Upload llama-1B/8_GPUS/dp-1_tp-2_pp-4_mbz-2
Browse files
llama-1B/8_GPUS/dp-1_tp-2_pp-4_mbz-2/bench.slurm
ADDED
|
@@ -0,0 +1,111 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
|
| 3 |
+
#SBATCH --job-name=bench_cluster
|
| 4 |
+
#SBATCH --time=02:00:00
|
| 5 |
+
#SBATCH --partition=hopper-prod
|
| 6 |
+
#SBATCH --nodes=1
|
| 7 |
+
#SBATCH --gres=gpu:8
|
| 8 |
+
#SBATCH --qos=normal
|
| 9 |
+
#SBATCH --ntasks-per-node=1
|
| 10 |
+
#SBATCH --cpus-per-task=96
|
| 11 |
+
#SBATCH --exclusive
|
| 12 |
+
#SBATCH --output=/fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-1_tp-2_pp-4_mbz-2/log.out
|
| 13 |
+
#SBATCH --error=/fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-1_tp-2_pp-4_mbz-2/log.out
|
| 14 |
+
|
| 15 |
+
# Function to update status based on squeue output
|
| 16 |
+
update_status() {
|
| 17 |
+
job_id=$1
|
| 18 |
+
status_file=$2
|
| 19 |
+
# For unknown reasons, it doenst update status for pending. It only works for running
|
| 20 |
+
while true; do
|
| 21 |
+
job_status=$(squeue --job $job_id --noheader --format=%T)
|
| 22 |
+
echo "Job status: $job_status"
|
| 23 |
+
if [ -z "$job_status" ]; then
|
| 24 |
+
# Job has finished or is not found
|
| 25 |
+
break
|
| 26 |
+
elif [ "$job_status" = "RUNNING" ]; then
|
| 27 |
+
printf "running" > $status_file
|
| 28 |
+
break
|
| 29 |
+
fi
|
| 30 |
+
sleep 10
|
| 31 |
+
done
|
| 32 |
+
}
|
| 33 |
+
|
| 34 |
+
# Misc initializations.
|
| 35 |
+
echo "========================"
|
| 36 |
+
echo "START TIME: $(date)"
|
| 37 |
+
source /fsx/ferdinandmom/miniforge3/etc/profile.d/conda.sh
|
| 38 |
+
conda activate /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster
|
| 39 |
+
echo python3 version = $(python3 --version)
|
| 40 |
+
echo "========================"
|
| 41 |
+
|
| 42 |
+
# Slurm stuff
|
| 43 |
+
export HOSTNAMES=$(scontrol show hostnames "$SLURM_JOB_NODELIST")
|
| 44 |
+
export MASTER_ADDR=$(scontrol show hostnames "$SLURM_JOB_NODELIST" | head -n 1)
|
| 45 |
+
export MASTER_PORT=$((1024 + RANDOM % 64511))
|
| 46 |
+
|
| 47 |
+
export TMPDIR=/scratch
|
| 48 |
+
export HF_DATASETS_CACHE="/admin/home/ferdinand_mom/.cache"
|
| 49 |
+
export CUBLAS_WORKSPACE_CONFIG=":4096:8"
|
| 50 |
+
export CUDA_DEVICE_MAX_CONNECTIONS="1"
|
| 51 |
+
|
| 52 |
+
huggingface-cli login --token $HUGGINGFACE_TOKEN
|
| 53 |
+
|
| 54 |
+
|
| 55 |
+
NANOTRON_REPO="/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron"
|
| 56 |
+
CMD="$NANOTRON_REPO/run_train.py --config-file /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-1_tp-2_pp-4_mbz-2/config.yaml"
|
| 57 |
+
|
| 58 |
+
LAUNCHER="torchrun \
|
| 59 |
+
--nproc_per_node 8 \
|
| 60 |
+
--nnodes 1 \
|
| 61 |
+
--rdzv_endpoint ${MASTER_ADDR}:${MASTER_PORT} \
|
| 62 |
+
--rdzv_backend c10d \
|
| 63 |
+
--max_restarts 0 \
|
| 64 |
+
--tee 3 \
|
| 65 |
+
--node_rank ${SLURM_PROCID}"
|
| 66 |
+
|
| 67 |
+
# Checkout the bench_cluster branch
|
| 68 |
+
cd $NANOTRON_REPO
|
| 69 |
+
git checkout bench_cluster
|
| 70 |
+
cd ..
|
| 71 |
+
# Get the current job ID
|
| 72 |
+
job_id=${SLURM_JOB_ID}
|
| 73 |
+
|
| 74 |
+
# Update status to "pending" or "running" in the background
|
| 75 |
+
update_status $job_id /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-1_tp-2_pp-4_mbz-2/status.txt &
|
| 76 |
+
|
| 77 |
+
# Run the main command
|
| 78 |
+
srun -u $LAUNCHER $CMD
|
| 79 |
+
exit_status=$?
|
| 80 |
+
|
| 81 |
+
# Update status based on the exit status of `srun`
|
| 82 |
+
if [ $exit_status -eq 0 ]; then
|
| 83 |
+
printf "completed" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-1_tp-2_pp-4_mbz-2/status.txt
|
| 84 |
+
else
|
| 85 |
+
if grep -q "OutOfMemoryError" /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-1_tp-2_pp-4_mbz-2/log.out; then
|
| 86 |
+
printf "oom" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-1_tp-2_pp-4_mbz-2/status.txt
|
| 87 |
+
elif grep -q " CUDA error: an illegal memory access" /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-1_tp-2_pp-4_mbz-2/log.out; then
|
| 88 |
+
printf "oom" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-1_tp-2_pp-4_mbz-2/status.txt
|
| 89 |
+
elif grep -q "Timeout at NCCL" /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-1_tp-2_pp-4_mbz-2/log.out; then
|
| 90 |
+
printf "timeout" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-1_tp-2_pp-4_mbz-2/status.txt
|
| 91 |
+
else
|
| 92 |
+
printf "fail" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-1_tp-2_pp-4_mbz-2/status.txt
|
| 93 |
+
fi
|
| 94 |
+
fi
|
| 95 |
+
|
| 96 |
+
# Run the report script if the job completed successfully
|
| 97 |
+
if [ $exit_status -eq 0 ]; then
|
| 98 |
+
python /fsx/ferdinandmom/ferdinand-hf/bench_cluster/main.py report --inp_dir /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-1_tp-2_pp-4_mbz-2 --is_logs
|
| 99 |
+
python /fsx/ferdinandmom/ferdinand-hf/bench_cluster/main.py report --inp_dir /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-1_tp-2_pp-4_mbz-2 --is_profiler
|
| 100 |
+
fi
|
| 101 |
+
|
| 102 |
+
|
| 103 |
+
# Push to hub the folder using huggingface_cli
|
| 104 |
+
huggingface-cli upload nanotron/bench_cluster /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-1_tp-2_pp-4_mbz-2 llama-1B/8_GPUS/dp-1_tp-2_pp-4_mbz-2 --commit-message "Upload llama-1B/8_GPUS/dp-1_tp-2_pp-4_mbz-2"
|
| 105 |
+
|
| 106 |
+
# Verify the upload
|
| 107 |
+
if [ $? -eq 0 ]; then
|
| 108 |
+
echo "Uploading to Huggingface Hub successful"
|
| 109 |
+
else
|
| 110 |
+
echo "Failed to upload to Huggingface Hub"
|
| 111 |
+
fi
|
llama-1B/8_GPUS/dp-1_tp-2_pp-4_mbz-2/config.yaml
ADDED
|
@@ -0,0 +1,90 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
general:
|
| 2 |
+
project: bench_cluster
|
| 3 |
+
seed: 42
|
| 4 |
+
model:
|
| 5 |
+
ddp_bucket_cap_mb: 25
|
| 6 |
+
dtype: bfloat16
|
| 7 |
+
init_method:
|
| 8 |
+
std: 0.025
|
| 9 |
+
make_vocab_size_divisible_by: 1
|
| 10 |
+
model_config:
|
| 11 |
+
bos_token_id: 1
|
| 12 |
+
eos_token_id: 2
|
| 13 |
+
hidden_act: silu
|
| 14 |
+
hidden_size: 2048
|
| 15 |
+
initializer_range: 0.02
|
| 16 |
+
intermediate_size: 4096
|
| 17 |
+
is_llama_config: true
|
| 18 |
+
max_position_embeddings: 4096
|
| 19 |
+
num_attention_heads: 32
|
| 20 |
+
num_hidden_layers: 24
|
| 21 |
+
num_key_value_heads: 32
|
| 22 |
+
pad_token_id: null
|
| 23 |
+
pretraining_tp: 1
|
| 24 |
+
rms_norm_eps: 1.0e-05
|
| 25 |
+
rope_scaling: null
|
| 26 |
+
rope_theta: 10000.0
|
| 27 |
+
tie_word_embeddings: true
|
| 28 |
+
use_cache: true
|
| 29 |
+
vocab_size: 50257
|
| 30 |
+
optimizer:
|
| 31 |
+
accumulate_grad_in_fp32: true
|
| 32 |
+
clip_grad: 1.0
|
| 33 |
+
learning_rate_scheduler:
|
| 34 |
+
learning_rate: 0.0001
|
| 35 |
+
lr_decay_style: linear
|
| 36 |
+
lr_warmup_style: linear
|
| 37 |
+
lr_warmup_steps: 1
|
| 38 |
+
min_decay_lr: 1.0e-05
|
| 39 |
+
optimizer_factory:
|
| 40 |
+
adam_beta1: 0.9
|
| 41 |
+
adam_beta2: 0.95
|
| 42 |
+
adam_eps: 1.0e-08
|
| 43 |
+
name: adamW
|
| 44 |
+
torch_adam_is_fused: true
|
| 45 |
+
weight_decay: 0.01
|
| 46 |
+
zero_stage: 1
|
| 47 |
+
parallelism:
|
| 48 |
+
dp: 1
|
| 49 |
+
expert_parallel_size: 1
|
| 50 |
+
pp: 4
|
| 51 |
+
pp_engine: 1f1b
|
| 52 |
+
tp: 2
|
| 53 |
+
tp_linear_async_communication: false
|
| 54 |
+
tp_mode: REDUCE_SCATTER
|
| 55 |
+
profiler:
|
| 56 |
+
profiler_export_path: /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-1_tp-2_pp-4_mbz-2
|
| 57 |
+
tokenizer:
|
| 58 |
+
tokenizer_max_length: null
|
| 59 |
+
tokenizer_name_or_path: openai-community/gpt2
|
| 60 |
+
tokenizer_revision: null
|
| 61 |
+
data_stages:
|
| 62 |
+
- name: Training Stage
|
| 63 |
+
start_training_step: 1
|
| 64 |
+
data:
|
| 65 |
+
dataset:
|
| 66 |
+
dataset_overwrite_cache: false
|
| 67 |
+
dataset_processing_num_proc_per_process: 64
|
| 68 |
+
hf_dataset_config_name: null
|
| 69 |
+
hf_dataset_or_datasets: roneneldan/TinyStories
|
| 70 |
+
hf_dataset_splits: train
|
| 71 |
+
text_column_name: text
|
| 72 |
+
num_loading_workers: 0
|
| 73 |
+
seed: 42
|
| 74 |
+
lighteval: null
|
| 75 |
+
tokens:
|
| 76 |
+
train_steps: 20
|
| 77 |
+
val_check_interval: -1
|
| 78 |
+
batch_accumulation_per_replica: 512
|
| 79 |
+
limit_test_batches: 0
|
| 80 |
+
limit_val_batches: 0
|
| 81 |
+
micro_batch_size: 2
|
| 82 |
+
sequence_length: 4096
|
| 83 |
+
logging:
|
| 84 |
+
iteration_step_info_interval: 1
|
| 85 |
+
log_level: info
|
| 86 |
+
log_level_replica: info
|
| 87 |
+
checkpoints:
|
| 88 |
+
checkpoint_interval: 100000
|
| 89 |
+
checkpoints_path: /dev/null
|
| 90 |
+
resume_checkpoint_path: null
|
llama-1B/8_GPUS/dp-1_tp-2_pp-4_mbz-2/log.out
ADDED
|
@@ -0,0 +1,608 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
========================
|
| 2 |
+
START TIME: Wed Jul 3 21:08:06 UTC 2024
|
| 3 |
+
python3 version = Python 3.10.14
|
| 4 |
+
========================
|
| 5 |
+
The token has not been saved to the git credentials helper. Pass `add_to_git_credential=True` in this function directly or `--add-to-git-credential` if using via `huggingface-cli` if you want to set the git credential as well.
|
| 6 |
+
Token is valid (permission: write).
|
| 7 |
+
Your token has been saved to /admin/home/ferdinand_mom/.cache/huggingface/token
|
| 8 |
+
Login successful
|
| 9 |
+
Already on 'bench_cluster'
|
| 10 |
+
M examples/config_tiny_llama.py
|
| 11 |
+
M examples/config_tiny_llama.yaml
|
| 12 |
+
M examples/train_tiny_llama.sh
|
| 13 |
+
M src/nanotron/models/llama.py
|
| 14 |
+
M src/nanotron/trainer.py
|
| 15 |
+
Your branch is up to date with 'origin/bench_cluster'.
|
| 16 |
+
Job status: RUNNING
|
| 17 |
+
W0703 21:08:08.929000 139794302568256 torch/distributed/run.py:757]
|
| 18 |
+
W0703 21:08:08.929000 139794302568256 torch/distributed/run.py:757] *****************************************
|
| 19 |
+
W0703 21:08:08.929000 139794302568256 torch/distributed/run.py:757] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed.
|
| 20 |
+
W0703 21:08:08.929000 139794302568256 torch/distributed/run.py:757] *****************************************
|
| 21 |
+
[default0]:07/03/2024 21:08:25 [WARNING|DP=0|PP=0|TP=0|ip-26-0-162-233]: [Vocab Size Padding] Padded vocab (size: 50257) with 1 dummy tokens (new size: 50258)
|
| 22 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: Config:
|
| 23 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: Config(general=GeneralArgs(project='bench_cluster',
|
| 24 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: run='%date_%jobid',
|
| 25 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: seed=42,
|
| 26 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: step=None,
|
| 27 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: consumed_train_samples=None,
|
| 28 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: benchmark_csv_path=None,
|
| 29 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: ignore_sanity_checks=True),
|
| 30 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: parallelism=ParallelismArgs(dp=1,
|
| 31 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: pp=4,
|
| 32 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: tp=2,
|
| 33 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: pp_engine=<nanotron.parallel.pipeline_parallel.engine.OneForwardOneBackwardPipelineEngine object at 0x7fbd74a78670>,
|
| 34 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: tp_mode=<TensorParallelLinearMode.REDUCE_SCATTER: 2>,
|
| 35 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: tp_linear_async_communication=False,
|
| 36 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: expert_parallel_size=1),
|
| 37 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: model=ModelArgs(model_config=LlamaConfig(bos_token_id=1,
|
| 38 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: eos_token_id=2,
|
| 39 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: hidden_act='silu',
|
| 40 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: hidden_size=2048,
|
| 41 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: initializer_range=0.02,
|
| 42 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: intermediate_size=4096,
|
| 43 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: is_llama_config=True,
|
| 44 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: max_position_embeddings=4096,
|
| 45 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: num_attention_heads=32,
|
| 46 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: num_hidden_layers=24,
|
| 47 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: num_key_value_heads=32,
|
| 48 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: pad_token_id=None,
|
| 49 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: pretraining_tp=1,
|
| 50 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: rms_norm_eps=1e-05,
|
| 51 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: rope_scaling=None,
|
| 52 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: rope_theta=10000.0,
|
| 53 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: tie_word_embeddings=True,
|
| 54 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: use_cache=True,
|
| 55 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: vocab_size=50258),
|
| 56 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: init_method=RandomInit(std=0.025),
|
| 57 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: dtype=torch.bfloat16,
|
| 58 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: make_vocab_size_divisible_by=1,
|
| 59 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: ddp_bucket_cap_mb=25),
|
| 60 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: tokenizer=TokenizerArgs(tokenizer_name_or_path='openai-community/gpt2',
|
| 61 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: tokenizer_revision=None,
|
| 62 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: tokenizer_max_length=None),
|
| 63 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: checkpoints=CheckpointsArgs(checkpoints_path=Path('/dev/null'),
|
| 64 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: checkpoint_interval=100000,
|
| 65 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: save_initial_state=False,
|
| 66 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: resume_checkpoint_path=None,
|
| 67 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: checkpoints_path_is_shared_file_system=False),
|
| 68 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: logging=LoggingArgs(log_level='info',
|
| 69 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: log_level_replica='info',
|
| 70 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: iteration_step_info_interval=1),
|
| 71 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: tokens=TokensArgs(sequence_length=4096,
|
| 72 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: train_steps=20,
|
| 73 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: micro_batch_size=2,
|
| 74 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: batch_accumulation_per_replica=512,
|
| 75 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: val_check_interval=-1,
|
| 76 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: limit_val_batches=0,
|
| 77 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: limit_test_batches=0),
|
| 78 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: optimizer=OptimizerArgs(optimizer_factory=AdamWOptimizerArgs(adam_eps=1e-08,
|
| 79 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: adam_beta1=0.9,
|
| 80 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: adam_beta2=0.95,
|
| 81 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: torch_adam_is_fused=True,
|
| 82 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: name='adamW'),
|
| 83 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: zero_stage=1,
|
| 84 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: weight_decay=0.01,
|
| 85 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: clip_grad=1.0,
|
| 86 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: accumulate_grad_in_fp32=True,
|
| 87 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: learning_rate_scheduler=LRSchedulerArgs(learning_rate=0.0001,
|
| 88 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: lr_warmup_steps=1,
|
| 89 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: lr_warmup_style='linear',
|
| 90 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: lr_decay_style='linear',
|
| 91 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: lr_decay_steps=19,
|
| 92 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: lr_decay_starting_step=None,
|
| 93 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: min_decay_lr=1e-05)),
|
| 94 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: data_stages=[DatasetStageArgs(name='Training Stage',
|
| 95 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: start_training_step=1,
|
| 96 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: data=DataArgs(dataset=PretrainDatasetsArgs(hf_dataset_or_datasets='roneneldan/TinyStories',
|
| 97 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: hf_dataset_splits='train',
|
| 98 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: hf_dataset_config_name=None,
|
| 99 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: dataset_processing_num_proc_per_process=64,
|
| 100 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: dataset_overwrite_cache=False,
|
| 101 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: text_column_name='text'),
|
| 102 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: seed=42,
|
| 103 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: num_loading_workers=0))],
|
| 104 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: profiler=ProfilerArgs(profiler_export_path=Path('/fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-1_tp-2_pp-4_mbz-2')),
|
| 105 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: lighteval=None)
|
| 106 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: Model Config:
|
| 107 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: LlamaConfig(bos_token_id=1,
|
| 108 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: eos_token_id=2,
|
| 109 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: hidden_act='silu',
|
| 110 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: hidden_size=2048,
|
| 111 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: initializer_range=0.02,
|
| 112 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: intermediate_size=4096,
|
| 113 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: is_llama_config=True,
|
| 114 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: max_position_embeddings=4096,
|
| 115 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: num_attention_heads=32,
|
| 116 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: num_hidden_layers=24,
|
| 117 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: num_key_value_heads=32,
|
| 118 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: pad_token_id=None,
|
| 119 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: pretraining_tp=1,
|
| 120 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: rms_norm_eps=1e-05,
|
| 121 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: rope_scaling=None,
|
| 122 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: rope_theta=10000.0,
|
| 123 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: tie_word_embeddings=True,
|
| 124 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: use_cache=True,
|
| 125 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: vocab_size=50258)
|
| 126 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: Building model..
|
| 127 |
+
[default0]:07/03/2024 21:08:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: Setting PP block ranks...
|
| 128 |
+
[default6]:07/03/2024 21:08:39 [INFO|DP=0|PP=3|TP=0|ip-26-0-162-233]: Local number of parameters: 135M (258.20MiB)
|
| 129 |
+
[default6]:07/03/2024 21:08:39 [INFO|DP=0|PP=3|TP=0|ip-26-0-162-233]: [After model building] Memory usage: 262.21MiB. Peak allocated: 264.24MiB Peak reserved: 280.00MiB
|
| 130 |
+
[default6]:07/03/2024 21:08:39 [INFO|DP=0|PP=3|TP=0|ip-26-0-162-233]: No checkpoint path provided.
|
| 131 |
+
[default1]:07/03/2024 21:08:39 [INFO|DP=0|PP=0|TP=1|ip-26-0-162-233]: Local number of parameters: 198M (378.21MiB)
|
| 132 |
+
[default1]:07/03/2024 21:08:39 [INFO|DP=0|PP=0|TP=1|ip-26-0-162-233]: [After model building] Memory usage: 385.23MiB. Peak allocated: 387.26MiB Peak reserved: 402.00MiB
|
| 133 |
+
[default1]:07/03/2024 21:08:39 [INFO|DP=0|PP=0|TP=1|ip-26-0-162-233]: No checkpoint path provided.
|
| 134 |
+
[default2]:07/03/2024 21:08:39 [INFO|DP=0|PP=1|TP=0|ip-26-0-162-233]: Local number of parameters: 147M (280.05MiB)
|
| 135 |
+
[default2]:07/03/2024 21:08:39 [INFO|DP=0|PP=1|TP=0|ip-26-0-162-233]: [After model building] Memory usage: 287.07MiB. Peak allocated: 289.10MiB Peak reserved: 302.00MiB
|
| 136 |
+
[default2]:07/03/2024 21:08:39 [INFO|DP=0|PP=1|TP=0|ip-26-0-162-233]: No checkpoint path provided.
|
| 137 |
+
[default5]:07/03/2024 21:08:39 [INFO|DP=0|PP=2|TP=1|ip-26-0-162-233]: Local number of parameters: 126M (240.05MiB)
|
| 138 |
+
[default5]:07/03/2024 21:08:39 [INFO|DP=0|PP=2|TP=1|ip-26-0-162-233]: [After model building] Memory usage: 246.06MiB. Peak allocated: 248.09MiB Peak reserved: 262.00MiB
|
| 139 |
+
[default7]:07/03/2024 21:08:39 [INFO|DP=0|PP=3|TP=1|ip-26-0-162-233]: Local number of parameters: 135M (258.20MiB)
|
| 140 |
+
[default7]:07/03/2024 21:08:39 [INFO|DP=0|PP=3|TP=1|ip-26-0-162-233]: [After model building] Memory usage: 262.21MiB. Peak allocated: 264.24MiB Peak reserved: 280.00MiB
|
| 141 |
+
[default0]:07/03/2024 21:08:39 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: Total number of parameters: 1.21G (2313.02MiB)
|
| 142 |
+
[default0]:07/03/2024 21:08:39 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: Local number of parameters: 198M (378.21MiB)
|
| 143 |
+
[default0]:07/03/2024 21:08:39 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: [After model building] Memory usage: 385.23MiB. Peak allocated: 387.26MiB Peak reserved: 402.00MiB
|
| 144 |
+
[default0]:07/03/2024 21:08:39 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: No checkpoint path provided.
|
| 145 |
+
[default0]:07/03/2024 21:08:39 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: Parametrizing model parameters using StandardParametrizator
|
| 146 |
+
[default5]:07/03/2024 21:08:39 [INFO|DP=0|PP=2|TP=1|ip-26-0-162-233]: No checkpoint path provided.
|
| 147 |
+
[default7]:07/03/2024 21:08:39 [INFO|DP=0|PP=3|TP=1|ip-26-0-162-233]: No checkpoint path provided.
|
| 148 |
+
[default4]:07/03/2024 21:08:39 [INFO|DP=0|PP=2|TP=0|ip-26-0-162-233]: Local number of parameters: 126M (240.05MiB)
|
| 149 |
+
[default3]:07/03/2024 21:08:39 [INFO|DP=0|PP=1|TP=1|ip-26-0-162-233]: Local number of parameters: 147M (280.05MiB)
|
| 150 |
+
[default3]:07/03/2024 21:08:39 [INFO|DP=0|PP=1|TP=1|ip-26-0-162-233]: [After model building] Memory usage: 287.07MiB. Peak allocated: 289.10MiB Peak reserved: 302.00MiB
|
| 151 |
+
[default4]:07/03/2024 21:08:39 [INFO|DP=0|PP=2|TP=0|ip-26-0-162-233]: [After model building] Memory usage: 246.06MiB. Peak allocated: 248.09MiB Peak reserved: 262.00MiB
|
| 152 |
+
[default4]:07/03/2024 21:08:39 [INFO|DP=0|PP=2|TP=0|ip-26-0-162-233]: No checkpoint path provided.
|
| 153 |
+
[default3]:07/03/2024 21:08:39 [INFO|DP=0|PP=1|TP=1|ip-26-0-162-233]: No checkpoint path provided.
|
| 154 |
+
[default0]:07/03/2024 21:08:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: [Optimizer Building] Using LearningRateForSP as learning rate
|
| 155 |
+
[default0]:07/03/2024 21:08:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: [ZeRO sharding] Size of optimizer params per rank:
|
| 156 |
+
[default0]:07/03/2024 21:08:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: [ZeRO sharding] DP Rank 0 has 198M out of 198M (100.00%) params' optimizer states
|
| 157 |
+
[default0]:07/03/2024 21:08:42 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: [Training Plan] Stage Training Stage has 19 remaining training steps and has consumed 0 samples
|
| 158 |
+
[default0]:07/03/2024 21:08:42 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: Using `datasets` library
|
| 159 |
+
[default0]:07/03/2024 21:08:42 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: Loading tokenizer from openai-community/gpt2 and transformers/hf_hub versions ('4.41.2', '0.23.4')
|
| 160 |
+
[default0]:Repo card metadata block was not found. Setting CardData to empty.
|
| 161 |
+
[default0]:07/03/2024 21:08:42 [WARNING|DP=0|PP=0|TP=0|ip-26-0-162-233]: Repo card metadata block was not found. Setting CardData to empty.
|
| 162 |
+
[default0]:07/03/2024 21:08:43 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: [Training Plan] There are 1 training stages
|
| 163 |
+
[default0]:07/03/2024 21:08:43 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: [Stage Training Stage] start from step 1
|
| 164 |
+
[default0]:07/03/2024 21:08:43 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]:
|
| 165 |
+
[default0]:07/03/2024 21:08:43 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: [Start training] datetime: 2024-07-03 21:08:43.202168 | mbs: 2 | grad_accum: 512 | global_batch_size: 1024 | sequence_length: 4096 | train_steps: 20 | start_iteration_step: 0 | consumed_train_samples: 0
|
| 166 |
+
[default0]:07/03/2024 21:08:43 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: Resuming training from stage Training Stage, it has trained for 0 samples and has 19 remaining train steps
|
| 167 |
+
[default0]:07/03/2024 21:08:43 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: Memory usage: 1898.09MiB. Peak allocated 1898.09MiB. Peak reserved: 1918.00MiB
|
| 168 |
+
[default2]:07/03/2024 21:08:43 [WARNING|DP=0|PP=1|TP=0|ip-26-0-162-233]: Repo card metadata block was not found. Setting CardData to empty.
|
| 169 |
+
[default5]:07/03/2024 21:08:43 [WARNING|DP=0|PP=2|TP=1|ip-26-0-162-233]: Repo card metadata block was not found. Setting CardData to empty.
|
| 170 |
+
[default7]:07/03/2024 21:08:43 [WARNING|DP=0|PP=3|TP=1|ip-26-0-162-233]: Repo card metadata block was not found. Setting CardData to empty.
|
| 171 |
+
[default4]:07/03/2024 21:08:43 [WARNING|DP=0|PP=2|TP=0|ip-26-0-162-233]: Repo card metadata block was not found. Setting CardData to empty.
|
| 172 |
+
[default4]:Repo card metadata block was not found. Setting CardData to empty.
|
| 173 |
+
[default3]:07/03/2024 21:08:43 [WARNING|DP=0|PP=1|TP=1|ip-26-0-162-233]: Repo card metadata block was not found. Setting CardData to empty.
|
| 174 |
+
[default2]:Repo card metadata block was not found. Setting CardData to empty.
|
| 175 |
+
[default5]:Repo card metadata block was not found. Setting CardData to empty.
|
| 176 |
+
[default3]:Repo card metadata block was not found. Setting CardData to empty.
|
| 177 |
+
[default7]:Repo card metadata block was not found. Setting CardData to empty.
|
| 178 |
+
[default6]:07/03/2024 21:08:43 [WARNING|DP=0|PP=3|TP=0|ip-26-0-162-233]: Repo card metadata block was not found. Setting CardData to empty.
|
| 179 |
+
[default1]:Repo card metadata block was not found. Setting CardData to empty.
|
| 180 |
+
[default1]:07/03/2024 21:08:43 [WARNING|DP=0|PP=0|TP=1|ip-26-0-162-233]: Repo card metadata block was not found. Setting CardData to empty.
|
| 181 |
+
[default6]:Repo card metadata block was not found. Setting CardData to empty.
|
| 182 |
+
[default6]:/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py:744: UserWarning: c10d::allreduce_: an autograd kernel was not registered to the Autograd key(s) but we are trying to backprop through it. This may lead to silently incorrect behavior. This behavior is deprecated and will be removed in a future version of PyTorch. If your operator is differentiable, please ensure you have registered an autograd kernel to the correct Autograd key (e.g. DispatchKey::Autograd, DispatchKey::CompositeImplicitAutograd). If your operator is not differentiable, or to squash this warning and use the previous behavior, please register torch::CppFunction::makeFallthrough() to DispatchKey::Autograd. (Triggered internally at ../torch/csrc/autograd/autograd_not_implemented_fallback.cpp:63.)
|
| 183 |
+
[default6]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
|
| 184 |
+
[default7]:/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py:744: UserWarning: c10d::allreduce_: an autograd kernel was not registered to the Autograd key(s) but we are trying to backprop through it. This may lead to silently incorrect behavior. This behavior is deprecated and will be removed in a future version of PyTorch. If your operator is differentiable, please ensure you have registered an autograd kernel to the correct Autograd key (e.g. DispatchKey::Autograd, DispatchKey::CompositeImplicitAutograd). If your operator is not differentiable, or to squash this warning and use the previous behavior, please register torch::CppFunction::makeFallthrough() to DispatchKey::Autograd. (Triggered internally at ../torch/csrc/autograd/autograd_not_implemented_fallback.cpp:63.)
|
| 185 |
+
[default7]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
|
| 186 |
+
[default5]:/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py:744: UserWarning: c10d::allreduce_: an autograd kernel was not registered to the Autograd key(s) but we are trying to backprop through it. This may lead to silently incorrect behavior. This behavior is deprecated and will be removed in a future version of PyTorch. If your operator is differentiable, please ensure you have registered an autograd kernel to the correct Autograd key (e.g. DispatchKey::Autograd, DispatchKey::CompositeImplicitAutograd). If your operator is not differentiable, or to squash this warning and use the previous behavior, please register torch::CppFunction::makeFallthrough() to DispatchKey::Autograd. (Triggered internally at ../torch/csrc/autograd/autograd_not_implemented_fallback.cpp:63.)
|
| 187 |
+
[default5]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
|
| 188 |
+
[default4]:/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py:744: UserWarning: c10d::allreduce_: an autograd kernel was not registered to the Autograd key(s) but we are trying to backprop through it. This may lead to silently incorrect behavior. This behavior is deprecated and will be removed in a future version of PyTorch. If your operator is differentiable, please ensure you have registered an autograd kernel to the correct Autograd key (e.g. DispatchKey::Autograd, DispatchKey::CompositeImplicitAutograd). If your operator is not differentiable, or to squash this warning and use the previous behavior, please register torch::CppFunction::makeFallthrough() to DispatchKey::Autograd. (Triggered internally at ../torch/csrc/autograd/autograd_not_implemented_fallback.cpp:63.)
|
| 189 |
+
[default4]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
|
| 190 |
+
[default2]:/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py:744: UserWarning: c10d::allreduce_: an autograd kernel was not registered to the Autograd key(s) but we are trying to backprop through it. This may lead to silently incorrect behavior. This behavior is deprecated and will be removed in a future version of PyTorch. If your operator is differentiable, please ensure you have registered an autograd kernel to the correct Autograd key (e.g. DispatchKey::Autograd, DispatchKey::CompositeImplicitAutograd). If your operator is not differentiable, or to squash this warning and use the previous behavior, please register torch::CppFunction::makeFallthrough() to DispatchKey::Autograd. (Triggered internally at ../torch/csrc/autograd/autograd_not_implemented_fallback.cpp:63.)
|
| 191 |
+
[default2]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
|
| 192 |
+
[default3]:/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py:744: UserWarning: c10d::allreduce_: an autograd kernel was not registered to the Autograd key(s) but we are trying to backprop through it. This may lead to silently incorrect behavior. This behavior is deprecated and will be removed in a future version of PyTorch. If your operator is differentiable, please ensure you have registered an autograd kernel to the correct Autograd key (e.g. DispatchKey::Autograd, DispatchKey::CompositeImplicitAutograd). If your operator is not differentiable, or to squash this warning and use the previous behavior, please register torch::CppFunction::makeFallthrough() to DispatchKey::Autograd. (Triggered internally at ../torch/csrc/autograd/autograd_not_implemented_fallback.cpp:63.)
|
| 193 |
+
[default3]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
|
| 194 |
+
[default1]:/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py:744: UserWarning: c10d::allreduce_: an autograd kernel was not registered to the Autograd key(s) but we are trying to backprop through it. This may lead to silently incorrect behavior. This behavior is deprecated and will be removed in a future version of PyTorch. If your operator is differentiable, please ensure you have registered an autograd kernel to the correct Autograd key (e.g. DispatchKey::Autograd, DispatchKey::CompositeImplicitAutograd). If your operator is not differentiable, or to squash this warning and use the previous behavior, please register torch::CppFunction::makeFallthrough() to DispatchKey::Autograd. (Triggered internally at ../torch/csrc/autograd/autograd_not_implemented_fallback.cpp:63.)
|
| 195 |
+
[default1]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
|
| 196 |
+
[default0]:/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py:744: UserWarning: c10d::allreduce_: an autograd kernel was not registered to the Autograd key(s) but we are trying to backprop through it. This may lead to silently incorrect behavior. This behavior is deprecated and will be removed in a future version of PyTorch. If your operator is differentiable, please ensure you have registered an autograd kernel to the correct Autograd key (e.g. DispatchKey::Autograd, DispatchKey::CompositeImplicitAutograd). If your operator is not differentiable, or to squash this warning and use the previous behavior, please register torch::CppFunction::makeFallthrough() to DispatchKey::Autograd. (Triggered internally at ../torch/csrc/autograd/autograd_not_implemented_fallback.cpp:63.)
|
| 197 |
+
[default0]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
|
| 198 |
+
[default6]:/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py:2261: UserWarning: torch.distributed.all_reduce_coalesced will be deprecated. If you must use it, please revisit our documentation later at https://pytorch.org/docs/master/distributed.html#collective-functions
|
| 199 |
+
[default6]: warnings.warn(
|
| 200 |
+
[default5]:/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py:2261: UserWarning: torch.distributed.all_reduce_coalesced will be deprecated. If you must use it, please revisit our documentation later at https://pytorch.org/docs/master/distributed.html#collective-functions
|
| 201 |
+
[default5]: warnings.warn(
|
| 202 |
+
[default3]:/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py:2261: UserWarning: torch.distributed.all_reduce_coalesced will be deprecated. If you must use it, please revisit our documentation later at https://pytorch.org/docs/master/distributed.html#collective-functions
|
| 203 |
+
[default3]: warnings.warn(
|
| 204 |
+
[default2]:/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py:2261: UserWarning: torch.distributed.all_reduce_coalesced will be deprecated. If you must use it, please revisit our documentation later at https://pytorch.org/docs/master/distributed.html#collective-functions
|
| 205 |
+
[default2]: warnings.warn(
|
| 206 |
+
[default4]:/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py:2261: UserWarning: torch.distributed.all_reduce_coalesced will be deprecated. If you must use it, please revisit our documentation later at https://pytorch.org/docs/master/distributed.html#collective-functions
|
| 207 |
+
[default4]: warnings.warn(
|
| 208 |
+
[default7]:/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py:2261: UserWarning: torch.distributed.all_reduce_coalesced will be deprecated. If you must use it, please revisit our documentation later at https://pytorch.org/docs/master/distributed.html#collective-functions
|
| 209 |
+
[default7]: warnings.warn(
|
| 210 |
+
[default1]:/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py:2261: UserWarning: torch.distributed.all_reduce_coalesced will be deprecated. If you must use it, please revisit our documentation later at https://pytorch.org/docs/master/distributed.html#collective-functions
|
| 211 |
+
[default1]: warnings.warn(
|
| 212 |
+
[default0]:07/03/2024 21:09:28 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: Memory usage: 1965.66MiB. Peak allocated 11546.57MiB. Peak reserved: 11756.00MiB
|
| 213 |
+
[default0]:/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py:2261: UserWarning: torch.distributed.all_reduce_coalesced will be deprecated. If you must use it, please revisit our documentation later at https://pytorch.org/docs/master/distributed.html#collective-functions
|
| 214 |
+
[default0]: warnings.warn(
|
| 215 |
+
[default6]:07/03/2024 21:09:35 [INFO|DP=0|PP=3|TP=0|ip-26-0-162-233]: iteration: 1 / 20 | consumed_tokens: 4.19M | elapsed_time_per_iteration_ms: 51.6K | tokens_per_sec: 81.3K | tokens_per_sec_per_gpu: 10.2K | global_batch_size: 1.02K | lm_loss: 11.2 | lr: 0.0001 | model_tflops_per_gpu: 92.2 | hardware_tflops_per_gpu: 92.2 | grad_norm: 14.8 | cuda_memory_allocated: 2.51G | cuda_max_memory_reserved: 5.82G | hd_total_memory_tb: 312G | hd_used_memory_tb: 66.1G | hd_free_memory_tb: 246G
|
| 216 |
+
[default0]:07/03/2024 21:09:35 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: Memory usage: 3478.54MiB. Peak allocated 3478.54MiB. Peak reserved: 11954.00MiB
|
| 217 |
+
[default6]:07/03/2024 21:10:02 [INFO|DP=0|PP=3|TP=0|ip-26-0-162-233]: iteration: 2 / 20 | consumed_tokens: 8.39M | elapsed_time_per_iteration_ms: 26.9K | tokens_per_sec: 156K | tokens_per_sec_per_gpu: 19.5K | global_batch_size: 1.02K | lm_loss: 11.2 | lr: 9.53e-05 | model_tflops_per_gpu: 177 | hardware_tflops_per_gpu: 177 | grad_norm: 14.9 | cuda_memory_allocated: 2.51G | cuda_max_memory_reserved: 5.82G | hd_total_memory_tb: 312G | hd_used_memory_tb: 66.1G | hd_free_memory_tb: 246G
|
| 218 |
+
[default0]:07/03/2024 21:10:02 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: Memory usage: 3478.54MiB. Peak allocated 12989.74MiB. Peak reserved: 13342.00MiB
|
| 219 |
+
[default0]:07/03/2024 21:10:02 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: Memory usage: 3478.54MiB. Peak allocated 3478.56MiB. Peak reserved: 13342.00MiB
|
| 220 |
+
[default6]:07/03/2024 21:10:29 [INFO|DP=0|PP=3|TP=0|ip-26-0-162-233]: iteration: 3 / 20 | consumed_tokens: 12.6M | elapsed_time_per_iteration_ms: 27K | tokens_per_sec: 155K | tokens_per_sec_per_gpu: 19.4K | global_batch_size: 1.02K | lm_loss: 9.53 | lr: 9.05e-05 | model_tflops_per_gpu: 176 | hardware_tflops_per_gpu: 176 | grad_norm: 35.8 | cuda_memory_allocated: 2.51G | cuda_max_memory_reserved: 5.82G | hd_total_memory_tb: 312G | hd_used_memory_tb: 66.1G | hd_free_memory_tb: 246G
|
| 221 |
+
[default0]:07/03/2024 21:10:29 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: Memory usage: 3478.54MiB. Peak allocated 12989.74MiB. Peak reserved: 13414.00MiB
|
| 222 |
+
[default0]:07/03/2024 21:10:29 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: Memory usage: 3478.54MiB. Peak allocated 3478.56MiB. Peak reserved: 13414.00MiB
|
| 223 |
+
[default0]:STAGE:2024-07-03 21:10:29 1812463:1812463 ActivityProfilerController.cpp:314] Completed Stage: Warm Up
|
| 224 |
+
[default0]:07/03/2024 21:11:03 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: Memory usage: 3478.54MiB. Peak allocated 12989.74MiB. Peak reserved: 13414.00MiB
|
| 225 |
+
[default0]:07/03/2024 21:11:03 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: Memory usage: 3478.54MiB. Peak allocated 3478.56MiB. Peak reserved: 13414.00MiB
|
| 226 |
+
[default6]:07/03/2024 21:11:03 [INFO|DP=0|PP=3|TP=0|ip-26-0-162-233]: iteration: 4 / 20 | consumed_tokens: 16.8M | elapsed_time_per_iteration_ms: 33.9K | tokens_per_sec: 124K | tokens_per_sec_per_gpu: 15.5K | global_batch_size: 1.02K | lm_loss: 12.3 | lr: 8.58e-05 | model_tflops_per_gpu: 140 | hardware_tflops_per_gpu: 140 | grad_norm: 37.4 | cuda_memory_allocated: 2.51G | cuda_max_memory_reserved: 5.82G | hd_total_memory_tb: 312G | hd_used_memory_tb: 66.1G | hd_free_memory_tb: 246G
|
| 227 |
+
[default6]:07/03/2024 21:11:37 [INFO|DP=0|PP=3|TP=0|ip-26-0-162-233]: iteration: 5 / 20 | consumed_tokens: 21M | elapsed_time_per_iteration_ms: 33.7K | tokens_per_sec: 124K | tokens_per_sec_per_gpu: 15.6K | global_batch_size: 1.02K | lm_loss: 9.94 | lr: 8.11e-05 | model_tflops_per_gpu: 141 | hardware_tflops_per_gpu: 141 | grad_norm: 14.1
|
| 228 |
+
[default0]:07/03/2024 21:11:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-162-233]: Memory usage: 3478.54MiB. Peak allocated 12989.74MiB. Peak reserved: 13414.00MiB
|
| 229 |
+
[default6]:07/03/2024 21:12:11 [INFO|DP=0|PP=3|TP=0|ip-26-0-162-233]: iteration: 6 / 20 | consumed_tokens: 25.2M | elapsed_time_per_iteration_ms: 33.7K | tokens_per_sec: 125K | tokens_per_sec_per_gpu: 15.6K | global_batch_size: 1.02K | lm_loss: 9.44 | lr: 7.63e-05 | model_tflops_per_gpu: 141 | hardware_tflops_per_gpu: 141 | grad_norm: 8.14
|
| 230 |
+
[default0]:STAGE:2024-07-03 21:13:39 1812463:1812463 ActivityProfilerController.cpp:320] Completed Stage: Collection
|
| 231 |
+
[default0]:STAGE:2024-07-03 21:13:48 1812463:1812463 ActivityProfilerController.cpp:324] Completed Stage: Post Processing
|
| 232 |
+
[default5]:[rank5]:[E ProcessGroupNCCL.cpp:563] [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=55299, OpType=RECV, NumelIn=7, NumelOut=7, Timeout(ms)=600000) ran for 600016 milliseconds before timing out.
|
| 233 |
+
[default7]:[rank7]:[E ProcessGroupNCCL.cpp:563] [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=27651, OpType=RECV, NumelIn=7, NumelOut=7, Timeout(ms)=600000) ran for 600017 milliseconds before timing out.
|
| 234 |
+
[default1]:[rank1]:[E ProcessGroupNCCL.cpp:563] [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=178199, OpType=_REDUCE_SCATTER_BASE, NumelIn=16777216, NumelOut=8388608, Timeout(ms)=600000) ran for 600000 milliseconds before timing out.
|
| 235 |
+
[default2]:[rank2]:[E ProcessGroupNCCL.cpp:563] [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=55299, OpType=RECV, NumelIn=7, NumelOut=7, Timeout(ms)=600000) ran for 600022 milliseconds before timing out.
|
| 236 |
+
[default4]:[rank4]:[E ProcessGroupNCCL.cpp:563] [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=55299, OpType=RECV, NumelIn=7, NumelOut=7, Timeout(ms)=600000) ran for 600026 milliseconds before timing out.
|
| 237 |
+
[default3]:[rank3]:[E ProcessGroupNCCL.cpp:563] [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=55299, OpType=RECV, NumelIn=7, NumelOut=7, Timeout(ms)=600000) ran for 600020 milliseconds before timing out.
|
| 238 |
+
[default6]:[rank6]:[E ProcessGroupNCCL.cpp:563] [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=27651, OpType=RECV, NumelIn=7, NumelOut=7, Timeout(ms)=600000) ran for 600018 milliseconds before timing out.
|
| 239 |
+
[default2]:[rank2]: Traceback (most recent call last):
|
| 240 |
+
[default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in <module>
|
| 241 |
+
[default4]:[rank4]: Traceback (most recent call last):
|
| 242 |
+
[default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in <module>
|
| 243 |
+
[default4]:[rank4]: trainer.train(dataloader)
|
| 244 |
+
[default2]:[rank2]: trainer.train(dataloader)
|
| 245 |
+
[default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train
|
| 246 |
+
[default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train
|
| 247 |
+
[default4]:[rank4]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader)
|
| 248 |
+
[default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step
|
| 249 |
+
[default4]:[rank4]: outputs = self.pipeline_engine.train_batch_iter(
|
| 250 |
+
[default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 252, in train_batch_iter
|
| 251 |
+
[default4]:[rank4]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model)
|
| 252 |
+
[default2]:[rank2]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader)
|
| 253 |
+
[default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step
|
| 254 |
+
[default2]:[rank2]: outputs = self.pipeline_engine.train_batch_iter(
|
| 255 |
+
[default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 252, in train_batch_iter
|
| 256 |
+
[default2]:[rank2]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model)
|
| 257 |
+
[default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward
|
| 258 |
+
[default2]:[rank2]: output = model(**micro_batch)
|
| 259 |
+
[default2]:[rank2]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
|
| 260 |
+
[default2]:[rank2]: return self._call_impl(*args, **kwargs)
|
| 261 |
+
[default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward
|
| 262 |
+
[default4]:[rank4]: output = model(**micro_batch)
|
| 263 |
+
[default4]:[rank4]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
|
| 264 |
+
[default4]:[rank4]: return self._call_impl(*args, **kwargs)
|
| 265 |
+
[default4]:[rank4]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
|
| 266 |
+
[default4]:[rank4]: return forward_call(*args, **kwargs)
|
| 267 |
+
[default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward
|
| 268 |
+
[default4]:[rank4]: sharded_logits = self.model(
|
| 269 |
+
[default2]:[rank2]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
|
| 270 |
+
[default2]:[rank2]: return forward_call(*args, **kwargs)
|
| 271 |
+
[default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward
|
| 272 |
+
[default2]:[rank2]: sharded_logits = self.model(
|
| 273 |
+
[default2]:[rank2]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
|
| 274 |
+
[default4]:[rank4]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
|
| 275 |
+
[default4]:[rank4]: return self._call_impl(*args, **kwargs)
|
| 276 |
+
[default4]:[rank4]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
|
| 277 |
+
[default4]:[rank4]: return forward_call(*args, **kwargs)
|
| 278 |
+
[default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward
|
| 279 |
+
[default4]:[rank4]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0]
|
| 280 |
+
[default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states
|
| 281 |
+
[default4]:[rank4]: hidden_encoder_states = encoder_block(**hidden_encoder_states)
|
| 282 |
+
[default4]:[rank4]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
|
| 283 |
+
[default4]:[rank4]: return self._call_impl(*args, **kwargs)
|
| 284 |
+
[default4]:[rank4]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
|
| 285 |
+
[default4]:[rank4]: return forward_call(*args, **kwargs)
|
| 286 |
+
[default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 126, in forward
|
| 287 |
+
[default2]:[rank2]: return self._call_impl(*args, **kwargs)
|
| 288 |
+
[default2]:[rank2]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
|
| 289 |
+
[default2]:[rank2]: return forward_call(*args, **kwargs)
|
| 290 |
+
[default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward
|
| 291 |
+
[default2]:[rank2]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0]
|
| 292 |
+
[default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states
|
| 293 |
+
[default4]:[rank4]: new_kwargs[name] = recv_from_pipeline_state_buffer(
|
| 294 |
+
[default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 117, in recv_from_pipeline_state_buffer
|
| 295 |
+
[default4]:[rank4]: pipeline_state.run_communication()
|
| 296 |
+
[default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 150, in run_communication
|
| 297 |
+
[default2]:[rank2]: hidden_encoder_states = encoder_block(**hidden_encoder_states)
|
| 298 |
+
[default2]:[rank2]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
|
| 299 |
+
[default2]:[rank2]: return self._call_impl(*args, **kwargs)
|
| 300 |
+
[default4]:[rank4]: recv_activation_tensor = recv_activation()
|
| 301 |
+
[default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 31, in __call__
|
| 302 |
+
[default4]:[rank4]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0]
|
| 303 |
+
[default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors
|
| 304 |
+
[default4]:[rank4]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag)
|
| 305 |
+
[default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors
|
| 306 |
+
[default4]:[rank4]: meta = self._recv_meta(from_rank=from_rank, tag=tag)
|
| 307 |
+
[default2]:[rank2]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
|
| 308 |
+
[default2]:[rank2]: return forward_call(*args, **kwargs)
|
| 309 |
+
[default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 126, in forward
|
| 310 |
+
[default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 267, in _recv_meta
|
| 311 |
+
[default2]:[rank2]: new_kwargs[name] = recv_from_pipeline_state_buffer(
|
| 312 |
+
[default4]:[rank4]: self.second_metadata = torch.empty(second_metadata_num_elements, dtype=torch.long, device=self.device)
|
| 313 |
+
[default4]:[rank4]: torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate more than 1EB memory.
|
| 314 |
+
[default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 117, in recv_from_pipeline_state_buffer
|
| 315 |
+
[default2]:[rank2]: pipeline_state.run_communication()
|
| 316 |
+
[default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 150, in run_communication
|
| 317 |
+
[default2]:[rank2]: recv_activation_tensor = recv_activation()
|
| 318 |
+
[default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 31, in __call__
|
| 319 |
+
[default2]:[rank2]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0]
|
| 320 |
+
[default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors
|
| 321 |
+
[default2]:[rank2]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag)
|
| 322 |
+
[default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors
|
| 323 |
+
[default2]:[rank2]: meta = self._recv_meta(from_rank=from_rank, tag=tag)
|
| 324 |
+
[default2]:[rank2]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 267, in _recv_meta
|
| 325 |
+
[default2]:[rank2]: self.second_metadata = torch.empty(second_metadata_num_elements, dtype=torch.long, device=self.device)
|
| 326 |
+
[default2]:[rank2]: torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate more than 1EB memory.
|
| 327 |
+
[default6]:[rank6]: Traceback (most recent call last):
|
| 328 |
+
[default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in <module>
|
| 329 |
+
[default6]:[rank6]: trainer.train(dataloader)
|
| 330 |
+
[default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train
|
| 331 |
+
[default6]:[rank6]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader)
|
| 332 |
+
[default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step
|
| 333 |
+
[default6]:[rank6]: outputs = self.pipeline_engine.train_batch_iter(
|
| 334 |
+
[default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 278, in train_batch_iter
|
| 335 |
+
[default6]:[rank6]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model)
|
| 336 |
+
[default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward
|
| 337 |
+
[default6]:[rank6]: output = model(**micro_batch)
|
| 338 |
+
[default6]:[rank6]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
|
| 339 |
+
[default6]:[rank6]: return self._call_impl(*args, **kwargs)
|
| 340 |
+
[default6]:[rank6]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
|
| 341 |
+
[default6]:[rank6]: return forward_call(*args, **kwargs)
|
| 342 |
+
[default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward
|
| 343 |
+
[default6]:[rank6]: sharded_logits = self.model(
|
| 344 |
+
[default6]:[rank6]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
|
| 345 |
+
[default6]:[rank6]: return self._call_impl(*args, **kwargs)
|
| 346 |
+
[default3]:[rank3]: Traceback (most recent call last):
|
| 347 |
+
[default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in <module>
|
| 348 |
+
[default6]:[rank6]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
|
| 349 |
+
[default3]:[rank3]: trainer.train(dataloader)
|
| 350 |
+
[default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train
|
| 351 |
+
[default3]:[rank3]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader)
|
| 352 |
+
[default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step
|
| 353 |
+
[default6]:[rank6]: return forward_call(*args, **kwargs)
|
| 354 |
+
[default3]:[rank3]: outputs = self.pipeline_engine.train_batch_iter(
|
| 355 |
+
[default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward
|
| 356 |
+
[default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 252, in train_batch_iter
|
| 357 |
+
[default6]:[rank6]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0]
|
| 358 |
+
[default3]:[rank3]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model)
|
| 359 |
+
[default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states
|
| 360 |
+
[default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward
|
| 361 |
+
[default6]:[rank6]: hidden_encoder_states = encoder_block(**hidden_encoder_states)
|
| 362 |
+
[default6]:[rank6]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
|
| 363 |
+
[default3]:[rank3]: output = model(**micro_batch)
|
| 364 |
+
[default3]:[rank3]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
|
| 365 |
+
[default6]:[rank6]: return self._call_impl(*args, **kwargs)
|
| 366 |
+
[default3]:[rank3]: return self._call_impl(*args, **kwargs)
|
| 367 |
+
[default3]:[rank3]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
|
| 368 |
+
[default3]:[rank3]: return forward_call(*args, **kwargs)
|
| 369 |
+
[default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward
|
| 370 |
+
[default3]:[rank3]: sharded_logits = self.model(
|
| 371 |
+
[default6]:[rank6]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
|
| 372 |
+
[default6]:[rank6]: return forward_call(*args, **kwargs)
|
| 373 |
+
[default3]:[rank3]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
|
| 374 |
+
[default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 126, in forward
|
| 375 |
+
[default6]:[rank6]: new_kwargs[name] = recv_from_pipeline_state_buffer(
|
| 376 |
+
[default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 117, in recv_from_pipeline_state_buffer
|
| 377 |
+
[default3]:[rank3]: return self._call_impl(*args, **kwargs)
|
| 378 |
+
[default3]:[rank3]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
|
| 379 |
+
[default6]:[rank6]: pipeline_state.run_communication()
|
| 380 |
+
[default3]:[rank3]: return forward_call(*args, **kwargs)
|
| 381 |
+
[default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward
|
| 382 |
+
[default3]:[rank3]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0]
|
| 383 |
+
[default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states
|
| 384 |
+
[default3]:[rank3]: hidden_encoder_states = encoder_block(**hidden_encoder_states)
|
| 385 |
+
[default3]:[rank3]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
|
| 386 |
+
[default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 150, in run_communication
|
| 387 |
+
[default6]:[rank6]: recv_activation_tensor = recv_activation()
|
| 388 |
+
[default3]:[rank3]: return self._call_impl(*args, **kwargs)
|
| 389 |
+
[default3]:[rank3]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
|
| 390 |
+
[default3]:[rank3]: return forward_call(*args, **kwargs)
|
| 391 |
+
[default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 31, in __call__
|
| 392 |
+
[default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 126, in forward
|
| 393 |
+
[default6]:[rank6]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0]
|
| 394 |
+
[default3]:[rank3]: new_kwargs[name] = recv_from_pipeline_state_buffer(
|
| 395 |
+
[default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors
|
| 396 |
+
[default6]:[rank6]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag)
|
| 397 |
+
[default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors
|
| 398 |
+
[default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 117, in recv_from_pipeline_state_buffer
|
| 399 |
+
[default6]:[rank6]: meta = self._recv_meta(from_rank=from_rank, tag=tag)
|
| 400 |
+
[default3]:[rank3]: pipeline_state.run_communication()
|
| 401 |
+
[default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 267, in _recv_meta
|
| 402 |
+
[default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 150, in run_communication
|
| 403 |
+
[default3]:[rank3]: recv_activation_tensor = recv_activation()
|
| 404 |
+
[default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 31, in __call__
|
| 405 |
+
[default3]:[rank3]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0]
|
| 406 |
+
[default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors
|
| 407 |
+
[default6]:[rank6]: self.second_metadata = torch.empty(second_metadata_num_elements, dtype=torch.long, device=self.device)
|
| 408 |
+
[default6]:[rank6]: torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate more than 1EB memory.
|
| 409 |
+
[default3]:[rank3]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag)
|
| 410 |
+
[default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors
|
| 411 |
+
[default3]:[rank3]: meta = self._recv_meta(from_rank=from_rank, tag=tag)
|
| 412 |
+
[default3]:[rank3]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 267, in _recv_meta
|
| 413 |
+
[default3]:[rank3]: self.second_metadata = torch.empty(second_metadata_num_elements, dtype=torch.long, device=self.device)
|
| 414 |
+
[default3]:[rank3]: torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate more than 1EB memory.
|
| 415 |
+
[default5]:[rank5]: Traceback (most recent call last):
|
| 416 |
+
[default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in <module>
|
| 417 |
+
[default5]:[rank5]: trainer.train(dataloader)
|
| 418 |
+
[default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train
|
| 419 |
+
[default5]:[rank5]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader)
|
| 420 |
+
[default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step
|
| 421 |
+
[default5]:[rank5]: outputs = self.pipeline_engine.train_batch_iter(
|
| 422 |
+
[default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 252, in train_batch_iter
|
| 423 |
+
[default5]:[rank5]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model)
|
| 424 |
+
[default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward
|
| 425 |
+
[default5]:[rank5]: output = model(**micro_batch)
|
| 426 |
+
[default5]:[rank5]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
|
| 427 |
+
[default5]:[rank5]: return self._call_impl(*args, **kwargs)
|
| 428 |
+
[default5]:[rank5]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
|
| 429 |
+
[default5]:[rank5]: return forward_call(*args, **kwargs)
|
| 430 |
+
[default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward
|
| 431 |
+
[default5]:[rank5]: sharded_logits = self.model(
|
| 432 |
+
[default5]:[rank5]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
|
| 433 |
+
[default5]:[rank5]: return self._call_impl(*args, **kwargs)
|
| 434 |
+
[default5]:[rank5]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
|
| 435 |
+
[default5]:[rank5]: return forward_call(*args, **kwargs)
|
| 436 |
+
[default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward
|
| 437 |
+
[default5]:[rank5]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0]
|
| 438 |
+
[default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states
|
| 439 |
+
[default5]:[rank5]: hidden_encoder_states = encoder_block(**hidden_encoder_states)
|
| 440 |
+
[default5]:[rank5]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
|
| 441 |
+
[default5]:[rank5]: return self._call_impl(*args, **kwargs)
|
| 442 |
+
[default5]:[rank5]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
|
| 443 |
+
[default5]:[rank5]: return forward_call(*args, **kwargs)
|
| 444 |
+
[default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 126, in forward
|
| 445 |
+
[default5]:[rank5]: new_kwargs[name] = recv_from_pipeline_state_buffer(
|
| 446 |
+
[default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 117, in recv_from_pipeline_state_buffer
|
| 447 |
+
[default5]:[rank5]: pipeline_state.run_communication()
|
| 448 |
+
[default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 150, in run_communication
|
| 449 |
+
[default5]:[rank5]: recv_activation_tensor = recv_activation()
|
| 450 |
+
[default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 31, in __call__
|
| 451 |
+
[default5]:[rank5]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0]
|
| 452 |
+
[default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors
|
| 453 |
+
[default5]:[rank5]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag)
|
| 454 |
+
[default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors
|
| 455 |
+
[default5]:[rank5]: meta = self._recv_meta(from_rank=from_rank, tag=tag)
|
| 456 |
+
[default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 267, in _recv_meta
|
| 457 |
+
[default5]:[rank5]: self.second_metadata = torch.empty(second_metadata_num_elements, dtype=torch.long, device=self.device)
|
| 458 |
+
[default5]:[rank5]: torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate more than 1EB memory.
|
| 459 |
+
[default7]:[rank7]: Traceback (most recent call last):
|
| 460 |
+
[default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in <module>
|
| 461 |
+
[default7]:[rank7]: trainer.train(dataloader)
|
| 462 |
+
[default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train
|
| 463 |
+
[default7]:[rank7]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader)
|
| 464 |
+
[default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step
|
| 465 |
+
[default7]:[rank7]: outputs = self.pipeline_engine.train_batch_iter(
|
| 466 |
+
[default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 278, in train_batch_iter
|
| 467 |
+
[default7]:[rank7]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model)
|
| 468 |
+
[default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward
|
| 469 |
+
[default7]:[rank7]: output = model(**micro_batch)
|
| 470 |
+
[default7]:[rank7]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
|
| 471 |
+
[default7]:[rank7]: return self._call_impl(*args, **kwargs)
|
| 472 |
+
[default7]:[rank7]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
|
| 473 |
+
[default7]:[rank7]: return forward_call(*args, **kwargs)
|
| 474 |
+
[default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward
|
| 475 |
+
[default7]:[rank7]: sharded_logits = self.model(
|
| 476 |
+
[default7]:[rank7]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
|
| 477 |
+
[default7]:[rank7]: return self._call_impl(*args, **kwargs)
|
| 478 |
+
[default7]:[rank7]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
|
| 479 |
+
[default7]:[rank7]: return forward_call(*args, **kwargs)
|
| 480 |
+
[default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward
|
| 481 |
+
[default7]:[rank7]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0]
|
| 482 |
+
[default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states
|
| 483 |
+
[default7]:[rank7]: hidden_encoder_states = encoder_block(**hidden_encoder_states)
|
| 484 |
+
[default7]:[rank7]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
|
| 485 |
+
[default7]:[rank7]: return self._call_impl(*args, **kwargs)
|
| 486 |
+
[default7]:[rank7]: File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
|
| 487 |
+
[default7]:[rank7]: return forward_call(*args, **kwargs)
|
| 488 |
+
[default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 126, in forward
|
| 489 |
+
[default7]:[rank7]: new_kwargs[name] = recv_from_pipeline_state_buffer(
|
| 490 |
+
[default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/functional.py", line 117, in recv_from_pipeline_state_buffer
|
| 491 |
+
[default7]:[rank7]: pipeline_state.run_communication()
|
| 492 |
+
[default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 150, in run_communication
|
| 493 |
+
[default7]:[rank7]: recv_activation_tensor = recv_activation()
|
| 494 |
+
[default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 31, in __call__
|
| 495 |
+
[default7]:[rank7]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0]
|
| 496 |
+
[default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors
|
| 497 |
+
[default7]:[rank7]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag)
|
| 498 |
+
[default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors
|
| 499 |
+
[default7]:[rank7]: meta = self._recv_meta(from_rank=from_rank, tag=tag)
|
| 500 |
+
[default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 267, in _recv_meta
|
| 501 |
+
[default7]:[rank7]: self.second_metadata = torch.empty(second_metadata_num_elements, dtype=torch.long, device=self.device)
|
| 502 |
+
[default7]:[rank7]: torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate more than 1EB memory.
|
| 503 |
+
[default7]:[rank7]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 3] Timeout at NCCL work: 27651, last enqueued NCCL work: 27651, last completed NCCL work: 27650.
|
| 504 |
+
[default7]:[rank7]:[E ProcessGroupNCCL.cpp:577] [Rank 3] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data.
|
| 505 |
+
[default7]:[rank7]:[E ProcessGroupNCCL.cpp:583] [Rank 3] To avoid data inconsistency, we are taking the entire process down.
|
| 506 |
+
[default7]:[rank7]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 3] Process group watchdog thread terminated with exception: [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=27651, OpType=RECV, NumelIn=7, NumelOut=7, Timeout(ms)=600000) ran for 600017 milliseconds before timing out.
|
| 507 |
+
[default7]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first):
|
| 508 |
+
[default7]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fe41d450897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so)
|
| 509 |
+
[default7]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional<std::chrono::duration<long, std::ratio<1l, 1000l> > >) + 0x1d2 (0x7fe41e729c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
|
| 510 |
+
[default7]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7fe41e72ea80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
|
| 511 |
+
[default7]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7fe41e72fdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
|
| 512 |
+
[default7]:frame #4: <unknown function> + 0xd3e95 (0x7fe46a1c8e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6)
|
| 513 |
+
[default7]:frame #5: <unknown function> + 0x8609 (0x7fe46f20f609 in /lib/x86_64-linux-gnu/libpthread.so.0)
|
| 514 |
+
[default7]:frame #6: clone + 0x43 (0x7fe46efda353 in /lib/x86_64-linux-gnu/libc.so.6)
|
| 515 |
+
[default7]:
|
| 516 |
+
[default7]:terminate called after throwing an instance of 'c10::DistBackendError'
|
| 517 |
+
[default7]: what(): [PG 4 Rank 3] Process group watchdog thread terminated with exception: [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=27651, OpType=RECV, NumelIn=7, NumelOut=7, Timeout(ms)=600000) ran for 600017 milliseconds before timing out.
|
| 518 |
+
[default7]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first):
|
| 519 |
+
[default7]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fe41d450897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so)
|
| 520 |
+
[default7]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional<std::chrono::duration<long, std::ratio<1l, 1000l> > >) + 0x1d2 (0x7fe41e729c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
|
| 521 |
+
[default7]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7fe41e72ea80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
|
| 522 |
+
[default7]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7fe41e72fdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
|
| 523 |
+
[default7]:frame #4: <unknown function> + 0xd3e95 (0x7fe46a1c8e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6)
|
| 524 |
+
[default7]:frame #5: <unknown function> + 0x8609 (0x7fe46f20f609 in /lib/x86_64-linux-gnu/libpthread.so.0)
|
| 525 |
+
[default7]:frame #6: clone + 0x43 (0x7fe46efda353 in /lib/x86_64-linux-gnu/libc.so.6)
|
| 526 |
+
[default7]:
|
| 527 |
+
[default7]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first):
|
| 528 |
+
[default7]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fe41d450897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so)
|
| 529 |
+
[default7]:frame #1: <unknown function> + 0xe32119 (0x7fe41e3b3119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
|
| 530 |
+
[default7]:frame #2: <unknown function> + 0xd3e95 (0x7fe46a1c8e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6)
|
| 531 |
+
[default7]:frame #3: <unknown function> + 0x8609 (0x7fe46f20f609 in /lib/x86_64-linux-gnu/libpthread.so.0)
|
| 532 |
+
[default7]:frame #4: clone + 0x43 (0x7fe46efda353 in /lib/x86_64-linux-gnu/libc.so.6)
|
| 533 |
+
[default7]:
|
| 534 |
+
[default6]:[rank6]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 3] Timeout at NCCL work: 27651, last enqueued NCCL work: 27651, last completed NCCL work: 27650.
|
| 535 |
+
[default6]:[rank6]:[E ProcessGroupNCCL.cpp:577] [Rank 3] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data.
|
| 536 |
+
[default6]:[rank6]:[E ProcessGroupNCCL.cpp:583] [Rank 3] To avoid data inconsistency, we are taking the entire process down.
|
| 537 |
+
[default6]:[rank6]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 3] Process group watchdog thread terminated with exception: [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=27651, OpType=RECV, NumelIn=7, NumelOut=7, Timeout(ms)=600000) ran for 600018 milliseconds before timing out.
|
| 538 |
+
[default6]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first):
|
| 539 |
+
[default6]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fe17f140897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so)
|
| 540 |
+
[default6]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional<std::chrono::duration<long, std::ratio<1l, 1000l> > >) + 0x1d2 (0x7fe180419c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
|
| 541 |
+
[default6]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7fe18041ea80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
|
| 542 |
+
[default6]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7fe18041fdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
|
| 543 |
+
[default6]:frame #4: <unknown function> + 0xd3e95 (0x7fe1cbeb8e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6)
|
| 544 |
+
[default6]:frame #5: <unknown function> + 0x8609 (0x7fe1d0eff609 in /lib/x86_64-linux-gnu/libpthread.so.0)
|
| 545 |
+
[default6]:frame #6: clone + 0x43 (0x7fe1d0cca353 in /lib/x86_64-linux-gnu/libc.so.6)
|
| 546 |
+
[default6]:
|
| 547 |
+
[default6]:terminate called after throwing an instance of 'c10::DistBackendError'
|
| 548 |
+
[default6]: what(): [PG 4 Rank 3] Process group watchdog thread terminated with exception: [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=27651, OpType=RECV, NumelIn=7, NumelOut=7, Timeout(ms)=600000) ran for 600018 milliseconds before timing out.
|
| 549 |
+
[default6]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first):
|
| 550 |
+
[default6]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fe17f140897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so)
|
| 551 |
+
[default6]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional<std::chrono::duration<long, std::ratio<1l, 1000l> > >) + 0x1d2 (0x7fe180419c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
|
| 552 |
+
[default6]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7fe18041ea80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
|
| 553 |
+
[default6]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7fe18041fdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
|
| 554 |
+
[default6]:frame #4: <unknown function> + 0xd3e95 (0x7fe1cbeb8e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6)
|
| 555 |
+
[default6]:frame #5: <unknown function> + 0x8609 (0x7fe1d0eff609 in /lib/x86_64-linux-gnu/libpthread.so.0)
|
| 556 |
+
[default6]:frame #6: clone + 0x43 (0x7fe1d0cca353 in /lib/x86_64-linux-gnu/libc.so.6)
|
| 557 |
+
[default6]:
|
| 558 |
+
[default6]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first):
|
| 559 |
+
[default6]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fe17f140897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so)
|
| 560 |
+
[default6]:frame #1: <unknown function> + 0xe32119 (0x7fe1800a3119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
|
| 561 |
+
[default6]:frame #2: <unknown function> + 0xd3e95 (0x7fe1cbeb8e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6)
|
| 562 |
+
[default6]:frame #3: <unknown function> + 0x8609 (0x7fe1d0eff609 in /lib/x86_64-linux-gnu/libpthread.so.0)
|
| 563 |
+
[default6]:frame #4: clone + 0x43 (0x7fe1d0cca353 in /lib/x86_64-linux-gnu/libc.so.6)
|
| 564 |
+
[default6]:
|
| 565 |
+
W0703 21:22:15.080000 139794302568256 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1812463 closing signal SIGTERM
|
| 566 |
+
W0703 21:22:15.080000 139794302568256 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1812464 closing signal SIGTERM
|
| 567 |
+
W0703 21:22:15.080000 139794302568256 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1812465 closing signal SIGTERM
|
| 568 |
+
W0703 21:22:15.080000 139794302568256 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1812466 closing signal SIGTERM
|
| 569 |
+
W0703 21:22:15.085000 139794302568256 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1812467 closing signal SIGTERM
|
| 570 |
+
W0703 21:22:15.085000 139794302568256 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 1812468 closing signal SIGTERM
|
| 571 |
+
E0703 21:22:19.843000 139794302568256 torch/distributed/elastic/multiprocessing/api.py:826] failed (exitcode: -6) local_rank: 6 (pid: 1812469) of binary: /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/python3.10
|
| 572 |
+
Traceback (most recent call last):
|
| 573 |
+
File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/torchrun", line 8, in <module>
|
| 574 |
+
sys.exit(main())
|
| 575 |
+
File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 347, in wrapper
|
| 576 |
+
return f(*args, **kwargs)
|
| 577 |
+
File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/run.py", line 879, in main
|
| 578 |
+
run(args)
|
| 579 |
+
File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/run.py", line 870, in run
|
| 580 |
+
elastic_launch(
|
| 581 |
+
File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 132, in __call__
|
| 582 |
+
return launch_agent(self._config, self._entrypoint, list(args))
|
| 583 |
+
File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 263, in launch_agent
|
| 584 |
+
raise ChildFailedError(
|
| 585 |
+
torch.distributed.elastic.multiprocessing.errors.ChildFailedError:
|
| 586 |
+
============================================================
|
| 587 |
+
/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py FAILED
|
| 588 |
+
------------------------------------------------------------
|
| 589 |
+
Failures:
|
| 590 |
+
[1]:
|
| 591 |
+
time : 2024-07-03_21:22:15
|
| 592 |
+
host : ip-26-0-162-233.ec2.internal
|
| 593 |
+
rank : 7 (local_rank: 7)
|
| 594 |
+
exitcode : -6 (pid: 1812470)
|
| 595 |
+
error_file: <N/A>
|
| 596 |
+
traceback : Signal 6 (SIGABRT) received by PID 1812470
|
| 597 |
+
------------------------------------------------------------
|
| 598 |
+
Root Cause (first observed failure):
|
| 599 |
+
[0]:
|
| 600 |
+
time : 2024-07-03_21:22:15
|
| 601 |
+
host : ip-26-0-162-233.ec2.internal
|
| 602 |
+
rank : 6 (local_rank: 6)
|
| 603 |
+
exitcode : -6 (pid: 1812469)
|
| 604 |
+
error_file: <N/A>
|
| 605 |
+
traceback : Signal 6 (SIGABRT) received by PID 1812469
|
| 606 |
+
============================================================
|
| 607 |
+
srun: error: ip-26-0-162-233: task 0: Exited with exit code 1
|
| 608 |
+
Consider using `hf_transfer` for faster uploads. This solution comes with some limitations. See https://huggingface.co/docs/huggingface_hub/hf_transfer for more details.
|
llama-1B/8_GPUS/dp-1_tp-2_pp-4_mbz-2/status.txt
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
oom
|