Upload llama-1B/8_GPUS/dp-1_tp-4_pp-2_mbz-4
Browse files- .gitattributes +1 -0
- llama-1B/8_GPUS/dp-1_tp-4_pp-2_mbz-4/bench.slurm +111 -0
- llama-1B/8_GPUS/dp-1_tp-4_pp-2_mbz-4/config.yaml +90 -0
- llama-1B/8_GPUS/dp-1_tp-4_pp-2_mbz-4/log.out +595 -0
- llama-1B/8_GPUS/dp-1_tp-4_pp-2_mbz-4/profiler/ip-26-0-169-86_2274198.1720050009949646955.pt.trace.json.tmp +3 -0
- llama-1B/8_GPUS/dp-1_tp-4_pp-2_mbz-4/status.txt +1 -0
.gitattributes
CHANGED
|
@@ -116,3 +116,4 @@ llama-1B/8_GPUS/dp-2_tp-1_pp-4_mbz-1/profiler/ip-26-0-174-36_243244.172004896077
|
|
| 116 |
llama-1B/8_GPUS/dp-1_tp-2_pp-4_mbz-4/profiler/ip-26-0-164-187_27331.1720048318217251415.pt.trace.json filter=lfs diff=lfs merge=lfs -text
|
| 117 |
llama-1B/8_GPUS/dp-2_tp-4_pp-1_mbz-4/profiler/ip-26-0-163-220_412762.1720048514231027509.pt.trace.json filter=lfs diff=lfs merge=lfs -text
|
| 118 |
llama-1B/8_GPUS/dp-1_tp-8_pp-1_mbz-32/profiler/ip-26-0-171-88_1097996.1720048094691939359.pt.trace.json filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
| 116 |
llama-1B/8_GPUS/dp-1_tp-2_pp-4_mbz-4/profiler/ip-26-0-164-187_27331.1720048318217251415.pt.trace.json filter=lfs diff=lfs merge=lfs -text
|
| 117 |
llama-1B/8_GPUS/dp-2_tp-4_pp-1_mbz-4/profiler/ip-26-0-163-220_412762.1720048514231027509.pt.trace.json filter=lfs diff=lfs merge=lfs -text
|
| 118 |
llama-1B/8_GPUS/dp-1_tp-8_pp-1_mbz-32/profiler/ip-26-0-171-88_1097996.1720048094691939359.pt.trace.json filter=lfs diff=lfs merge=lfs -text
|
| 119 |
+
llama-1B/8_GPUS/dp-1_tp-4_pp-2_mbz-4/profiler/ip-26-0-169-86_2274198.1720050009949646955.pt.trace.json.tmp filter=lfs diff=lfs merge=lfs -text
|
llama-1B/8_GPUS/dp-1_tp-4_pp-2_mbz-4/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-4_pp-2_mbz-4/log.out
|
| 13 |
+
#SBATCH --error=/fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-1_tp-4_pp-2_mbz-4/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-4_pp-2_mbz-4/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-4_pp-2_mbz-4/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-4_pp-2_mbz-4/status.txt
|
| 84 |
+
else
|
| 85 |
+
if grep -q "OutOfMemoryError" /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-1_tp-4_pp-2_mbz-4/log.out; then
|
| 86 |
+
printf "oom" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-1_tp-4_pp-2_mbz-4/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-4_pp-2_mbz-4/log.out; then
|
| 88 |
+
printf "oom" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-1_tp-4_pp-2_mbz-4/status.txt
|
| 89 |
+
elif grep -q "Timeout at NCCL" /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-1_tp-4_pp-2_mbz-4/log.out; then
|
| 90 |
+
printf "timeout" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-1_tp-4_pp-2_mbz-4/status.txt
|
| 91 |
+
else
|
| 92 |
+
printf "fail" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-1_tp-4_pp-2_mbz-4/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-4_pp-2_mbz-4 --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-4_pp-2_mbz-4 --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-4_pp-2_mbz-4 llama-1B/8_GPUS/dp-1_tp-4_pp-2_mbz-4 --commit-message "Upload llama-1B/8_GPUS/dp-1_tp-4_pp-2_mbz-4"
|
| 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-4_pp-2_mbz-4/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: 2
|
| 51 |
+
pp_engine: 1f1b
|
| 52 |
+
tp: 4
|
| 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-4_pp-2_mbz-4
|
| 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: 256
|
| 79 |
+
limit_test_batches: 0
|
| 80 |
+
limit_val_batches: 0
|
| 81 |
+
micro_batch_size: 4
|
| 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-4_pp-2_mbz-4/log.out
ADDED
|
@@ -0,0 +1,595 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
========================
|
| 2 |
+
START TIME: Wed Jul 3 23:26:01 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 23:26:10.004000 139973789022016 torch/distributed/run.py:757]
|
| 18 |
+
W0703 23:26:10.004000 139973789022016 torch/distributed/run.py:757] *****************************************
|
| 19 |
+
W0703 23:26:10.004000 139973789022016 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 23:26:10.004000 139973789022016 torch/distributed/run.py:757] *****************************************
|
| 21 |
+
[default0]:07/03/2024 23:26:32 [WARNING|DP=0|PP=0|TP=0|ip-26-0-169-86]: [Vocab Size Padding] Padded vocab (size: 50257) with 3 dummy tokens (new size: 50260)
|
| 22 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: Config:
|
| 23 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: Config(general=GeneralArgs(project='bench_cluster',
|
| 24 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: run='%date_%jobid',
|
| 25 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: seed=42,
|
| 26 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: step=None,
|
| 27 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: consumed_train_samples=None,
|
| 28 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: benchmark_csv_path=None,
|
| 29 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: ignore_sanity_checks=True),
|
| 30 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: parallelism=ParallelismArgs(dp=1,
|
| 31 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: pp=2,
|
| 32 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: tp=4,
|
| 33 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: pp_engine=<nanotron.parallel.pipeline_parallel.engine.OneForwardOneBackwardPipelineEngine object at 0x7f94edff88e0>,
|
| 34 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: tp_mode=<TensorParallelLinearMode.REDUCE_SCATTER: 2>,
|
| 35 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: tp_linear_async_communication=False,
|
| 36 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: expert_parallel_size=1),
|
| 37 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: model=ModelArgs(model_config=LlamaConfig(bos_token_id=1,
|
| 38 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: eos_token_id=2,
|
| 39 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: hidden_act='silu',
|
| 40 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: hidden_size=2048,
|
| 41 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: initializer_range=0.02,
|
| 42 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: intermediate_size=4096,
|
| 43 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: is_llama_config=True,
|
| 44 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: max_position_embeddings=4096,
|
| 45 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: num_attention_heads=32,
|
| 46 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: num_hidden_layers=24,
|
| 47 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: num_key_value_heads=32,
|
| 48 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: pad_token_id=None,
|
| 49 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: pretraining_tp=1,
|
| 50 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: rms_norm_eps=1e-05,
|
| 51 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: rope_scaling=None,
|
| 52 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: rope_theta=10000.0,
|
| 53 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: tie_word_embeddings=True,
|
| 54 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: use_cache=True,
|
| 55 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: vocab_size=50260),
|
| 56 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: init_method=RandomInit(std=0.025),
|
| 57 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: dtype=torch.bfloat16,
|
| 58 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: make_vocab_size_divisible_by=1,
|
| 59 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: ddp_bucket_cap_mb=25),
|
| 60 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: tokenizer=TokenizerArgs(tokenizer_name_or_path='openai-community/gpt2',
|
| 61 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: tokenizer_revision=None,
|
| 62 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: tokenizer_max_length=None),
|
| 63 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: checkpoints=CheckpointsArgs(checkpoints_path=Path('/dev/null'),
|
| 64 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: checkpoint_interval=100000,
|
| 65 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: save_initial_state=False,
|
| 66 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: resume_checkpoint_path=None,
|
| 67 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: checkpoints_path_is_shared_file_system=False),
|
| 68 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: logging=LoggingArgs(log_level='info',
|
| 69 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: log_level_replica='info',
|
| 70 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: iteration_step_info_interval=1),
|
| 71 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: tokens=TokensArgs(sequence_length=4096,
|
| 72 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: train_steps=20,
|
| 73 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: micro_batch_size=4,
|
| 74 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: batch_accumulation_per_replica=256,
|
| 75 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: val_check_interval=-1,
|
| 76 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: limit_val_batches=0,
|
| 77 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: limit_test_batches=0),
|
| 78 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: optimizer=OptimizerArgs(optimizer_factory=AdamWOptimizerArgs(adam_eps=1e-08,
|
| 79 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: adam_beta1=0.9,
|
| 80 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: adam_beta2=0.95,
|
| 81 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: torch_adam_is_fused=True,
|
| 82 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: name='adamW'),
|
| 83 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: zero_stage=1,
|
| 84 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: weight_decay=0.01,
|
| 85 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: clip_grad=1.0,
|
| 86 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: accumulate_grad_in_fp32=True,
|
| 87 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: learning_rate_scheduler=LRSchedulerArgs(learning_rate=0.0001,
|
| 88 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: lr_warmup_steps=1,
|
| 89 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: lr_warmup_style='linear',
|
| 90 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: lr_decay_style='linear',
|
| 91 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: lr_decay_steps=19,
|
| 92 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: lr_decay_starting_step=None,
|
| 93 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: min_decay_lr=1e-05)),
|
| 94 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: data_stages=[DatasetStageArgs(name='Training Stage',
|
| 95 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: start_training_step=1,
|
| 96 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: data=DataArgs(dataset=PretrainDatasetsArgs(hf_dataset_or_datasets='roneneldan/TinyStories',
|
| 97 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: hf_dataset_splits='train',
|
| 98 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: hf_dataset_config_name=None,
|
| 99 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: dataset_processing_num_proc_per_process=64,
|
| 100 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: dataset_overwrite_cache=False,
|
| 101 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: text_column_name='text'),
|
| 102 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: seed=42,
|
| 103 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: num_loading_workers=0))],
|
| 104 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: profiler=ProfilerArgs(profiler_export_path=Path('/fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-1_tp-4_pp-2_mbz-4')),
|
| 105 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: lighteval=None)
|
| 106 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: Model Config:
|
| 107 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: LlamaConfig(bos_token_id=1,
|
| 108 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: eos_token_id=2,
|
| 109 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: hidden_act='silu',
|
| 110 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: hidden_size=2048,
|
| 111 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: initializer_range=0.02,
|
| 112 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: intermediate_size=4096,
|
| 113 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: is_llama_config=True,
|
| 114 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: max_position_embeddings=4096,
|
| 115 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: num_attention_heads=32,
|
| 116 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: num_hidden_layers=24,
|
| 117 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: num_key_value_heads=32,
|
| 118 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: pad_token_id=None,
|
| 119 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: pretraining_tp=1,
|
| 120 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: rms_norm_eps=1e-05,
|
| 121 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: rope_scaling=None,
|
| 122 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: rope_theta=10000.0,
|
| 123 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: tie_word_embeddings=True,
|
| 124 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: use_cache=True,
|
| 125 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: vocab_size=50260)
|
| 126 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: Building model..
|
| 127 |
+
[default0]:07/03/2024 23:26:32 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: Setting PP block ranks...
|
| 128 |
+
[default5]:07/03/2024 23:26:47 [INFO|DP=0|PP=1|TP=1|ip-26-0-169-86]: Local number of parameters: 131M (249.16MiB)
|
| 129 |
+
[default5]:07/03/2024 23:26:47 [INFO|DP=0|PP=1|TP=1|ip-26-0-169-86]: [After model building] Memory usage: 260.10MiB. Peak allocated: 262.13MiB Peak reserved: 264.00MiB
|
| 130 |
+
[default5]:07/03/2024 23:26:47 [INFO|DP=0|PP=1|TP=1|ip-26-0-169-86]: No checkpoint path provided.
|
| 131 |
+
[default1]:07/03/2024 23:26:47 [INFO|DP=0|PP=0|TP=1|ip-26-0-169-86]: Local number of parameters: 173M (329.19MiB)
|
| 132 |
+
[default1]:07/03/2024 23:26:47 [INFO|DP=0|PP=0|TP=1|ip-26-0-169-86]: [After model building] Memory usage: 344.13MiB. Peak allocated: 346.16MiB Peak reserved: 348.00MiB
|
| 133 |
+
[default1]:07/03/2024 23:26:47 [INFO|DP=0|PP=0|TP=1|ip-26-0-169-86]: No checkpoint path provided.
|
| 134 |
+
[default0]:07/03/2024 23:26:47 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: Total number of parameters: 1.21G (2313.42MiB)
|
| 135 |
+
[default0]:07/03/2024 23:26:47 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: Local number of parameters: 173M (329.19MiB)
|
| 136 |
+
[default0]:07/03/2024 23:26:47 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: [After model building] Memory usage: 344.13MiB. Peak allocated: 346.16MiB Peak reserved: 348.00MiB
|
| 137 |
+
[default0]:07/03/2024 23:26:47 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: No checkpoint path provided.
|
| 138 |
+
[default0]:07/03/2024 23:26:47 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: Parametrizing model parameters using StandardParametrizator
|
| 139 |
+
[default7]:07/03/2024 23:26:47 [INFO|DP=0|PP=1|TP=3|ip-26-0-169-86]: Local number of parameters: 131M (249.16MiB)
|
| 140 |
+
[default7]:07/03/2024 23:26:47 [INFO|DP=0|PP=1|TP=3|ip-26-0-169-86]: [After model building] Memory usage: 260.10MiB. Peak allocated: 262.13MiB Peak reserved: 264.00MiB
|
| 141 |
+
[default7]:07/03/2024 23:26:47 [INFO|DP=0|PP=1|TP=3|ip-26-0-169-86]: No checkpoint path provided.
|
| 142 |
+
[default2]:07/03/2024 23:26:47 [INFO|DP=0|PP=0|TP=2|ip-26-0-169-86]: Local number of parameters: 173M (329.19MiB)
|
| 143 |
+
[default2]:07/03/2024 23:26:47 [INFO|DP=0|PP=0|TP=2|ip-26-0-169-86]: [After model building] Memory usage: 344.13MiB. Peak allocated: 346.16MiB Peak reserved: 348.00MiB
|
| 144 |
+
[default2]:07/03/2024 23:26:47 [INFO|DP=0|PP=0|TP=2|ip-26-0-169-86]: No checkpoint path provided.
|
| 145 |
+
[default4]:07/03/2024 23:26:47 [INFO|DP=0|PP=1|TP=0|ip-26-0-169-86]: Local number of parameters: 131M (249.16MiB)
|
| 146 |
+
[default4]:07/03/2024 23:26:47 [INFO|DP=0|PP=1|TP=0|ip-26-0-169-86]: [After model building] Memory usage: 260.10MiB. Peak allocated: 262.13MiB Peak reserved: 264.00MiB
|
| 147 |
+
[default4]:07/03/2024 23:26:47 [INFO|DP=0|PP=1|TP=0|ip-26-0-169-86]: No checkpoint path provided.
|
| 148 |
+
[default6]:07/03/2024 23:26:47 [INFO|DP=0|PP=1|TP=2|ip-26-0-169-86]: Local number of parameters: 131M (249.16MiB)
|
| 149 |
+
[default6]:07/03/2024 23:26:47 [INFO|DP=0|PP=1|TP=2|ip-26-0-169-86]: [After model building] Memory usage: 260.10MiB. Peak allocated: 262.13MiB Peak reserved: 264.00MiB
|
| 150 |
+
[default6]:07/03/2024 23:26:47 [INFO|DP=0|PP=1|TP=2|ip-26-0-169-86]: No checkpoint path provided.
|
| 151 |
+
[default3]:07/03/2024 23:26:47 [INFO|DP=0|PP=0|TP=3|ip-26-0-169-86]: Local number of parameters: 173M (329.19MiB)
|
| 152 |
+
[default3]:07/03/2024 23:26:47 [INFO|DP=0|PP=0|TP=3|ip-26-0-169-86]: [After model building] Memory usage: 344.13MiB. Peak allocated: 346.16MiB Peak reserved: 348.00MiB
|
| 153 |
+
[default3]:07/03/2024 23:26:47 [INFO|DP=0|PP=0|TP=3|ip-26-0-169-86]: No checkpoint path provided.
|
| 154 |
+
[default0]:07/03/2024 23:26:47 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: [Optimizer Building] Using LearningRateForSP as learning rate
|
| 155 |
+
[default0]:07/03/2024 23:26:47 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: [ZeRO sharding] Size of optimizer params per rank:
|
| 156 |
+
[default0]:07/03/2024 23:26:47 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: [ZeRO sharding] DP Rank 0 has 173M out of 173M (100.00%) params' optimizer states
|
| 157 |
+
[default0]:07/03/2024 23:26:49 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: [Training Plan] Stage Training Stage has 19 remaining training steps and has consumed 0 samples
|
| 158 |
+
[default0]:07/03/2024 23:26:49 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: Using `datasets` library
|
| 159 |
+
[default0]:07/03/2024 23:26:49 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: Loading tokenizer from openai-community/gpt2 and transformers/hf_hub versions ('4.41.2', '0.23.4')
|
| 160 |
+
[default0]:07/03/2024 23:26:49 [WARNING|DP=0|PP=0|TP=0|ip-26-0-169-86]: Repo card metadata block was not found. Setting CardData to empty.
|
| 161 |
+
[default0]:Repo card metadata block was not found. Setting CardData to empty.
|
| 162 |
+
[default0]:07/03/2024 23:26:52 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: [Training Plan] There are 1 training stages
|
| 163 |
+
[default0]:07/03/2024 23:26:52 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: [Stage Training Stage] start from step 1
|
| 164 |
+
[default0]:07/03/2024 23:26:52 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]:
|
| 165 |
+
[default0]:07/03/2024 23:26:52 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: [Start training] datetime: 2024-07-03 23:26:52.108515 | mbs: 4 | grad_accum: 256 | global_batch_size: 1024 | sequence_length: 4096 | train_steps: 20 | start_iteration_step: 0 | consumed_train_samples: 0
|
| 166 |
+
[default0]:07/03/2024 23:26:52 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: Resuming training from stage Training Stage, it has trained for 0 samples and has 19 remaining train steps
|
| 167 |
+
[default0]:07/03/2024 23:26:52 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: Memory usage: 1660.89MiB. Peak allocated 1660.89MiB. Peak reserved: 1668.00MiB
|
| 168 |
+
[default1]:07/03/2024 23:26:52 [WARNING|DP=0|PP=0|TP=1|ip-26-0-169-86]: Repo card metadata block was not found. Setting CardData to empty.
|
| 169 |
+
[default4]:Repo card metadata block was not found. Setting CardData to empty.
|
| 170 |
+
[default7]:07/03/2024 23:26:52 [WARNING|DP=0|PP=1|TP=3|ip-26-0-169-86]: Repo card metadata block was not found. Setting CardData to empty.
|
| 171 |
+
[default2]:07/03/2024 23:26:52 [WARNING|DP=0|PP=0|TP=2|ip-26-0-169-86]: Repo card metadata block was not found. Setting CardData to empty.
|
| 172 |
+
[default6]:07/03/2024 23:26:52 [WARNING|DP=0|PP=1|TP=2|ip-26-0-169-86]: Repo card metadata block was not found. Setting CardData to empty.
|
| 173 |
+
[default4]:07/03/2024 23:26:52 [WARNING|DP=0|PP=1|TP=0|ip-26-0-169-86]: Repo card metadata block was not found. Setting CardData to empty.
|
| 174 |
+
[default3]:07/03/2024 23:26:52 [WARNING|DP=0|PP=0|TP=3|ip-26-0-169-86]: Repo card metadata block was not found. Setting CardData to empty.
|
| 175 |
+
[default7]:Repo card metadata block was not found. Setting CardData to empty.
|
| 176 |
+
[default2]:Repo card metadata block was not found. Setting CardData to empty.
|
| 177 |
+
[default3]:Repo card metadata block was not found. Setting CardData to empty.
|
| 178 |
+
[default6]: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 |
+
[default5]:07/03/2024 23:26:57 [WARNING|DP=0|PP=1|TP=1|ip-26-0-169-86]: Repo card metadata block was not found. Setting CardData to empty.
|
| 181 |
+
[default5]:Repo card metadata block was not found. Setting CardData to empty.
|
| 182 |
+
[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.)
|
| 183 |
+
[default4]: 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 |
+
[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.)
|
| 189 |
+
[default6]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
|
| 190 |
+
[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.)
|
| 191 |
+
[default0]: 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 |
+
[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.)
|
| 195 |
+
[default2]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
|
| 196 |
+
[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.)
|
| 197 |
+
[default1]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
|
| 198 |
+
[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
|
| 199 |
+
[default4]: warnings.warn(
|
| 200 |
+
[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
|
| 201 |
+
[default2]: warnings.warn(
|
| 202 |
+
[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
|
| 203 |
+
[default7]: warnings.warn(
|
| 204 |
+
[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
|
| 205 |
+
[default5]: warnings.warn(
|
| 206 |
+
[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
|
| 207 |
+
[default6]: warnings.warn(
|
| 208 |
+
[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
|
| 209 |
+
[default1]: warnings.warn(
|
| 210 |
+
[default0]:07/03/2024 23:27:36 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: Memory usage: 1732.03MiB. Peak allocated 13063.24MiB. Peak reserved: 13348.00MiB
|
| 211 |
+
[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
|
| 212 |
+
[default0]: warnings.warn(
|
| 213 |
+
[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
|
| 214 |
+
[default3]: warnings.warn(
|
| 215 |
+
[default0]:07/03/2024 23:27:44 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: Memory usage: 3048.84MiB. Peak allocated 3048.84MiB. Peak reserved: 13348.00MiB
|
| 216 |
+
[default4]:07/03/2024 23:27:44 [INFO|DP=0|PP=1|TP=0|ip-26-0-169-86]: iteration: 1 / 20 | consumed_tokens: 4.19M | elapsed_time_per_iteration_ms: 45.1K | tokens_per_sec: 92.9K | tokens_per_sec_per_gpu: 11.6K | global_batch_size: 1.02K | lm_loss: 11.1 | lr: 0.0001 | model_tflops_per_gpu: 105 | hardware_tflops_per_gpu: 105 | grad_norm: 15 | cuda_memory_allocated: 2.44G | cuda_max_memory_reserved: 8.57G | hd_total_memory_tb: 312G | hd_used_memory_tb: 68.8G | hd_free_memory_tb: 243G
|
| 217 |
+
[default0]:07/03/2024 23:28:07 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: Memory usage: 3048.84MiB. Peak allocated 14380.05MiB. Peak reserved: 14534.00MiB
|
| 218 |
+
[default0]:07/03/2024 23:28:07 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: Memory usage: 3048.84MiB. Peak allocated 3048.88MiB. Peak reserved: 14534.00MiB
|
| 219 |
+
[default4]:07/03/2024 23:28:07 [INFO|DP=0|PP=1|TP=0|ip-26-0-169-86]: iteration: 2 / 20 | consumed_tokens: 8.39M | elapsed_time_per_iteration_ms: 23.1K | tokens_per_sec: 181K | tokens_per_sec_per_gpu: 22.7K | global_batch_size: 1.02K | lm_loss: 11.1 | lr: 9.53e-05 | model_tflops_per_gpu: 206 | hardware_tflops_per_gpu: 206 | grad_norm: 15.1 | cuda_memory_allocated: 2.44G | cuda_max_memory_reserved: 8.57G | hd_total_memory_tb: 312G | hd_used_memory_tb: 68.8G | hd_free_memory_tb: 243G
|
| 220 |
+
[default4]:07/03/2024 23:28:31 [INFO|DP=0|PP=1|TP=0|ip-26-0-169-86]: iteration: 3 / 20 | consumed_tokens: 12.6M | elapsed_time_per_iteration_ms: 24.8K | tokens_per_sec: 169K | tokens_per_sec_per_gpu: 21.1K | global_batch_size: 1.02K | lm_loss: 11.4 | lr: 9.05e-05 | model_tflops_per_gpu: 192 | hardware_tflops_per_gpu: 192 | grad_norm: 106 | cuda_memory_allocated: 2.44G | cuda_max_memory_reserved: 8.57G | hd_total_memory_tb: 312G | hd_used_memory_tb: 68.8G | hd_free_memory_tb: 243G
|
| 221 |
+
[default0]:07/03/2024 23:28:31 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: Memory usage: 3048.84MiB. Peak allocated 14380.05MiB. Peak reserved: 14646.00MiB
|
| 222 |
+
[default0]:07/03/2024 23:28:31 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: Memory usage: 3048.84MiB. Peak allocated 3048.88MiB. Peak reserved: 14646.00MiB
|
| 223 |
+
[default0]:STAGE:2024-07-03 23:28:31 2274198:2274198 ActivityProfilerController.cpp:314] Completed Stage: Warm Up
|
| 224 |
+
[default4]:07/03/2024 23:29:04 [INFO|DP=0|PP=1|TP=0|ip-26-0-169-86]: iteration: 4 / 20 | consumed_tokens: 16.8M | elapsed_time_per_iteration_ms: 33K | tokens_per_sec: 127K | tokens_per_sec_per_gpu: 15.9K | global_batch_size: 1.02K | lm_loss: 11.7 | lr: 8.58e-05 | model_tflops_per_gpu: 144 | hardware_tflops_per_gpu: 144 | grad_norm: 24.5 | cuda_memory_allocated: 2.44G | cuda_max_memory_reserved: 8.57G | hd_total_memory_tb: 312G | hd_used_memory_tb: 68.8G | hd_free_memory_tb: 243G
|
| 225 |
+
[default0]:07/03/2024 23:29:04 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: Memory usage: 3048.84MiB. Peak allocated 14380.05MiB. Peak reserved: 14646.00MiB
|
| 226 |
+
[default0]:07/03/2024 23:29:04 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: Memory usage: 3048.84MiB. Peak allocated 3048.88MiB. Peak reserved: 14646.00MiB
|
| 227 |
+
[default4]:07/03/2024 23:29:38 [INFO|DP=0|PP=1|TP=0|ip-26-0-169-86]: iteration: 5 / 20 | consumed_tokens: 21M | elapsed_time_per_iteration_ms: 33.3K | tokens_per_sec: 126K | tokens_per_sec_per_gpu: 15.7K | global_batch_size: 1.02K | lm_loss: 10 | lr: 8.11e-05 | model_tflops_per_gpu: 143 | hardware_tflops_per_gpu: 143 | grad_norm: 11
|
| 228 |
+
[default0]:07/03/2024 23:29:38 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-86]: Memory usage: 3048.84MiB. Peak allocated 14380.05MiB. Peak reserved: 14646.00MiB
|
| 229 |
+
[default4]:07/03/2024 23:30:11 [INFO|DP=0|PP=1|TP=0|ip-26-0-169-86]: iteration: 6 / 20 | consumed_tokens: 25.2M | elapsed_time_per_iteration_ms: 33.1K | tokens_per_sec: 127K | tokens_per_sec_per_gpu: 15.9K | global_batch_size: 1.02K | lm_loss: 9.46 | lr: 7.63e-05 | model_tflops_per_gpu: 144 | hardware_tflops_per_gpu: 144 | grad_norm: 7.2
|
| 230 |
+
[default0]:STAGE:2024-07-03 23:31:36 2274198:2274198 ActivityProfilerController.cpp:320] Completed Stage: Collection
|
| 231 |
+
[default0]:STAGE:2024-07-03 23:31:45 2274198:2274198 ActivityProfilerController.cpp:324] Completed Stage: Post Processing
|
| 232 |
+
[default1]:[rank1]:[E ProcessGroupNCCL.cpp:563] [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=175141, OpType=_REDUCE_SCATTER_BASE, NumelIn=33554432, NumelOut=8388608, Timeout(ms)=600000) ran for 600016 milliseconds before timing out.
|
| 233 |
+
[default6]:[rank6]:[E ProcessGroupNCCL.cpp:563] [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=13834, OpType=RECV, NumelIn=7, NumelOut=7, Timeout(ms)=600000) ran for 600052 milliseconds before timing out.
|
| 234 |
+
[default2]:[rank2]:[E ProcessGroupNCCL.cpp:563] [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=175141, OpType=_REDUCE_SCATTER_BASE, NumelIn=33554432, NumelOut=8388608, Timeout(ms)=600000) ran for 600054 milliseconds before timing out.
|
| 235 |
+
[default7]:[rank7]:[E ProcessGroupNCCL.cpp:563] [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=13834, OpType=RECV, NumelIn=7, NumelOut=7, Timeout(ms)=600000) ran for 600052 milliseconds before timing out.
|
| 236 |
+
[default3]:[rank3]:[E ProcessGroupNCCL.cpp:563] [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=175141, OpType=_REDUCE_SCATTER_BASE, NumelIn=33554432, NumelOut=8388608, Timeout(ms)=600000) ran for 600092 milliseconds before timing out.
|
| 237 |
+
[default5]:[rank5]:[E ProcessGroupNCCL.cpp:563] [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=13834, OpType=RECV, NumelIn=7, NumelOut=7, Timeout(ms)=600000) ran for 600055 milliseconds before timing out.
|
| 238 |
+
[default4]:[rank4]:[E ProcessGroupNCCL.cpp:563] [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=13834, OpType=RECV, NumelIn=7, NumelOut=7, Timeout(ms)=600000) ran for 600065 milliseconds before timing out.
|
| 239 |
+
[default6]:[rank6]: Traceback (most recent call last):
|
| 240 |
+
[default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in <module>
|
| 241 |
+
[default6]:[rank6]: trainer.train(dataloader)
|
| 242 |
+
[default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train
|
| 243 |
+
[default6]:[rank6]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader)
|
| 244 |
+
[default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step
|
| 245 |
+
[default6]:[rank6]: outputs = self.pipeline_engine.train_batch_iter(
|
| 246 |
+
[default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 278, in train_batch_iter
|
| 247 |
+
[default6]:[rank6]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model)
|
| 248 |
+
[default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward
|
| 249 |
+
[default6]:[rank6]: output = model(**micro_batch)
|
| 250 |
+
[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
|
| 251 |
+
[default6]:[rank6]: return self._call_impl(*args, **kwargs)
|
| 252 |
+
[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
|
| 253 |
+
[default6]:[rank6]: return forward_call(*args, **kwargs)
|
| 254 |
+
[default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward
|
| 255 |
+
[default6]:[rank6]: sharded_logits = self.model(
|
| 256 |
+
[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
|
| 257 |
+
[default6]:[rank6]: return self._call_impl(*args, **kwargs)
|
| 258 |
+
[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
|
| 259 |
+
[default6]:[rank6]: return forward_call(*args, **kwargs)
|
| 260 |
+
[default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward
|
| 261 |
+
[default6]:[rank6]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0]
|
| 262 |
+
[default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states
|
| 263 |
+
[default6]:[rank6]: hidden_encoder_states = encoder_block(**hidden_encoder_states)
|
| 264 |
+
[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
|
| 265 |
+
[default6]:[rank6]: return self._call_impl(*args, **kwargs)
|
| 266 |
+
[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
|
| 267 |
+
[default6]:[rank6]: return forward_call(*args, **kwargs)
|
| 268 |
+
[default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 126, in forward
|
| 269 |
+
[default6]:[rank6]: new_kwargs[name] = recv_from_pipeline_state_buffer(
|
| 270 |
+
[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
|
| 271 |
+
[default6]:[rank6]: pipeline_state.run_communication()
|
| 272 |
+
[default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 150, in run_communication
|
| 273 |
+
[default6]:[rank6]: recv_activation_tensor = recv_activation()
|
| 274 |
+
[default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 31, in __call__
|
| 275 |
+
[default6]:[rank6]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0]
|
| 276 |
+
[default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors
|
| 277 |
+
[default6]:[rank6]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag)
|
| 278 |
+
[default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors
|
| 279 |
+
[default6]:[rank6]: meta = self._recv_meta(from_rank=from_rank, tag=tag)
|
| 280 |
+
[default6]:[rank6]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 267, in _recv_meta
|
| 281 |
+
[default6]:[rank6]: self.second_metadata = torch.empty(second_metadata_num_elements, dtype=torch.long, device=self.device)
|
| 282 |
+
[default6]:[rank6]: torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate more than 1EB memory.
|
| 283 |
+
[default7]:[rank7]: Traceback (most recent call last):
|
| 284 |
+
[default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in <module>
|
| 285 |
+
[default7]:[rank7]: trainer.train(dataloader)
|
| 286 |
+
[default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train
|
| 287 |
+
[default7]:[rank7]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader)
|
| 288 |
+
[default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step
|
| 289 |
+
[default7]:[rank7]: outputs = self.pipeline_engine.train_batch_iter(
|
| 290 |
+
[default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 278, in train_batch_iter
|
| 291 |
+
[default7]:[rank7]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model)
|
| 292 |
+
[default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward
|
| 293 |
+
[default7]:[rank7]: output = model(**micro_batch)
|
| 294 |
+
[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
|
| 295 |
+
[default7]:[rank7]: return self._call_impl(*args, **kwargs)
|
| 296 |
+
[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
|
| 297 |
+
[default7]:[rank7]: return forward_call(*args, **kwargs)
|
| 298 |
+
[default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward
|
| 299 |
+
[default7]:[rank7]: sharded_logits = self.model(
|
| 300 |
+
[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
|
| 301 |
+
[default7]:[rank7]: return self._call_impl(*args, **kwargs)
|
| 302 |
+
[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
|
| 303 |
+
[default7]:[rank7]: return forward_call(*args, **kwargs)
|
| 304 |
+
[default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward
|
| 305 |
+
[default7]:[rank7]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0]
|
| 306 |
+
[default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states
|
| 307 |
+
[default7]:[rank7]: hidden_encoder_states = encoder_block(**hidden_encoder_states)
|
| 308 |
+
[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
|
| 309 |
+
[default7]:[rank7]: return self._call_impl(*args, **kwargs)
|
| 310 |
+
[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
|
| 311 |
+
[default7]:[rank7]: return forward_call(*args, **kwargs)
|
| 312 |
+
[default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 126, in forward
|
| 313 |
+
[default7]:[rank7]: new_kwargs[name] = recv_from_pipeline_state_buffer(
|
| 314 |
+
[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
|
| 315 |
+
[default7]:[rank7]: pipeline_state.run_communication()
|
| 316 |
+
[default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 150, in run_communication
|
| 317 |
+
[default7]:[rank7]: recv_activation_tensor = recv_activation()
|
| 318 |
+
[default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 31, in __call__
|
| 319 |
+
[default7]:[rank7]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0]
|
| 320 |
+
[default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors
|
| 321 |
+
[default7]:[rank7]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag)
|
| 322 |
+
[default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors
|
| 323 |
+
[default7]:[rank7]: meta = self._recv_meta(from_rank=from_rank, tag=tag)
|
| 324 |
+
[default7]:[rank7]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 267, in _recv_meta
|
| 325 |
+
[default7]:[rank7]: self.second_metadata = torch.empty(second_metadata_num_elements, dtype=torch.long, device=self.device)
|
| 326 |
+
[default7]:[rank7]: torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate more than 1EB memory.
|
| 327 |
+
[default5]:[rank5]: Traceback (most recent call last):
|
| 328 |
+
[default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in <module>
|
| 329 |
+
[default5]:[rank5]: trainer.train(dataloader)
|
| 330 |
+
[default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train
|
| 331 |
+
[default5]:[rank5]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader)
|
| 332 |
+
[default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step
|
| 333 |
+
[default5]:[rank5]: outputs = self.pipeline_engine.train_batch_iter(
|
| 334 |
+
[default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 278, in train_batch_iter
|
| 335 |
+
[default5]:[rank5]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model)
|
| 336 |
+
[default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward
|
| 337 |
+
[default5]:[rank5]: output = model(**micro_batch)
|
| 338 |
+
[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
|
| 339 |
+
[default5]:[rank5]: return self._call_impl(*args, **kwargs)
|
| 340 |
+
[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
|
| 341 |
+
[default5]:[rank5]: return forward_call(*args, **kwargs)
|
| 342 |
+
[default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward
|
| 343 |
+
[default5]:[rank5]: sharded_logits = self.model(
|
| 344 |
+
[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
|
| 345 |
+
[default5]:[rank5]: return self._call_impl(*args, **kwargs)
|
| 346 |
+
[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
|
| 347 |
+
[default5]:[rank5]: return forward_call(*args, **kwargs)
|
| 348 |
+
[default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward
|
| 349 |
+
[default5]:[rank5]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0]
|
| 350 |
+
[default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states
|
| 351 |
+
[default5]:[rank5]: hidden_encoder_states = encoder_block(**hidden_encoder_states)
|
| 352 |
+
[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
|
| 353 |
+
[default5]:[rank5]: return self._call_impl(*args, **kwargs)
|
| 354 |
+
[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
|
| 355 |
+
[default5]:[rank5]: return forward_call(*args, **kwargs)
|
| 356 |
+
[default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 126, in forward
|
| 357 |
+
[default5]:[rank5]: new_kwargs[name] = recv_from_pipeline_state_buffer(
|
| 358 |
+
[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
|
| 359 |
+
[default5]:[rank5]: pipeline_state.run_communication()
|
| 360 |
+
[default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 150, in run_communication
|
| 361 |
+
[default5]:[rank5]: recv_activation_tensor = recv_activation()
|
| 362 |
+
[default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 31, in __call__
|
| 363 |
+
[default5]:[rank5]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0]
|
| 364 |
+
[default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors
|
| 365 |
+
[default5]:[rank5]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag)
|
| 366 |
+
[default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors
|
| 367 |
+
[default5]:[rank5]: meta = self._recv_meta(from_rank=from_rank, tag=tag)
|
| 368 |
+
[default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 267, in _recv_meta
|
| 369 |
+
[default5]:[rank5]: self.second_metadata = torch.empty(second_metadata_num_elements, dtype=torch.long, device=self.device)
|
| 370 |
+
[default5]:[rank5]: torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate more than 1EB memory.
|
| 371 |
+
[default4]:[rank4]: Traceback (most recent call last):
|
| 372 |
+
[default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in <module>
|
| 373 |
+
[default4]:[rank4]: trainer.train(dataloader)
|
| 374 |
+
[default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train
|
| 375 |
+
[default4]:[rank4]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader)
|
| 376 |
+
[default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step
|
| 377 |
+
[default4]:[rank4]: outputs = self.pipeline_engine.train_batch_iter(
|
| 378 |
+
[default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 278, in train_batch_iter
|
| 379 |
+
[default4]:[rank4]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model)
|
| 380 |
+
[default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward
|
| 381 |
+
[default4]:[rank4]: output = model(**micro_batch)
|
| 382 |
+
[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
|
| 383 |
+
[default4]:[rank4]: return self._call_impl(*args, **kwargs)
|
| 384 |
+
[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
|
| 385 |
+
[default4]:[rank4]: return forward_call(*args, **kwargs)
|
| 386 |
+
[default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward
|
| 387 |
+
[default4]:[rank4]: sharded_logits = self.model(
|
| 388 |
+
[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
|
| 389 |
+
[default4]:[rank4]: return self._call_impl(*args, **kwargs)
|
| 390 |
+
[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
|
| 391 |
+
[default4]:[rank4]: return forward_call(*args, **kwargs)
|
| 392 |
+
[default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward
|
| 393 |
+
[default4]:[rank4]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0]
|
| 394 |
+
[default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states
|
| 395 |
+
[default4]:[rank4]: hidden_encoder_states = encoder_block(**hidden_encoder_states)
|
| 396 |
+
[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
|
| 397 |
+
[default4]:[rank4]: return self._call_impl(*args, **kwargs)
|
| 398 |
+
[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
|
| 399 |
+
[default4]:[rank4]: return forward_call(*args, **kwargs)
|
| 400 |
+
[default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 126, in forward
|
| 401 |
+
[default4]:[rank4]: new_kwargs[name] = recv_from_pipeline_state_buffer(
|
| 402 |
+
[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
|
| 403 |
+
[default4]:[rank4]: pipeline_state.run_communication()
|
| 404 |
+
[default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 150, in run_communication
|
| 405 |
+
[default4]:[rank4]: recv_activation_tensor = recv_activation()
|
| 406 |
+
[default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 31, in __call__
|
| 407 |
+
[default4]:[rank4]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0]
|
| 408 |
+
[default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors
|
| 409 |
+
[default4]:[rank4]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag)
|
| 410 |
+
[default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors
|
| 411 |
+
[default4]:[rank4]: meta = self._recv_meta(from_rank=from_rank, tag=tag)
|
| 412 |
+
[default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 267, in _recv_meta
|
| 413 |
+
[default4]:[rank4]: self.second_metadata = torch.empty(second_metadata_num_elements, dtype=torch.long, device=self.device)
|
| 414 |
+
[default4]:[rank4]: torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate more than 1EB memory.
|
| 415 |
+
[default6]:[rank6]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 1] Timeout at NCCL work: 13834, last enqueued NCCL work: 13834, last completed NCCL work: 13833.
|
| 416 |
+
[default6]:[rank6]:[E ProcessGroupNCCL.cpp:577] [Rank 1] 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.
|
| 417 |
+
[default6]:[rank6]:[E ProcessGroupNCCL.cpp:583] [Rank 1] To avoid data inconsistency, we are taking the entire process down.
|
| 418 |
+
[default6]:[rank6]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=13834, OpType=RECV, NumelIn=7, NumelOut=7, Timeout(ms)=600000) ran for 600052 milliseconds before timing out.
|
| 419 |
+
[default6]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first):
|
| 420 |
+
[default6]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7ffa8926d897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so)
|
| 421 |
+
[default6]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional<std::chrono::duration<long, std::ratio<1l, 1000l> > >) + 0x1d2 (0x7ffa8a546c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
|
| 422 |
+
[default6]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7ffa8a54ba80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
|
| 423 |
+
[default6]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7ffa8a54cdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
|
| 424 |
+
[default6]:frame #4: <unknown function> + 0xd3e95 (0x7ffad5fe5e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6)
|
| 425 |
+
[default6]:frame #5: <unknown function> + 0x8609 (0x7ffadb02c609 in /lib/x86_64-linux-gnu/libpthread.so.0)
|
| 426 |
+
[default6]:frame #6: clone + 0x43 (0x7ffadadf7353 in /lib/x86_64-linux-gnu/libc.so.6)
|
| 427 |
+
[default6]:
|
| 428 |
+
[default6]:terminate called after throwing an instance of 'c10::DistBackendError'
|
| 429 |
+
[default6]: what(): [PG 4 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=13834, OpType=RECV, NumelIn=7, NumelOut=7, Timeout(ms)=600000) ran for 600052 milliseconds before timing out.
|
| 430 |
+
[default6]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first):
|
| 431 |
+
[default6]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7ffa8926d897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so)
|
| 432 |
+
[default6]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional<std::chrono::duration<long, std::ratio<1l, 1000l> > >) + 0x1d2 (0x7ffa8a546c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
|
| 433 |
+
[default6]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7ffa8a54ba80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
|
| 434 |
+
[default6]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7ffa8a54cdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
|
| 435 |
+
[default6]:frame #4: <unknown function> + 0xd3e95 (0x7ffad5fe5e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6)
|
| 436 |
+
[default6]:frame #5: <unknown function> + 0x8609 (0x7ffadb02c609 in /lib/x86_64-linux-gnu/libpthread.so.0)
|
| 437 |
+
[default6]:frame #6: clone + 0x43 (0x7ffadadf7353 in /lib/x86_64-linux-gnu/libc.so.6)
|
| 438 |
+
[default6]:
|
| 439 |
+
[default6]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first):
|
| 440 |
+
[default6]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7ffa8926d897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so)
|
| 441 |
+
[default6]:frame #1: <unknown function> + 0xe32119 (0x7ffa8a1d0119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
|
| 442 |
+
[default6]:frame #2: <unknown function> + 0xd3e95 (0x7ffad5fe5e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6)
|
| 443 |
+
[default6]:frame #3: <unknown function> + 0x8609 (0x7ffadb02c609 in /lib/x86_64-linux-gnu/libpthread.so.0)
|
| 444 |
+
[default6]:frame #4: clone + 0x43 (0x7ffadadf7353 in /lib/x86_64-linux-gnu/libc.so.6)
|
| 445 |
+
[default6]:
|
| 446 |
+
[default7]:[rank7]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 1] Timeout at NCCL work: 13834, last enqueued NCCL work: 13834, last completed NCCL work: 13833.
|
| 447 |
+
[default7]:[rank7]:[E ProcessGroupNCCL.cpp:577] [Rank 1] 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.
|
| 448 |
+
[default7]:[rank7]:[E ProcessGroupNCCL.cpp:583] [Rank 1] To avoid data inconsistency, we are taking the entire process down.
|
| 449 |
+
[default7]:[rank7]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=13834, OpType=RECV, NumelIn=7, NumelOut=7, Timeout(ms)=600000) ran for 600052 milliseconds before timing out.
|
| 450 |
+
[default7]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first):
|
| 451 |
+
[default7]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f08104fd897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so)
|
| 452 |
+
[default7]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional<std::chrono::duration<long, std::ratio<1l, 1000l> > >) + 0x1d2 (0x7f08117d6c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
|
| 453 |
+
[default7]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f08117dba80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
|
| 454 |
+
[default7]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f08117dcdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
|
| 455 |
+
[default7]:frame #4: <unknown function> + 0xd3e95 (0x7f085d275e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6)
|
| 456 |
+
[default7]:frame #5: <unknown function> + 0x8609 (0x7f08622bc609 in /lib/x86_64-linux-gnu/libpthread.so.0)
|
| 457 |
+
[default7]:frame #6: clone + 0x43 (0x7f0862087353 in /lib/x86_64-linux-gnu/libc.so.6)
|
| 458 |
+
[default7]:
|
| 459 |
+
[default7]:terminate called after throwing an instance of 'c10::DistBackendError'
|
| 460 |
+
[default7]: what(): [PG 4 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=13834, OpType=RECV, NumelIn=7, NumelOut=7, Timeout(ms)=600000) ran for 600052 milliseconds before timing out.
|
| 461 |
+
[default7]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first):
|
| 462 |
+
[default7]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f08104fd897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so)
|
| 463 |
+
[default7]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional<std::chrono::duration<long, std::ratio<1l, 1000l> > >) + 0x1d2 (0x7f08117d6c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
|
| 464 |
+
[default7]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f08117dba80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
|
| 465 |
+
[default7]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f08117dcdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
|
| 466 |
+
[default7]:frame #4: <unknown function> + 0xd3e95 (0x7f085d275e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6)
|
| 467 |
+
[default7]:frame #5: <unknown function> + 0x8609 (0x7f08622bc609 in /lib/x86_64-linux-gnu/libpthread.so.0)
|
| 468 |
+
[default7]:frame #6: clone + 0x43 (0x7f0862087353 in /lib/x86_64-linux-gnu/libc.so.6)
|
| 469 |
+
[default7]:
|
| 470 |
+
[default7]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first):
|
| 471 |
+
[default7]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f08104fd897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so)
|
| 472 |
+
[default7]:frame #1: <unknown function> + 0xe32119 (0x7f0811460119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
|
| 473 |
+
[default7]:frame #2: <unknown function> + 0xd3e95 (0x7f085d275e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6)
|
| 474 |
+
[default7]:frame #3: <unknown function> + 0x8609 (0x7f08622bc609 in /lib/x86_64-linux-gnu/libpthread.so.0)
|
| 475 |
+
[default7]:frame #4: clone + 0x43 (0x7f0862087353 in /lib/x86_64-linux-gnu/libc.so.6)
|
| 476 |
+
[default7]:
|
| 477 |
+
[default5]:[rank5]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 1] Timeout at NCCL work: 13834, last enqueued NCCL work: 13834, last completed NCCL work: 13833.
|
| 478 |
+
[default5]:[rank5]:[E ProcessGroupNCCL.cpp:577] [Rank 1] 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.
|
| 479 |
+
[default5]:[rank5]:[E ProcessGroupNCCL.cpp:583] [Rank 1] To avoid data inconsistency, we are taking the entire process down.
|
| 480 |
+
[default5]:[rank5]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=13834, OpType=RECV, NumelIn=7, NumelOut=7, Timeout(ms)=600000) ran for 600055 milliseconds before timing out.
|
| 481 |
+
[default5]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first):
|
| 482 |
+
[default5]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f80e58fb897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so)
|
| 483 |
+
[default5]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional<std::chrono::duration<long, std::ratio<1l, 1000l> > >) + 0x1d2 (0x7f80e6bd4c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
|
| 484 |
+
[default5]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f80e6bd9a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
|
| 485 |
+
[default5]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f80e6bdadcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
|
| 486 |
+
[default5]:frame #4: <unknown function> + 0xd3e95 (0x7f8132673e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6)
|
| 487 |
+
[default5]:frame #5: <unknown function> + 0x8609 (0x7f81376ba609 in /lib/x86_64-linux-gnu/libpthread.so.0)
|
| 488 |
+
[default5]:frame #6: clone + 0x43 (0x7f8137485353 in /lib/x86_64-linux-gnu/libc.so.6)
|
| 489 |
+
[default5]:
|
| 490 |
+
[default5]:terminate called after throwing an instance of 'c10::DistBackendError'
|
| 491 |
+
[default5]: what(): [PG 4 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=13834, OpType=RECV, NumelIn=7, NumelOut=7, Timeout(ms)=600000) ran for 600055 milliseconds before timing out.
|
| 492 |
+
[default5]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first):
|
| 493 |
+
[default5]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f80e58fb897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so)
|
| 494 |
+
[default5]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional<std::chrono::duration<long, std::ratio<1l, 1000l> > >) + 0x1d2 (0x7f80e6bd4c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
|
| 495 |
+
[default5]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f80e6bd9a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
|
| 496 |
+
[default5]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f80e6bdadcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
|
| 497 |
+
[default5]:frame #4: <unknown function> + 0xd3e95 (0x7f8132673e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6)
|
| 498 |
+
[default5]:frame #5: <unknown function> + 0x8609 (0x7f81376ba609 in /lib/x86_64-linux-gnu/libpthread.so.0)
|
| 499 |
+
[default5]:frame #6: clone + 0x43 (0x7f8137485353 in /lib/x86_64-linux-gnu/libc.so.6)
|
| 500 |
+
[default5]:
|
| 501 |
+
[default5]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first):
|
| 502 |
+
[default5]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f80e58fb897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so)
|
| 503 |
+
[default5]:frame #1: <unknown function> + 0xe32119 (0x7f80e685e119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
|
| 504 |
+
[default5]:frame #2: <unknown function> + 0xd3e95 (0x7f8132673e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6)
|
| 505 |
+
[default5]:frame #3: <unknown function> + 0x8609 (0x7f81376ba609 in /lib/x86_64-linux-gnu/libpthread.so.0)
|
| 506 |
+
[default5]:frame #4: clone + 0x43 (0x7f8137485353 in /lib/x86_64-linux-gnu/libc.so.6)
|
| 507 |
+
[default5]:
|
| 508 |
+
[default4]:[rank4]:[E ProcessGroupNCCL.cpp:1537] [PG 4 Rank 1] Timeout at NCCL work: 13834, last enqueued NCCL work: 13834, last completed NCCL work: 13833.
|
| 509 |
+
[default4]:[rank4]:[E ProcessGroupNCCL.cpp:577] [Rank 1] 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.
|
| 510 |
+
[default4]:[rank4]:[E ProcessGroupNCCL.cpp:583] [Rank 1] To avoid data inconsistency, we are taking the entire process down.
|
| 511 |
+
[default4]:[rank4]:[E ProcessGroupNCCL.cpp:1414] [PG 4 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=13834, OpType=RECV, NumelIn=7, NumelOut=7, Timeout(ms)=600000) ran for 600065 milliseconds before timing out.
|
| 512 |
+
[default4]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first):
|
| 513 |
+
[default4]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f4c6224c897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so)
|
| 514 |
+
[default4]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional<std::chrono::duration<long, std::ratio<1l, 1000l> > >) + 0x1d2 (0x7f4c63525c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
|
| 515 |
+
[default4]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f4c6352aa80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
|
| 516 |
+
[default4]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f4c6352bdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
|
| 517 |
+
[default4]:frame #4: <unknown function> + 0xd3e95 (0x7f4caefc4e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6)
|
| 518 |
+
[default4]:frame #5: <unknown function> + 0x8609 (0x7f4cb400b609 in /lib/x86_64-linux-gnu/libpthread.so.0)
|
| 519 |
+
[default4]:frame #6: clone + 0x43 (0x7f4cb3dd6353 in /lib/x86_64-linux-gnu/libc.so.6)
|
| 520 |
+
[default4]:
|
| 521 |
+
[default4]:terminate called after throwing an instance of 'c10::DistBackendError'
|
| 522 |
+
[default4]: what(): [PG 4 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=13834, OpType=RECV, NumelIn=7, NumelOut=7, Timeout(ms)=600000) ran for 600065 milliseconds before timing out.
|
| 523 |
+
[default4]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first):
|
| 524 |
+
[default4]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f4c6224c897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so)
|
| 525 |
+
[default4]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional<std::chrono::duration<long, std::ratio<1l, 1000l> > >) + 0x1d2 (0x7f4c63525c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
|
| 526 |
+
[default4]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f4c6352aa80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
|
| 527 |
+
[default4]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f4c6352bdcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
|
| 528 |
+
[default4]:frame #4: <unknown function> + 0xd3e95 (0x7f4caefc4e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6)
|
| 529 |
+
[default4]:frame #5: <unknown function> + 0x8609 (0x7f4cb400b609 in /lib/x86_64-linux-gnu/libpthread.so.0)
|
| 530 |
+
[default4]:frame #6: clone + 0x43 (0x7f4cb3dd6353 in /lib/x86_64-linux-gnu/libc.so.6)
|
| 531 |
+
[default4]:
|
| 532 |
+
[default4]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first):
|
| 533 |
+
[default4]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f4c6224c897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so)
|
| 534 |
+
[default4]:frame #1: <unknown function> + 0xe32119 (0x7f4c631af119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
|
| 535 |
+
[default4]:frame #2: <unknown function> + 0xd3e95 (0x7f4caefc4e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6)
|
| 536 |
+
[default4]:frame #3: <unknown function> + 0x8609 (0x7f4cb400b609 in /lib/x86_64-linux-gnu/libpthread.so.0)
|
| 537 |
+
[default4]:frame #4: clone + 0x43 (0x7f4cb3dd6353 in /lib/x86_64-linux-gnu/libc.so.6)
|
| 538 |
+
[default4]:
|
| 539 |
+
W0703 23:40:16.055000 139973789022016 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 2274198 closing signal SIGTERM
|
| 540 |
+
W0703 23:40:16.055000 139973789022016 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 2274199 closing signal SIGTERM
|
| 541 |
+
W0703 23:40:16.055000 139973789022016 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 2274200 closing signal SIGTERM
|
| 542 |
+
W0703 23:40:16.055000 139973789022016 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 2274201 closing signal SIGTERM
|
| 543 |
+
E0703 23:40:20.862000 139973789022016 torch/distributed/elastic/multiprocessing/api.py:826] failed (exitcode: -6) local_rank: 4 (pid: 2274202) of binary: /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/python3.10
|
| 544 |
+
Traceback (most recent call last):
|
| 545 |
+
File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/torchrun", line 8, in <module>
|
| 546 |
+
sys.exit(main())
|
| 547 |
+
File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 347, in wrapper
|
| 548 |
+
return f(*args, **kwargs)
|
| 549 |
+
File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/run.py", line 879, in main
|
| 550 |
+
run(args)
|
| 551 |
+
File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/run.py", line 870, in run
|
| 552 |
+
elastic_launch(
|
| 553 |
+
File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 132, in __call__
|
| 554 |
+
return launch_agent(self._config, self._entrypoint, list(args))
|
| 555 |
+
File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 263, in launch_agent
|
| 556 |
+
raise ChildFailedError(
|
| 557 |
+
torch.distributed.elastic.multiprocessing.errors.ChildFailedError:
|
| 558 |
+
============================================================
|
| 559 |
+
/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py FAILED
|
| 560 |
+
------------------------------------------------------------
|
| 561 |
+
Failures:
|
| 562 |
+
[1]:
|
| 563 |
+
time : 2024-07-03_23:40:16
|
| 564 |
+
host : ip-26-0-169-86.ec2.internal
|
| 565 |
+
rank : 5 (local_rank: 5)
|
| 566 |
+
exitcode : -6 (pid: 2274203)
|
| 567 |
+
error_file: <N/A>
|
| 568 |
+
traceback : Signal 6 (SIGABRT) received by PID 2274203
|
| 569 |
+
[2]:
|
| 570 |
+
time : 2024-07-03_23:40:16
|
| 571 |
+
host : ip-26-0-169-86.ec2.internal
|
| 572 |
+
rank : 6 (local_rank: 6)
|
| 573 |
+
exitcode : -6 (pid: 2274204)
|
| 574 |
+
error_file: <N/A>
|
| 575 |
+
traceback : Signal 6 (SIGABRT) received by PID 2274204
|
| 576 |
+
[3]:
|
| 577 |
+
time : 2024-07-03_23:40:16
|
| 578 |
+
host : ip-26-0-169-86.ec2.internal
|
| 579 |
+
rank : 7 (local_rank: 7)
|
| 580 |
+
exitcode : -6 (pid: 2274205)
|
| 581 |
+
error_file: <N/A>
|
| 582 |
+
traceback : Signal 6 (SIGABRT) received by PID 2274205
|
| 583 |
+
------------------------------------------------------------
|
| 584 |
+
Root Cause (first observed failure):
|
| 585 |
+
[0]:
|
| 586 |
+
time : 2024-07-03_23:40:16
|
| 587 |
+
host : ip-26-0-169-86.ec2.internal
|
| 588 |
+
rank : 4 (local_rank: 4)
|
| 589 |
+
exitcode : -6 (pid: 2274202)
|
| 590 |
+
error_file: <N/A>
|
| 591 |
+
traceback : Signal 6 (SIGABRT) received by PID 2274202
|
| 592 |
+
============================================================
|
| 593 |
+
srun: error: ip-26-0-169-86: task 0: Exited with exit code 1
|
| 594 |
+
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.
|
| 595 |
+
|
llama-1B/8_GPUS/dp-1_tp-4_pp-2_mbz-4/profiler/ip-26-0-169-86_2274198.1720050009949646955.pt.trace.json.tmp
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:781117f0f97d49fb807e3801f4455014e1c21e39c28f6f79eb5423cb8f7b16e3
|
| 3 |
+
size 848259944
|
llama-1B/8_GPUS/dp-1_tp-4_pp-2_mbz-4/status.txt
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
oom
|