir=$DATA_DIR/$dataset \
+  --imagenet_v2_data_dir=$DATA_DIR/$dataset \
+  --librispeech_tokenizer_vocab_path=$tokenizer_path \
+  --submission_path=$submission \
+  --tuning_search_space=$search_space \
+  --num_tuning_trials=$num_tuning_trials \
+  --hparam_start_index=0 \
+  --hparam_end_index=1 \
+  --max_pct_of_global_steps=0.1 \
+  --experiment_dir=$EXP_DIR  \
+  --experiment_name=$experiment_name \
+  --save_intermediate_checkpoints=False \
+  --save_checkpoints=False \
+  --use_wandb \
+  --rng_seed=$rng_seed \
+  --torch_compile=True \
+  --allow_tf32=$allow_tf_32_flag \
+  --halve_CUDA_mem=$halve_cuda_mem_flag \
+  --pytorch_eval_num_workers=$eval_num_workers \
+  --overwrite