Hugging Face's logo

nanotron
/
old_bench

Model card Files Community
3outeille HF staff commited on
Commit
2348abf
·
verified ·
1 Parent(s): 45adef9

Upload llama-1B/8_GPUS/dp-4_tp-2_pp-1_mbz-1

Browse files
Files changed (4) hide show
  1. llama-1B/8_GPUS/dp-4_tp-2_pp-1_mbz-1/bench.slurm +111 -0
  2. llama-1B/8_GPUS/dp-4_tp-2_pp-1_mbz-1/config.yaml +90 -0
  3. llama-1B/8_GPUS/dp-4_tp-2_pp-1_mbz-1/log.out +305 -0
  4. llama-1B/8_GPUS/dp-4_tp-2_pp-1_mbz-1/status.txt +1 -0
llama-1B/8_GPUS/dp-4_tp-2_pp-1_mbz-1/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-4_tp-2_pp-1_mbz-1/log.out
13
+ #SBATCH --error=/fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-4_tp-2_pp-1_mbz-1/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-4_tp-2_pp-1_mbz-1/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-4_tp-2_pp-1_mbz-1/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-4_tp-2_pp-1_mbz-1/status.txt
84
+ else
85
+ if grep -q "OutOfMemoryError" /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-4_tp-2_pp-1_mbz-1/log.out; then
86
+ printf "oom" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-4_tp-2_pp-1_mbz-1/status.txt
87
+ elif grep -q " CUDA error: an illegal memory access" /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-4_tp-2_pp-1_mbz-1/log.out; then
88
+ printf "oom" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-4_tp-2_pp-1_mbz-1/status.txt
89
+ elif grep -q "Timeout at NCCL" /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-4_tp-2_pp-1_mbz-1/log.out; then
90
+ printf "timeout" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-4_tp-2_pp-1_mbz-1/status.txt
91
+ else
92
+ printf "fail" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-4_tp-2_pp-1_mbz-1/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-4_tp-2_pp-1_mbz-1 --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-4_tp-2_pp-1_mbz-1 --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-4_tp-2_pp-1_mbz-1 llama-1B/8_GPUS/dp-4_tp-2_pp-1_mbz-1 --commit-message "Upload llama-1B/8_GPUS/dp-4_tp-2_pp-1_mbz-1"
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-4_tp-2_pp-1_mbz-1/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: 4
49
+ expert_parallel_size: 1
50
+ pp: 1
51
+ pp_engine: 1f1b
52
+ tp: 2
53
+ tp_linear_async_communication: false
54
+ tp_mode: REDUCE_SCATTER
55
+ profiler:
56
+ profiler_export_path: /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-4_tp-2_pp-1_mbz-1
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: 1
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-4_tp-2_pp-1_mbz-1/log.out ADDED
@@ -0,0 +1,305 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ========================
2
+ START TIME: Wed Jul 3 22:51:41 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 22:51:44.289000 139636250072896 torch/distributed/run.py:757]
18
+ W0703 22:51:44.289000 139636250072896 torch/distributed/run.py:757] *****************************************
19
+ W0703 22:51:44.289000 139636250072896 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 22:51:44.289000 139636250072896 torch/distributed/run.py:757] *****************************************
21
+ [default0]:07/03/2024 22:52:00 [WARNING|DP=0|PP=0|TP=0|ip-26-0-160-225]: [Vocab Size Padding] Padded vocab (size: 50257) with 1 dummy tokens (new size: 50258)
22
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Config:
23
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Config(general=GeneralArgs(project='bench_cluster',
24
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: run='%date_%jobid',
25
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: seed=42,
26
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: step=None,
27
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: consumed_train_samples=None,
28
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: benchmark_csv_path=None,
29
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: ignore_sanity_checks=True),
30
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: parallelism=ParallelismArgs(dp=4,
31
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: pp=1,
32
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: tp=2,
33
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: pp_engine=<nanotron.parallel.pipeline_parallel.engine.OneForwardOneBackwardPipelineEngine object at 0x7fc6c3bdc700>,
34
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: tp_mode=<TensorParallelLinearMode.REDUCE_SCATTER: 2>,
35
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: tp_linear_async_communication=False,
36
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: expert_parallel_size=1),
37
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: model=ModelArgs(model_config=LlamaConfig(bos_token_id=1,
38
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: eos_token_id=2,
39
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: hidden_act='silu',
40
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: hidden_size=2048,
41
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: initializer_range=0.02,
42
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: intermediate_size=4096,
43
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: is_llama_config=True,
44
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: max_position_embeddings=4096,
45
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: num_attention_heads=32,
46
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: num_hidden_layers=24,
47
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: num_key_value_heads=32,
48
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: pad_token_id=None,
49
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: pretraining_tp=1,
50
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: rms_norm_eps=1e-05,
51
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: rope_scaling=None,
52
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: rope_theta=10000.0,
53
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: tie_word_embeddings=True,
54
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: use_cache=True,
55
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: vocab_size=50258),
56
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: init_method=RandomInit(std=0.025),
57
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: dtype=torch.bfloat16,
58
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: make_vocab_size_divisible_by=1,
59
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: ddp_bucket_cap_mb=25),
60
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: tokenizer=TokenizerArgs(tokenizer_name_or_path='openai-community/gpt2',
61
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: tokenizer_revision=None,
62
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: tokenizer_max_length=None),
63
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: checkpoints=CheckpointsArgs(checkpoints_path=Path('/dev/null'),
64
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: checkpoint_interval=100000,
65
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: save_initial_state=False,
66
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: resume_checkpoint_path=None,
67
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: checkpoints_path_is_shared_file_system=False),
68
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: logging=LoggingArgs(log_level='info',
69
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: log_level_replica='info',
70
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: iteration_step_info_interval=1),
71
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: tokens=TokensArgs(sequence_length=4096,
72
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: train_steps=20,
73
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: micro_batch_size=1,
74
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: batch_accumulation_per_replica=256,
75
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: val_check_interval=-1,
76
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: limit_val_batches=0,
77
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: limit_test_batches=0),
78
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: optimizer=OptimizerArgs(optimizer_factory=AdamWOptimizerArgs(adam_eps=1e-08,
79
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: adam_beta1=0.9,
80
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: adam_beta2=0.95,
81
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: torch_adam_is_fused=True,
82
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: name='adamW'),
83
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: zero_stage=1,
84
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: weight_decay=0.01,
85
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: clip_grad=1.0,
86
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: accumulate_grad_in_fp32=True,
87
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: learning_rate_scheduler=LRSchedulerArgs(learning_rate=0.0001,
88
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: lr_warmup_steps=1,
89
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: lr_warmup_style='linear',
90
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: lr_decay_style='linear',
91
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: lr_decay_steps=19,
92
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: lr_decay_starting_step=None,
93
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: min_decay_lr=1e-05)),
94
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: data_stages=[DatasetStageArgs(name='Training Stage',
95
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: start_training_step=1,
96
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: data=DataArgs(dataset=PretrainDatasetsArgs(hf_dataset_or_datasets='roneneldan/TinyStories',
97
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: hf_dataset_splits='train',
98
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: hf_dataset_config_name=None,
99
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: dataset_processing_num_proc_per_process=64,
100
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: dataset_overwrite_cache=False,
101
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: text_column_name='text'),
102
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: seed=42,
103
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: num_loading_workers=0))],
104
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: profiler=ProfilerArgs(profiler_export_path=Path('/fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-4_tp-2_pp-1_mbz-1')),
105
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: lighteval=None)
106
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Model Config:
107
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: LlamaConfig(bos_token_id=1,
108
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: eos_token_id=2,
109
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: hidden_act='silu',
110
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: hidden_size=2048,
111
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: initializer_range=0.02,
112
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: intermediate_size=4096,
113
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: is_llama_config=True,
114
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: max_position_embeddings=4096,
115
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: num_attention_heads=32,
116
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: num_hidden_layers=24,
117
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: num_key_value_heads=32,
118
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: pad_token_id=None,
119
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: pretraining_tp=1,
120
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: rms_norm_eps=1e-05,
121
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: rope_scaling=None,
122
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: rope_theta=10000.0,
123
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: tie_word_embeddings=True,
124
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: use_cache=True,
125
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: vocab_size=50258)
126
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Building model..
127
+ [default0]:07/03/2024 22:52:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Setting PP block ranks...
128
+ [default1]:07/03/2024 22:52:10 [INFO|DP=0|PP=0|TP=1|ip-26-0-160-225]: Local number of parameters: 555M (1058.35MiB)
129
+ [default1]:07/03/2024 22:52:10 [INFO|DP=0|PP=0|TP=1|ip-26-0-160-225]: [After model building] Memory usage: 1082.37MiB. Peak allocated: 1182.56MiB Peak reserved: 1200.00MiB
130
+ [default1]:07/03/2024 22:52:10 [INFO|DP=0|PP=0|TP=1|ip-26-0-160-225]: No checkpoint path provided.
131
+ [default0]:07/03/2024 22:52:10 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Total number of parameters: 1.11G (2116.70MiB)
132
+ [default0]:07/03/2024 22:52:10 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Local number of parameters: 555M (1058.35MiB)
133
+ [default0]:07/03/2024 22:52:10 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: [After model building] Memory usage: 1082.37MiB. Peak allocated: 1182.56MiB Peak reserved: 1200.00MiB
134
+ [default0]:07/03/2024 22:52:10 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: No checkpoint path provided.
135
+ [default0]:07/03/2024 22:52:10 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Parametrizing model parameters using StandardParametrizator
136
+ [default3]:07/03/2024 22:52:10 [INFO|DP=1|PP=0|TP=1|ip-26-0-160-225]: No checkpoint path provided.
137
+ [default2]:07/03/2024 22:52:10 [INFO|DP=1|PP=0|TP=0|ip-26-0-160-225]: No checkpoint path provided.
138
+ [default7]:07/03/2024 22:52:10 [INFO|DP=3|PP=0|TP=1|ip-26-0-160-225]: No checkpoint path provided.
139
+ [default6]:07/03/2024 22:52:10 [INFO|DP=3|PP=0|TP=0|ip-26-0-160-225]: No checkpoint path provided.
140
+ [default4]:07/03/2024 22:52:10 [INFO|DP=2|PP=0|TP=0|ip-26-0-160-225]: No checkpoint path provided.
141
+ [default5]:07/03/2024 22:52:10 [INFO|DP=2|PP=0|TP=1|ip-26-0-160-225]: No checkpoint path provided.
142
+ [default0]:07/03/2024 22:52:15 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: [Optimizer Building] Using LearningRateForSP as learning rate
143
+ [default0]:07/03/2024 22:52:15 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: [ZeRO sharding] Size of optimizer params per rank:
144
+ [default0]:07/03/2024 22:52:15 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: [ZeRO sharding] DP Rank 0 has 139M out of 555M (25.00%) params' optimizer states
145
+ [default0]:07/03/2024 22:52:15 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: [ZeRO sharding] DP Rank 1 has 139M out of 555M (25.00%) params' optimizer states
146
+ [default0]:07/03/2024 22:52:15 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: [ZeRO sharding] DP Rank 2 has 139M out of 555M (25.00%) params' optimizer states
147
+ [default0]:07/03/2024 22:52:15 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: [ZeRO sharding] DP Rank 3 has 139M out of 555M (25.00%) params' optimizer states
148
+ [default0]:07/03/2024 22:52:16 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: [Training Plan] Stage Training Stage has 19 remaining training steps and has consumed 0 samples
149
+ [default0]:07/03/2024 22:52:16 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Using `datasets` library
150
+ [default0]:07/03/2024 22:52:16 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Loading tokenizer from openai-community/gpt2 and transformers/hf_hub versions ('4.41.2', '0.23.4')
151
+ [default0]:07/03/2024 22:52:16 [WARNING|DP=0|PP=0|TP=0|ip-26-0-160-225]: Repo card metadata block was not found. Setting CardData to empty.
152
+ [default0]:Repo card metadata block was not found. Setting CardData to empty.
153
+ [default0]:07/03/2024 22:52:17 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: [Training Plan] There are 1 training stages
154
+ [default0]:07/03/2024 22:52:17 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: [Stage Training Stage] start from step 1
155
+ [default0]:07/03/2024 22:52:17 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]:
156
+ [default0]:07/03/2024 22:52:17 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: [Start training] datetime: 2024-07-03 22:52:17.276501 | mbs: 1 | grad_accum: 256 | global_batch_size: 1024 | sequence_length: 4096 | train_steps: 20 | start_iteration_step: 0 | consumed_train_samples: 0
157
+ [default0]:07/03/2024 22:52:17 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Resuming training from stage Training Stage, it has trained for 0 samples and has 19 remaining train steps
158
+ [default0]:07/03/2024 22:52:17 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Memory usage: 3729.08MiB. Peak allocated 3729.08MiB. Peak reserved: 3848.00MiB
159
+ [default1]:07/03/2024 22:52:17 [WARNING|DP=0|PP=0|TP=1|ip-26-0-160-225]: Repo card metadata block was not found. Setting CardData to empty.
160
+ [default2]:Repo card metadata block was not found. Setting CardData to empty.
161
+ [default6]:07/03/2024 22:52:17 [WARNING|DP=3|PP=0|TP=0|ip-26-0-160-225]: Repo card metadata block was not found. Setting CardData to empty.
162
+ [default7]:07/03/2024 22:52:17 [WARNING|DP=3|PP=0|TP=1|ip-26-0-160-225]: Repo card metadata block was not found. Setting CardData to empty.
163
+ [default2]:07/03/2024 22:52:17 [WARNING|DP=1|PP=0|TP=0|ip-26-0-160-225]: Repo card metadata block was not found. Setting CardData to empty.
164
+ [default6]:Repo card metadata block was not found. Setting CardData to empty.
165
+ [default4]:Repo card metadata block was not found. Setting CardData to empty.
166
+ [default1]:Repo card metadata block was not found. Setting CardData to empty.
167
+ [default4]:07/03/2024 22:52:17 [WARNING|DP=2|PP=0|TP=0|ip-26-0-160-225]: Repo card metadata block was not found. Setting CardData to empty.
168
+ [default5]:07/03/2024 22:52:17 [WARNING|DP=2|PP=0|TP=1|ip-26-0-160-225]: Repo card metadata block was not found. Setting CardData to empty.
169
+ [default5]:Repo card metadata block was not found. Setting CardData to empty.
170
+ [default7]:Repo card metadata block was not found. Setting CardData to empty.
171
+ [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.)
172
+ [default0]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
173
+ [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.)
174
+ [default1]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
175
+ [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.)
176
+ [default4]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
177
+ [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.)
178
+ [default5]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
179
+ [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.)
180
+ [default7]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
181
+ [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.)
182
+ [default6]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
183
+ [default3]:07/03/2024 22:52:35 [WARNING|DP=1|PP=0|TP=1|ip-26-0-160-225]: Repo card metadata block was not found. Setting CardData to empty.
184
+ [default3]:Repo card metadata block was not found. Setting CardData to empty.
185
+ [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.)
186
+ [default2]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
187
+ [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.)
188
+ [default3]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
189
+ [default0]:07/03/2024 22:52:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Memory usage: 3805.27MiB. Peak allocated 8461.15MiB. Peak reserved: 9066.00MiB
190
+ [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
191
+ [default0]: warnings.warn(
192
+ [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
193
+ [default1]: warnings.warn(
194
+ [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
195
+ [default5]: warnings.warn(
196
+ [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
197
+ [default4]: warnings.warn(
198
+ [default6]:/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py:2261: UserWarning: torch.distributed.all_reduce_coalesced will be deprecated. If you must use it, please revisit our documentation later at https://pytorch.org/docs/master/distributed.html#collective-functions
199
+ [default6]: warnings.warn(
200
+ [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
201
+ [default7]: warnings.warn(
202
+ [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
203
+ [default2]: warnings.warn(
204
+ [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
205
+ [default3]: warnings.warn(
206
+ [default0]:07/03/2024 22:53:13 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: iteration: 1 / 20 | consumed_tokens: 4.19M | elapsed_time_per_iteration_ms: 56.4K | tokens_per_sec: 74.3K | tokens_per_sec_per_gpu: 9.29K | global_batch_size: 1.02K | lm_loss: 11.5 | lr: 0.0001 | model_tflops_per_gpu: 84.3 | hardware_tflops_per_gpu: 84.3 | grad_norm: 26.4 | cuda_memory_allocated: 5.1G | cuda_max_memory_reserved: 11.5G | hd_total_memory_tb: 312G | hd_used_memory_tb: 66.5G | hd_free_memory_tb: 246G
207
+ [default0]:07/03/2024 22:53:13 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Memory usage: 4863.69MiB. Peak allocated 7244.99MiB. Peak reserved: 10924.00MiB
208
+ [default0]:07/03/2024 22:53:46 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Memory usage: 4863.82MiB. Peak allocated 9519.70MiB. Peak reserved: 10924.00MiB
209
+ [default0]:07/03/2024 22:53:47 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: iteration: 2 / 20 | consumed_tokens: 8.39M | elapsed_time_per_iteration_ms: 33.8K | tokens_per_sec: 124K | tokens_per_sec_per_gpu: 15.5K | global_batch_size: 1.02K | lm_loss: 11.5 | lr: 9.53e-05 | model_tflops_per_gpu: 141 | hardware_tflops_per_gpu: 141 | grad_norm: 26.6 | cuda_memory_allocated: 5.1G | cuda_max_memory_reserved: 11.5G | hd_total_memory_tb: 312G | hd_used_memory_tb: 66.5G | hd_free_memory_tb: 246G
210
+ [default0]:07/03/2024 22:53:47 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Memory usage: 4863.69MiB. Peak allocated 7245.11MiB. Peak reserved: 10924.00MiB
211
+ [default0]:07/03/2024 22:54:27 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Memory usage: 4863.82MiB. Peak allocated 9519.70MiB. Peak reserved: 10924.00MiB
212
+ [default0]:07/03/2024 22:54:27 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: iteration: 3 / 20 | consumed_tokens: 12.6M | elapsed_time_per_iteration_ms: 39.5K | tokens_per_sec: 106K | tokens_per_sec_per_gpu: 13.3K | global_batch_size: 1.02K | lm_loss: 11.5 | lr: 9.05e-05 | model_tflops_per_gpu: 121 | hardware_tflops_per_gpu: 121 | grad_norm: 262 | cuda_memory_allocated: 5.1G | cuda_max_memory_reserved: 11.5G | hd_total_memory_tb: 312G | hd_used_memory_tb: 66.5G | hd_free_memory_tb: 246G
213
+ [default0]:07/03/2024 22:54:27 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Memory usage: 4863.69MiB. Peak allocated 7245.11MiB. Peak reserved: 10924.00MiB
214
+ [default0]:STAGE:2024-07-03 22:54:27 347101:347101 ActivityProfilerController.cpp:314] Completed Stage: Warm Up
215
+ [default0]:07/03/2024 22:55:19 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Memory usage: 4863.82MiB. Peak allocated 9519.70MiB. Peak reserved: 10924.00MiB
216
+ [default0]:07/03/2024 22:55:19 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: iteration: 4 / 20 | consumed_tokens: 16.8M | elapsed_time_per_iteration_ms: 52.6K | tokens_per_sec: 79.7K | tokens_per_sec_per_gpu: 9.97K | global_batch_size: 1.02K | lm_loss: 14.6 | lr: 8.58e-05 | model_tflops_per_gpu: 90.4 | hardware_tflops_per_gpu: 90.4 | grad_norm: 29.1 | cuda_memory_allocated: 5.1G | cuda_max_memory_reserved: 11.5G | hd_total_memory_tb: 312G | hd_used_memory_tb: 66.5G | hd_free_memory_tb: 246G
217
+ [default0]:07/03/2024 22:55:19 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Memory usage: 4863.69MiB. Peak allocated 7245.11MiB. Peak reserved: 10924.00MiB
218
+ [default0]:07/03/2024 22:56:12 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: iteration: 5 / 20 | consumed_tokens: 21M | elapsed_time_per_iteration_ms: 52.9K | tokens_per_sec: 79.3K | tokens_per_sec_per_gpu: 9.91K | global_batch_size: 1.02K | lm_loss: 10.8 | lr: 8.11e-05 | model_tflops_per_gpu: 89.9 | hardware_tflops_per_gpu: 89.9 | grad_norm: 31.2
219
+ [default0]:07/03/2024 22:56:12 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Memory usage: 4863.69MiB. Peak allocated 9519.70MiB. Peak reserved: 10924.00MiB
220
+ [default0]:07/03/2024 22:57:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: iteration: 6 / 20 | consumed_tokens: 25.2M | elapsed_time_per_iteration_ms: 53.3K | tokens_per_sec: 78.7K | tokens_per_sec_per_gpu: 9.83K | global_batch_size: 1.02K | lm_loss: 10.6 | lr: 7.63e-05 | model_tflops_per_gpu: 89.2 | hardware_tflops_per_gpu: 89.2 | grad_norm: 27
221
+ [default0]:STAGE:2024-07-03 22:59:28 347101:347101 ActivityProfilerController.cpp:320] Completed Stage: Collection
222
+ [default0]:STAGE:2024-07-03 22:59:43 347101:347101 ActivityProfilerController.cpp:324] Completed Stage: Post Processing
223
+ [default1]:[rank1]:[E ProcessGroupNCCL.cpp:563] [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=305728, OpType=_REDUCE_SCATTER_BASE, NumelIn=8388608, NumelOut=4194304, Timeout(ms)=600000) ran for 600041 milliseconds before timing out.
224
+ [default2]:[rank2]:[E ProcessGroupNCCL.cpp:563] [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=165, OpType=ALLREDUCE, NumelIn=554881024, NumelOut=554881024, Timeout(ms)=600000) ran for 600084 milliseconds before timing out.
225
+ [default3]:[rank3]:[E ProcessGroupNCCL.cpp:563] [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=356672, OpType=COALESCED, NumelIn=2048, NumelOut=2048, Timeout(ms)=600000) ran for 600048 milliseconds before timing out.
226
+ [default3]:[rank3]:[E ProcessGroupNCCL.cpp:563] [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=165, OpType=ALLREDUCE, NumelIn=554881024, NumelOut=554881024, Timeout(ms)=600000) ran for 600077 milliseconds before timing out.
227
+ [default5]:[rank5]:[E ProcessGroupNCCL.cpp:563] [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=165, OpType=ALLREDUCE, NumelIn=554881024, NumelOut=554881024, Timeout(ms)=600000) ran for 600055 milliseconds before timing out.
228
+ [default2]:[rank2]:[E ProcessGroupNCCL.cpp:563] [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=356589, OpType=_ALLGATHER_BASE, NumelIn=4194304, NumelOut=8388608, Timeout(ms)=600000) ran for 600643 milliseconds before timing out.
229
+ [default4]:[rank4]:[E ProcessGroupNCCL.cpp:563] [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=165, OpType=ALLREDUCE, NumelIn=554881024, NumelOut=554881024, Timeout(ms)=600000) ran for 600019 milliseconds before timing out.
230
+ [default6]:[rank6]:[E ProcessGroupNCCL.cpp:563] [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=165, OpType=ALLREDUCE, NumelIn=554881024, NumelOut=554881024, Timeout(ms)=600000) ran for 600096 milliseconds before timing out.
231
+ [default4]:[rank4]:[E ProcessGroupNCCL.cpp:563] [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=356667, OpType=_ALLGATHER_BASE, NumelIn=4194304, NumelOut=8388608, Timeout(ms)=600000) ran for 600260 milliseconds before timing out.
232
+ [default7]:[rank7]:[E ProcessGroupNCCL.cpp:563] [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=165, OpType=ALLREDUCE, NumelIn=554881024, NumelOut=554881024, Timeout(ms)=600000) ran for 600086 milliseconds before timing out.
233
+ [default5]:[rank5]:[E ProcessGroupNCCL.cpp:563] [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=356617, OpType=_ALLGATHER_BASE, NumelIn=4194304, NumelOut=8388608, Timeout(ms)=600000) ran for 600337 milliseconds before timing out.
234
+ [default7]:[rank7]:[E ProcessGroupNCCL.cpp:563] [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=356671, OpType=_ALLGATHER_BASE, NumelIn=4194304, NumelOut=8388608, Timeout(ms)=600000) ran for 600322 milliseconds before timing out.
235
+ [default6]:[rank6]:[E ProcessGroupNCCL.cpp:563] [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=356619, OpType=_ALLGATHER_BASE, NumelIn=4194304, NumelOut=8388608, Timeout(ms)=600000) ran for 600376 milliseconds before timing out.
236
+ [default1]:[rank1]:[E ProcessGroupNCCL.cpp:1537] [PG 2 Rank 1] Timeout at NCCL work: 305728, last enqueued NCCL work: 305848, last completed NCCL work: 305727.
237
+ [default1]:[rank1]:[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.
238
+ [default1]:[rank1]:[E ProcessGroupNCCL.cpp:583] [Rank 1] To avoid data inconsistency, we are taking the entire process down.
239
+ [default1]:[rank1]:[E ProcessGroupNCCL.cpp:1414] [PG 2 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=305728, OpType=_REDUCE_SCATTER_BASE, NumelIn=8388608, NumelOut=4194304, Timeout(ms)=600000) ran for 600041 milliseconds before timing out.
240
+ [default1]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first):
241
+ [default1]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f841ba9b897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so)
242
+ [default1]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional<std::chrono::duration<long, std::ratio<1l, 1000l> > >) + 0x1d2 (0x7f841cd74c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
243
+ [default1]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f841cd79a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
244
+ [default1]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f841cd7adcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
245
+ [default1]:frame #4: <unknown function> + 0xd3e95 (0x7f8468813e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6)
246
+ [default1]:frame #5: <unknown function> + 0x8609 (0x7f846d85a609 in /lib/x86_64-linux-gnu/libpthread.so.0)
247
+ [default1]:frame #6: clone + 0x43 (0x7f846d625353 in /lib/x86_64-linux-gnu/libc.so.6)
248
+ [default1]:
249
+ [default1]:terminate called after throwing an instance of 'c10::DistBackendError'
250
+ [default1]: what(): [PG 2 Rank 1] Process group watchdog thread terminated with exception: [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=305728, OpType=_REDUCE_SCATTER_BASE, NumelIn=8388608, NumelOut=4194304, Timeout(ms)=600000) ran for 600041 milliseconds before timing out.
251
+ [default1]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first):
252
+ [default1]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f841ba9b897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so)
253
+ [default1]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional<std::chrono::duration<long, std::ratio<1l, 1000l> > >) + 0x1d2 (0x7f841cd74c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
254
+ [default1]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f841cd79a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
255
+ [default1]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f841cd7adcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
256
+ [default1]:frame #4: <unknown function> + 0xd3e95 (0x7f8468813e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6)
257
+ [default1]:frame #5: <unknown function> + 0x8609 (0x7f846d85a609 in /lib/x86_64-linux-gnu/libpthread.so.0)
258
+ [default1]:frame #6: clone + 0x43 (0x7f846d625353 in /lib/x86_64-linux-gnu/libc.so.6)
259
+ [default1]:
260
+ [default1]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first):
261
+ [default1]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f841ba9b897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so)
262
+ [default1]:frame #1: <unknown function> + 0xe32119 (0x7f841c9fe119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
263
+ [default1]:frame #2: <unknown function> + 0xd3e95 (0x7f8468813e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6)
264
+ [default1]:frame #3: <unknown function> + 0x8609 (0x7f846d85a609 in /lib/x86_64-linux-gnu/libpthread.so.0)
265
+ [default1]:frame #4: clone + 0x43 (0x7f846d625353 in /lib/x86_64-linux-gnu/libc.so.6)
266
+ [default1]:
267
+ W0703 23:08:10.250000 139636250072896 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 347101 closing signal SIGTERM
268
+ W0703 23:08:10.250000 139636250072896 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 347103 closing signal SIGTERM
269
+ W0703 23:08:10.251000 139636250072896 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 347104 closing signal SIGTERM
270
+ W0703 23:08:10.251000 139636250072896 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 347105 closing signal SIGTERM
271
+ W0703 23:08:10.251000 139636250072896 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 347106 closing signal SIGTERM
272
+ W0703 23:08:10.254000 139636250072896 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 347107 closing signal SIGTERM
273
+ W0703 23:08:10.254000 139636250072896 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 347108 closing signal SIGTERM
274
+ E0703 23:08:16.149000 139636250072896 torch/distributed/elastic/multiprocessing/api.py:826] failed (exitcode: -6) local_rank: 1 (pid: 347102) of binary: /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/python3.10
275
+ Traceback (most recent call last):
276
+ File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/torchrun", line 8, in <module>
277
+ sys.exit(main())
278
+ File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 347, in wrapper
279
+ return f(*args, **kwargs)
280
+ File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/run.py", line 879, in main
281
+ run(args)
282
+ File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/run.py", line 870, in run
283
+ elastic_launch(
284
+ File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 132, in __call__
285
+ return launch_agent(self._config, self._entrypoint, list(args))
286
+ File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 263, in launch_agent
287
+ raise ChildFailedError(
288
+ torch.distributed.elastic.multiprocessing.errors.ChildFailedError:
289
+ ============================================================
290
+ /fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py FAILED
291
+ ------------------------------------------------------------
292
+ Failures:
293
+ <NO_OTHER_FAILURES>
294
+ ------------------------------------------------------------
295
+ Root Cause (first observed failure):
296
+ [0]:
297
+ time : 2024-07-03_23:08:10
298
+ host : ip-26-0-160-225.ec2.internal
299
+ rank : 1 (local_rank: 1)
300
+ exitcode : -6 (pid: 347102)
301
+ error_file: <N/A>
302
+ traceback : Signal 6 (SIGABRT) received by PID 347102
303
+ ============================================================
304
+ srun: error: ip-26-0-160-225: task 0: Exited with exit code 1
305
+ Consider using `hf_transfer` for faster uploads. This solution comes with some limitations. See https://huggingface.co/docs/huggingface_hub/hf_transfer for more details.
llama-1B/8_GPUS/dp-4_tp-2_pp-1_mbz-1/status.txt ADDED
@@ -0,0 +1 @@
 
 
1
+ timeout