Hugging Face's logo

nanotron
/
old_bench

Model card Files Community
3outeille HF staff commited on
Commit
f9472bf
•
1 Parent(s): e7d410e

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

Browse files
Files changed (6) hide show
  1. .gitattributes +1 -0
  2. llama-1B/8_GPUS/dp-2_tp-2_pp-2_mbz-2/bench.slurm +111 -0
  3. llama-1B/8_GPUS/dp-2_tp-2_pp-2_mbz-2/config.yaml +90 -0
  4. llama-1B/8_GPUS/dp-2_tp-2_pp-2_mbz-2/log.out +422 -0
  5. llama-1B/8_GPUS/dp-2_tp-2_pp-2_mbz-2/profiler/ip-26-0-169-239_2654346.1720051209254005232.pt.trace.json.tmp +3 -0
  6. llama-1B/8_GPUS/dp-2_tp-2_pp-2_mbz-2/status.txt +1 -0
.gitattributes CHANGED
@@ -121,3 +121,4 @@ llama-1B/8_GPUS/dp-1_tp-4_pp-2_mbz-16/profiler/ip-26-0-174-36_488999.17200499548
121
  llama-1B/8_GPUS/dp-8_tp-1_pp-1_mbz-1/profiler/ip-26-0-160-225_351538.1720048932925630211.pt.trace.json filter=lfs diff=lfs merge=lfs -text
122
  llama-1B/8_GPUS/dp-4_tp-1_pp-2_mbz-1/profiler/ip-26-0-169-86_2278697.1720051013128234436.pt.trace.json.tmp filter=lfs diff=lfs merge=lfs -text
123
  llama-1B/8_GPUS/dp-2_tp-2_pp-2_mbz-8/profiler/ip-26-0-169-139_673212.1720050658975812000.pt.trace.json filter=lfs diff=lfs merge=lfs -text
 
 
121
  llama-1B/8_GPUS/dp-8_tp-1_pp-1_mbz-1/profiler/ip-26-0-160-225_351538.1720048932925630211.pt.trace.json filter=lfs diff=lfs merge=lfs -text
122
  llama-1B/8_GPUS/dp-4_tp-1_pp-2_mbz-1/profiler/ip-26-0-169-86_2278697.1720051013128234436.pt.trace.json.tmp filter=lfs diff=lfs merge=lfs -text
123
  llama-1B/8_GPUS/dp-2_tp-2_pp-2_mbz-8/profiler/ip-26-0-169-139_673212.1720050658975812000.pt.trace.json filter=lfs diff=lfs merge=lfs -text
124
+ llama-1B/8_GPUS/dp-2_tp-2_pp-2_mbz-2/profiler/ip-26-0-169-239_2654346.1720051209254005232.pt.trace.json.tmp filter=lfs diff=lfs merge=lfs -text
llama-1B/8_GPUS/dp-2_tp-2_pp-2_mbz-2/bench.slurm ADDED
@@ -0,0 +1,111 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/bin/bash
2
+
3
+ #SBATCH --job-name=bench_cluster
4
+ #SBATCH --time=02:00:00
5
+ #SBATCH --partition=hopper-prod
6
+ #SBATCH --nodes=1
7
+ #SBATCH --gres=gpu:8
8
+ #SBATCH --qos=normal
9
+ #SBATCH --ntasks-per-node=1
10
+ #SBATCH --cpus-per-task=96
11
+ #SBATCH --exclusive
12
+ #SBATCH --output=/fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-2_tp-2_pp-2_mbz-2/log.out
13
+ #SBATCH --error=/fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-2_tp-2_pp-2_mbz-2/log.out
14
+
15
+ # Function to update status based on squeue output
16
+ update_status() {
17
+ job_id=$1
18
+ status_file=$2
19
+ # For unknown reasons, it doenst update status for pending. It only works for running
20
+ while true; do
21
+ job_status=$(squeue --job $job_id --noheader --format=%T)
22
+ echo "Job status: $job_status"
23
+ if [ -z "$job_status" ]; then
24
+ # Job has finished or is not found
25
+ break
26
+ elif [ "$job_status" = "RUNNING" ]; then
27
+ printf "running" > $status_file
28
+ break
29
+ fi
30
+ sleep 10
31
+ done
32
+ }
33
+
34
+ # Misc initializations.
35
+ echo "========================"
36
+ echo "START TIME: $(date)"
37
+ source /fsx/ferdinandmom/miniforge3/etc/profile.d/conda.sh
38
+ conda activate /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster
39
+ echo python3 version = $(python3 --version)
40
+ echo "========================"
41
+
42
+ # Slurm stuff
43
+ export HOSTNAMES=$(scontrol show hostnames "$SLURM_JOB_NODELIST")
44
+ export MASTER_ADDR=$(scontrol show hostnames "$SLURM_JOB_NODELIST" | head -n 1)
45
+ export MASTER_PORT=$((1024 + RANDOM % 64511))
46
+
47
+ export TMPDIR=/scratch
48
+ export HF_DATASETS_CACHE="/admin/home/ferdinand_mom/.cache"
49
+ export CUBLAS_WORKSPACE_CONFIG=":4096:8"
50
+ export CUDA_DEVICE_MAX_CONNECTIONS="1"
51
+
52
+ huggingface-cli login --token $HUGGINGFACE_TOKEN
53
+
54
+
55
+ NANOTRON_REPO="/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron"
56
+ CMD="$NANOTRON_REPO/run_train.py --config-file /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-2_tp-2_pp-2_mbz-2/config.yaml"
57
+
58
+ LAUNCHER="torchrun \
59
+ --nproc_per_node 8 \
60
+ --nnodes 1 \
61
+ --rdzv_endpoint ${MASTER_ADDR}:${MASTER_PORT} \
62
+ --rdzv_backend c10d \
63
+ --max_restarts 0 \
64
+ --tee 3 \
65
+ --node_rank ${SLURM_PROCID}"
66
+
67
+ # Checkout the bench_cluster branch
68
+ cd $NANOTRON_REPO
69
+ git checkout bench_cluster
70
+ cd ..
71
+ # Get the current job ID
72
+ job_id=${SLURM_JOB_ID}
73
+
74
+ # Update status to "pending" or "running" in the background
75
+ update_status $job_id /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-2_tp-2_pp-2_mbz-2/status.txt &
76
+
77
+ # Run the main command
78
+ srun -u $LAUNCHER $CMD
79
+ exit_status=$?
80
+
81
+ # Update status based on the exit status of `srun`
82
+ if [ $exit_status -eq 0 ]; then
83
+ printf "completed" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-2_tp-2_pp-2_mbz-2/status.txt
84
+ else
85
+ if grep -q "OutOfMemoryError" /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-2_tp-2_pp-2_mbz-2/log.out; then
86
+ printf "oom" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-2_tp-2_pp-2_mbz-2/status.txt
87
+ elif grep -q " CUDA error: an illegal memory access" /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-2_tp-2_pp-2_mbz-2/log.out; then
88
+ printf "oom" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-2_tp-2_pp-2_mbz-2/status.txt
89
+ elif grep -q "Timeout at NCCL" /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-2_tp-2_pp-2_mbz-2/log.out; then
90
+ printf "timeout" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-2_tp-2_pp-2_mbz-2/status.txt
91
+ else
92
+ printf "fail" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-2_tp-2_pp-2_mbz-2/status.txt
93
+ fi
94
+ fi
95
+
96
+ # Run the report script if the job completed successfully
97
+ if [ $exit_status -eq 0 ]; then
98
+ python /fsx/ferdinandmom/ferdinand-hf/bench_cluster/main.py report --inp_dir /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-2_tp-2_pp-2_mbz-2 --is_logs
99
+ python /fsx/ferdinandmom/ferdinand-hf/bench_cluster/main.py report --inp_dir /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-2_tp-2_pp-2_mbz-2 --is_profiler
100
+ fi
101
+
102
+
103
+ # Push to hub the folder using huggingface_cli
104
+ huggingface-cli upload nanotron/bench_cluster /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-2_tp-2_pp-2_mbz-2 llama-1B/8_GPUS/dp-2_tp-2_pp-2_mbz-2 --commit-message "Upload llama-1B/8_GPUS/dp-2_tp-2_pp-2_mbz-2"
105
+
106
+ # Verify the upload
107
+ if [ $? -eq 0 ]; then
108
+ echo "Uploading to Huggingface Hub successful"
109
+ else
110
+ echo "Failed to upload to Huggingface Hub"
111
+ fi
llama-1B/8_GPUS/dp-2_tp-2_pp-2_mbz-2/config.yaml ADDED
@@ -0,0 +1,90 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ general:
2
+ project: bench_cluster
3
+ seed: 42
4
+ model:
5
+ ddp_bucket_cap_mb: 25
6
+ dtype: bfloat16
7
+ init_method:
8
+ std: 0.025
9
+ make_vocab_size_divisible_by: 1
10
+ model_config:
11
+ bos_token_id: 1
12
+ eos_token_id: 2
13
+ hidden_act: silu
14
+ hidden_size: 2048
15
+ initializer_range: 0.02
16
+ intermediate_size: 4096
17
+ is_llama_config: true
18
+ max_position_embeddings: 4096
19
+ num_attention_heads: 32
20
+ num_hidden_layers: 24
21
+ num_key_value_heads: 32
22
+ pad_token_id: null
23
+ pretraining_tp: 1
24
+ rms_norm_eps: 1.0e-05
25
+ rope_scaling: null
26
+ rope_theta: 10000.0
27
+ tie_word_embeddings: true
28
+ use_cache: true
29
+ vocab_size: 50257
30
+ optimizer:
31
+ accumulate_grad_in_fp32: true
32
+ clip_grad: 1.0
33
+ learning_rate_scheduler:
34
+ learning_rate: 0.0001
35
+ lr_decay_style: linear
36
+ lr_warmup_style: linear
37
+ lr_warmup_steps: 1
38
+ min_decay_lr: 1.0e-05
39
+ optimizer_factory:
40
+ adam_beta1: 0.9
41
+ adam_beta2: 0.95
42
+ adam_eps: 1.0e-08
43
+ name: adamW
44
+ torch_adam_is_fused: true
45
+ weight_decay: 0.01
46
+ zero_stage: 1
47
+ parallelism:
48
+ dp: 2
49
+ expert_parallel_size: 1
50
+ pp: 2
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-2_tp-2_pp-2_mbz-2
57
+ tokenizer:
58
+ tokenizer_max_length: null
59
+ tokenizer_name_or_path: openai-community/gpt2
60
+ tokenizer_revision: null
61
+ data_stages:
62
+ - name: Training Stage
63
+ start_training_step: 1
64
+ data:
65
+ dataset:
66
+ dataset_overwrite_cache: false
67
+ dataset_processing_num_proc_per_process: 64
68
+ hf_dataset_config_name: null
69
+ hf_dataset_or_datasets: roneneldan/TinyStories
70
+ hf_dataset_splits: train
71
+ text_column_name: text
72
+ num_loading_workers: 0
73
+ seed: 42
74
+ lighteval: null
75
+ tokens:
76
+ train_steps: 20
77
+ val_check_interval: -1
78
+ batch_accumulation_per_replica: 256
79
+ limit_test_batches: 0
80
+ limit_val_batches: 0
81
+ micro_batch_size: 2
82
+ sequence_length: 4096
83
+ logging:
84
+ iteration_step_info_interval: 1
85
+ log_level: info
86
+ log_level_replica: info
87
+ checkpoints:
88
+ checkpoint_interval: 100000
89
+ checkpoints_path: /dev/null
90
+ resume_checkpoint_path: null
llama-1B/8_GPUS/dp-2_tp-2_pp-2_mbz-2/log.out ADDED
@@ -0,0 +1,422 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ========================
2
+ START TIME: Wed Jul 3 23:46:11 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 23:46:19.526000 139684507649856 torch/distributed/run.py:757]
18
+ W0703 23:46:19.526000 139684507649856 torch/distributed/run.py:757] *****************************************
19
+ W0703 23:46:19.526000 139684507649856 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 23:46:19.526000 139684507649856 torch/distributed/run.py:757] *****************************************
21
+ [default0]:07/03/2024 23:46:41 [WARNING|DP=0|PP=0|TP=0|ip-26-0-169-239]: [Vocab Size Padding] Padded vocab (size: 50257) with 1 dummy tokens (new size: 50258)
22
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: Config:
23
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: Config(general=GeneralArgs(project='bench_cluster',
24
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: run='%date_%jobid',
25
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: seed=42,
26
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: step=None,
27
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: consumed_train_samples=None,
28
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: benchmark_csv_path=None,
29
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: ignore_sanity_checks=True),
30
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: parallelism=ParallelismArgs(dp=2,
31
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: pp=2,
32
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: tp=2,
33
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: pp_engine=<nanotron.parallel.pipeline_parallel.engine.OneForwardOneBackwardPipelineEngine object at 0x7fb6f028c8b0>,
34
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: tp_mode=<TensorParallelLinearMode.REDUCE_SCATTER: 2>,
35
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: tp_linear_async_communication=False,
36
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: expert_parallel_size=1),
37
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: model=ModelArgs(model_config=LlamaConfig(bos_token_id=1,
38
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: eos_token_id=2,
39
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: hidden_act='silu',
40
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: hidden_size=2048,
41
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: initializer_range=0.02,
42
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: intermediate_size=4096,
43
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: is_llama_config=True,
44
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: max_position_embeddings=4096,
45
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: num_attention_heads=32,
46
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: num_hidden_layers=24,
47
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: num_key_value_heads=32,
48
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: pad_token_id=None,
49
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: pretraining_tp=1,
50
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: rms_norm_eps=1e-05,
51
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: rope_scaling=None,
52
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: rope_theta=10000.0,
53
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: tie_word_embeddings=True,
54
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: use_cache=True,
55
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: vocab_size=50258),
56
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: init_method=RandomInit(std=0.025),
57
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: dtype=torch.bfloat16,
58
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: make_vocab_size_divisible_by=1,
59
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: ddp_bucket_cap_mb=25),
60
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: tokenizer=TokenizerArgs(tokenizer_name_or_path='openai-community/gpt2',
61
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: tokenizer_revision=None,
62
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: tokenizer_max_length=None),
63
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: checkpoints=CheckpointsArgs(checkpoints_path=Path('/dev/null'),
64
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: checkpoint_interval=100000,
65
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: save_initial_state=False,
66
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: resume_checkpoint_path=None,
67
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: checkpoints_path_is_shared_file_system=False),
68
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: logging=LoggingArgs(log_level='info',
69
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: log_level_replica='info',
70
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: iteration_step_info_interval=1),
71
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: tokens=TokensArgs(sequence_length=4096,
72
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: train_steps=20,
73
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: micro_batch_size=2,
74
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: batch_accumulation_per_replica=256,
75
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: val_check_interval=-1,
76
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: limit_val_batches=0,
77
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: limit_test_batches=0),
78
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: optimizer=OptimizerArgs(optimizer_factory=AdamWOptimizerArgs(adam_eps=1e-08,
79
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: adam_beta1=0.9,
80
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: adam_beta2=0.95,
81
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: torch_adam_is_fused=True,
82
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: name='adamW'),
83
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: zero_stage=1,
84
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: weight_decay=0.01,
85
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: clip_grad=1.0,
86
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: accumulate_grad_in_fp32=True,
87
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: learning_rate_scheduler=LRSchedulerArgs(learning_rate=0.0001,
88
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: lr_warmup_steps=1,
89
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: lr_warmup_style='linear',
90
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: lr_decay_style='linear',
91
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: lr_decay_steps=19,
92
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: lr_decay_starting_step=None,
93
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: min_decay_lr=1e-05)),
94
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: data_stages=[DatasetStageArgs(name='Training Stage',
95
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: start_training_step=1,
96
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: data=DataArgs(dataset=PretrainDatasetsArgs(hf_dataset_or_datasets='roneneldan/TinyStories',
97
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: hf_dataset_splits='train',
98
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: hf_dataset_config_name=None,
99
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: dataset_processing_num_proc_per_process=64,
100
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: dataset_overwrite_cache=False,
101
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: text_column_name='text'),
102
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: seed=42,
103
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: num_loading_workers=0))],
104
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: profiler=ProfilerArgs(profiler_export_path=Path('/fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-2_tp-2_pp-2_mbz-2')),
105
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: lighteval=None)
106
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: Model Config:
107
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: LlamaConfig(bos_token_id=1,
108
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: eos_token_id=2,
109
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: hidden_act='silu',
110
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: hidden_size=2048,
111
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: initializer_range=0.02,
112
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: intermediate_size=4096,
113
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: is_llama_config=True,
114
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: max_position_embeddings=4096,
115
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: num_attention_heads=32,
116
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: num_hidden_layers=24,
117
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: num_key_value_heads=32,
118
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: pad_token_id=None,
119
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: pretraining_tp=1,
120
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: rms_norm_eps=1e-05,
121
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: rope_scaling=None,
122
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: rope_theta=10000.0,
123
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: tie_word_embeddings=True,
124
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: use_cache=True,
125
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: vocab_size=50258)
126
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: Building model..
127
+ [default0]:07/03/2024 23:46:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: Setting PP block ranks...
128
+ [default0]:07/03/2024 23:46:53 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: Total number of parameters: 1.21G (2313.02MiB)
129
+ [default0]:07/03/2024 23:46:53 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: Local number of parameters: 345M (658.27MiB)
130
+ [default0]:07/03/2024 23:46:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: [After model building] Memory usage: 672.29MiB. Peak allocated: 674.32MiB Peak reserved: 690.00MiB
131
+ [default0]:07/03/2024 23:46:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: No checkpoint path provided.
132
+ [default0]:07/03/2024 23:46:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: Parametrizing model parameters using StandardParametrizator
133
+ [default7]:07/03/2024 23:46:54 [INFO|DP=1|PP=1|TP=1|ip-26-0-169-239]: No checkpoint path provided.
134
+ [default1]:07/03/2024 23:46:53 [INFO|DP=0|PP=0|TP=1|ip-26-0-169-239]: Local number of parameters: 345M (658.27MiB)
135
+ [default1]:07/03/2024 23:46:54 [INFO|DP=0|PP=0|TP=1|ip-26-0-169-239]: [After model building] Memory usage: 672.29MiB. Peak allocated: 674.32MiB Peak reserved: 690.00MiB
136
+ [default1]:07/03/2024 23:46:54 [INFO|DP=0|PP=0|TP=1|ip-26-0-169-239]: No checkpoint path provided.
137
+ [default5]:07/03/2024 23:46:53 [INFO|DP=0|PP=1|TP=1|ip-26-0-169-239]: Local number of parameters: 261M (498.24MiB)
138
+ [default5]:07/03/2024 23:46:54 [INFO|DP=0|PP=1|TP=1|ip-26-0-169-239]: [After model building] Memory usage: 508.26MiB. Peak allocated: 510.29MiB Peak reserved: 526.00MiB
139
+ [default4]:07/03/2024 23:46:53 [INFO|DP=0|PP=1|TP=0|ip-26-0-169-239]: Local number of parameters: 261M (498.24MiB)
140
+ [default4]:07/03/2024 23:46:54 [INFO|DP=0|PP=1|TP=0|ip-26-0-169-239]: [After model building] Memory usage: 508.26MiB. Peak allocated: 510.29MiB Peak reserved: 526.00MiB
141
+ [default4]:07/03/2024 23:46:54 [INFO|DP=0|PP=1|TP=0|ip-26-0-169-239]: No checkpoint path provided.
142
+ [default5]:07/03/2024 23:46:54 [INFO|DP=0|PP=1|TP=1|ip-26-0-169-239]: No checkpoint path provided.
143
+ [default2]:07/03/2024 23:46:54 [INFO|DP=1|PP=0|TP=0|ip-26-0-169-239]: No checkpoint path provided.
144
+ [default6]:07/03/2024 23:46:54 [INFO|DP=1|PP=1|TP=0|ip-26-0-169-239]: No checkpoint path provided.
145
+ [default3]:07/03/2024 23:46:54 [INFO|DP=1|PP=0|TP=1|ip-26-0-169-239]: No checkpoint path provided.
146
+ [default0]:07/03/2024 23:46:56 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: [Optimizer Building] Using LearningRateForSP as learning rate
147
+ [default0]:07/03/2024 23:46:56 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: [ZeRO sharding] Size of optimizer params per rank:
148
+ [default0]:07/03/2024 23:46:56 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: [ZeRO sharding] DP Rank 0 has 173M out of 345M (50.00%) params' optimizer states
149
+ [default0]:07/03/2024 23:46:56 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: [ZeRO sharding] DP Rank 1 has 173M out of 345M (50.00%) params' optimizer states
150
+ [default0]:07/03/2024 23:46:57 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: [Training Plan] Stage Training Stage has 19 remaining training steps and has consumed 0 samples
151
+ [default0]:07/03/2024 23:46:57 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: Using `datasets` library
152
+ [default0]:07/03/2024 23:46:57 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: Loading tokenizer from openai-community/gpt2 and transformers/hf_hub versions ('4.41.2', '0.23.4')
153
+ [default0]:07/03/2024 23:46:58 [WARNING|DP=0|PP=0|TP=0|ip-26-0-169-239]: Repo card metadata block was not found. Setting CardData to empty.
154
+ [default0]:Repo card metadata block was not found. Setting CardData to empty.
155
+ [default0]:07/03/2024 23:47:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: [Training Plan] There are 1 training stages
156
+ [default0]:07/03/2024 23:47:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: [Stage Training Stage] start from step 1
157
+ [default0]:07/03/2024 23:47:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]:
158
+ [default0]:07/03/2024 23:47:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: [Start training] datetime: 2024-07-03 23:47:00.050277 | mbs: 2 | grad_accum: 256 | global_batch_size: 1024 | sequence_length: 4096 | train_steps: 20 | start_iteration_step: 0 | consumed_train_samples: 0
159
+ [default0]:07/03/2024 23:47:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: Resuming training from stage Training Stage, it has trained for 0 samples and has 19 remaining train steps
160
+ [default0]:07/03/2024 23:47:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: Memory usage: 2647.09MiB. Peak allocated 2647.09MiB. Peak reserved: 2668.00MiB
161
+ [default1]:07/03/2024 23:47:00 [WARNING|DP=0|PP=0|TP=1|ip-26-0-169-239]: Repo card metadata block was not found. Setting CardData to empty.
162
+ [default1]:Repo card metadata block was not found. Setting CardData to empty.
163
+ [default3]:Repo card metadata block was not found. Setting CardData to empty.
164
+ [default3]:07/03/2024 23:47:00 [WARNING|DP=1|PP=0|TP=1|ip-26-0-169-239]: Repo card metadata block was not found. Setting CardData to empty.
165
+ [default5]:Repo card metadata block was not found. Setting CardData to empty.
166
+ [default6]:07/03/2024 23:47:00 [WARNING|DP=1|PP=1|TP=0|ip-26-0-169-239]: Repo card metadata block was not found. Setting CardData to empty.
167
+ [default6]:Repo card metadata block was not found. Setting CardData to empty.
168
+ [default7]:07/03/2024 23:47:00 [WARNING|DP=1|PP=1|TP=1|ip-26-0-169-239]: Repo card metadata block was not found. Setting CardData to empty.
169
+ [default4]:07/03/2024 23:47:00 [WARNING|DP=0|PP=1|TP=0|ip-26-0-169-239]: Repo card metadata block was not found. Setting CardData to empty.
170
+ [default5]:07/03/2024 23:47:00 [WARNING|DP=0|PP=1|TP=1|ip-26-0-169-239]: Repo card metadata block was not found. Setting CardData to empty.
171
+ [default4]:Repo card metadata block was not found. Setting CardData to empty.
172
+ [default2]:07/03/2024 23:47:00 [WARNING|DP=1|PP=0|TP=0|ip-26-0-169-239]: Repo card metadata block was not found. Setting CardData to empty.
173
+ [default2]:Repo card metadata block was not found. Setting CardData to empty.
174
+ [default7]:Repo card metadata block was not found. Setting CardData to empty.
175
+ [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.)
176
+ [default5]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
177
+ [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.)
178
+ [default4]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
179
+ [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.)
180
+ [default1]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
181
+ [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.)
182
+ [default0]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
183
+ [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.)
184
+ [default7]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
185
+ [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.)
186
+ [default6]: 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
+ [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.)
190
+ [default2]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
191
+ [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
192
+ [default4]: warnings.warn(
193
+ [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
194
+ [default5]: warnings.warn(
195
+ [default0]:07/03/2024 23:47:34 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: Memory usage: 2718.16MiB. Peak allocated 12298.83MiB. Peak reserved: 12508.00MiB
196
+ [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
197
+ [default1]: warnings.warn(
198
+ [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
199
+ [default0]: warnings.warn(
200
+ [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
201
+ [default6]: 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
+ [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
+ [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
207
+ [default2]: warnings.warn(
208
+ [default0]:07/03/2024 23:47:41 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: Memory usage: 4034.75MiB. Peak allocated 5683.01MiB. Peak reserved: 14158.00MiB
209
+ [default4]:07/03/2024 23:47:41 [INFO|DP=0|PP=1|TP=0|ip-26-0-169-239]: iteration: 1 / 20 | consumed_tokens: 4.19M | elapsed_time_per_iteration_ms: 41.1K | tokens_per_sec: 102K | tokens_per_sec_per_gpu: 12.8K | global_batch_size: 1.02K | lm_loss: 11.2 | lr: 0.0001 | model_tflops_per_gpu: 116 | hardware_tflops_per_gpu: 116 | grad_norm: 21.2 | cuda_memory_allocated: 3.22G | cuda_max_memory_reserved: 8.68G | hd_total_memory_tb: 312G | hd_used_memory_tb: 65.6G | hd_free_memory_tb: 247G
210
+ [default0]:07/03/2024 23:48:03 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: Memory usage: 4034.75MiB. Peak allocated 13545.81MiB. Peak reserved: 14158.00MiB
211
+ [default4]:07/03/2024 23:48:03 [INFO|DP=0|PP=1|TP=0|ip-26-0-169-239]: iteration: 2 / 20 | consumed_tokens: 8.39M | elapsed_time_per_iteration_ms: 22K | tokens_per_sec: 191K | tokens_per_sec_per_gpu: 23.9K | global_batch_size: 1.02K | lm_loss: 11.2 | lr: 9.53e-05 | model_tflops_per_gpu: 217 | hardware_tflops_per_gpu: 217 | grad_norm: 21.3 | cuda_memory_allocated: 3.22G | cuda_max_memory_reserved: 8.68G | hd_total_memory_tb: 312G | hd_used_memory_tb: 65.6G | hd_free_memory_tb: 247G
212
+ [default0]:07/03/2024 23:48:03 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: Memory usage: 4034.75MiB. Peak allocated 5683.01MiB. Peak reserved: 14158.00MiB
213
+ [default0]:07/03/2024 23:48:28 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: Memory usage: 4034.75MiB. Peak allocated 13545.81MiB. Peak reserved: 14158.00MiB
214
+ [default0]:07/03/2024 23:48:28 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: Memory usage: 4034.75MiB. Peak allocated 5683.01MiB. Peak reserved: 14158.00MiB
215
+ [default4]:07/03/2024 23:48:28 [INFO|DP=0|PP=1|TP=0|ip-26-0-169-239]: iteration: 3 / 20 | consumed_tokens: 12.6M | elapsed_time_per_iteration_ms: 24.4K | tokens_per_sec: 172K | tokens_per_sec_per_gpu: 21.5K | global_batch_size: 1.02K | lm_loss: 9.94 | lr: 9.05e-05 | model_tflops_per_gpu: 195 | hardware_tflops_per_gpu: 195 | grad_norm: 114 | cuda_memory_allocated: 3.22G | cuda_max_memory_reserved: 8.68G | hd_total_memory_tb: 312G | hd_used_memory_tb: 65.6G | hd_free_memory_tb: 247G
216
+ [default0]:STAGE:2024-07-03 23:48:28 2654346:2654346 ActivityProfilerController.cpp:314] Completed Stage: Warm Up
217
+ [default0]:07/03/2024 23:49:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: Memory usage: 4034.75MiB. Peak allocated 13545.81MiB. Peak reserved: 14158.00MiB
218
+ [default0]:07/03/2024 23:49:00 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: Memory usage: 4034.75MiB. Peak allocated 5683.01MiB. Peak reserved: 14158.00MiB
219
+ [default4]:07/03/2024 23:49:00 [INFO|DP=0|PP=1|TP=0|ip-26-0-169-239]: iteration: 4 / 20 | consumed_tokens: 16.8M | elapsed_time_per_iteration_ms: 32.6K | tokens_per_sec: 129K | tokens_per_sec_per_gpu: 16.1K | global_batch_size: 1.02K | lm_loss: 13.3 | lr: 8.58e-05 | model_tflops_per_gpu: 146 | hardware_tflops_per_gpu: 146 | grad_norm: 22.8 | cuda_memory_allocated: 3.22G | cuda_max_memory_reserved: 8.68G | hd_total_memory_tb: 312G | hd_used_memory_tb: 65.6G | hd_free_memory_tb: 247G
220
+ [default0]:07/03/2024 23:49:33 [INFO|DP=0|PP=0|TP=0|ip-26-0-169-239]: Memory usage: 4034.75MiB. Peak allocated 13545.81MiB. Peak reserved: 14158.00MiB
221
+ [default4]:07/03/2024 23:49:33 [INFO|DP=0|PP=1|TP=0|ip-26-0-169-239]: iteration: 5 / 20 | consumed_tokens: 21M | elapsed_time_per_iteration_ms: 33K | tokens_per_sec: 127K | tokens_per_sec_per_gpu: 15.9K | global_batch_size: 1.02K | lm_loss: 10.2 | lr: 8.11e-05 | model_tflops_per_gpu: 144 | hardware_tflops_per_gpu: 144 | grad_norm: 10.4
222
+ [default4]:07/03/2024 23:50:06 [INFO|DP=0|PP=1|TP=0|ip-26-0-169-239]: iteration: 6 / 20 | consumed_tokens: 25.2M | elapsed_time_per_iteration_ms: 32.7K | tokens_per_sec: 128K | tokens_per_sec_per_gpu: 16K | global_batch_size: 1.02K | lm_loss: 9.36 | lr: 7.63e-05 | model_tflops_per_gpu: 146 | hardware_tflops_per_gpu: 146 | grad_norm: 15.4
223
+ [default0]:STAGE:2024-07-03 23:51:32 2654346:2654346 ActivityProfilerController.cpp:320] Completed Stage: Collection
224
+ [default0]:STAGE:2024-07-03 23:51:41 2654346:2654346 ActivityProfilerController.cpp:324] Completed Stage: Post Processing
225
+ [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 600019 milliseconds before timing out.
226
+ [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 600009 milliseconds before timing out.
227
+ [default1]:[rank1]:[E ProcessGroupNCCL.cpp:563] [Rank 1] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=175141, OpType=_REDUCE_SCATTER_BASE, NumelIn=16777216, NumelOut=8388608, Timeout(ms)=600000) ran for 600054 milliseconds before timing out.
228
+ [default5]:[rank5]: Traceback (most recent call last):
229
+ [default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in <module>
230
+ [default5]:[rank5]: trainer.train(dataloader)
231
+ [default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train
232
+ [default5]:[rank5]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader)
233
+ [default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step
234
+ [default5]:[rank5]: outputs = self.pipeline_engine.train_batch_iter(
235
+ [default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 278, in train_batch_iter
236
+ [default5]:[rank5]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model)
237
+ [default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward
238
+ [default5]:[rank5]: output = model(**micro_batch)
239
+ [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
240
+ [default5]:[rank5]: return self._call_impl(*args, **kwargs)
241
+ [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
242
+ [default5]:[rank5]: return forward_call(*args, **kwargs)
243
+ [default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward
244
+ [default5]:[rank5]: sharded_logits = self.model(
245
+ [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
246
+ [default5]:[rank5]: return self._call_impl(*args, **kwargs)
247
+ [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
248
+ [default5]:[rank5]: return forward_call(*args, **kwargs)
249
+ [default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward
250
+ [default5]:[rank5]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0]
251
+ [default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states
252
+ [default5]:[rank5]: hidden_encoder_states = encoder_block(**hidden_encoder_states)
253
+ [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
254
+ [default5]:[rank5]: return self._call_impl(*args, **kwargs)
255
+ [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
256
+ [default5]:[rank5]: return forward_call(*args, **kwargs)
257
+ [default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 126, in forward
258
+ [default5]:[rank5]: new_kwargs[name] = recv_from_pipeline_state_buffer(
259
+ [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
260
+ [default5]:[rank5]: pipeline_state.run_communication()
261
+ [default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 150, in run_communication
262
+ [default5]:[rank5]: recv_activation_tensor = recv_activation()
263
+ [default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 31, in __call__
264
+ [default5]:[rank5]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0]
265
+ [default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors
266
+ [default5]:[rank5]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag)
267
+ [default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors
268
+ [default5]:[rank5]: meta = self._recv_meta(from_rank=from_rank, tag=tag)
269
+ [default5]:[rank5]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 267, in _recv_meta
270
+ [default5]:[rank5]: self.second_metadata = torch.empty(second_metadata_num_elements, dtype=torch.long, device=self.device)
271
+ [default5]:[rank5]: torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate more than 1EB memory.
272
+ [default4]:[rank4]: Traceback (most recent call last):
273
+ [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py", line 237, in <module>
274
+ [default4]:[rank4]: trainer.train(dataloader)
275
+ [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 429, in train
276
+ [default4]:[rank4]: outputs, loss_avg = self.training_step(dataloader=self.current_dataloader)
277
+ [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/trainer.py", line 462, in training_step
278
+ [default4]:[rank4]: outputs = self.pipeline_engine.train_batch_iter(
279
+ [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 278, in train_batch_iter
280
+ [default4]:[rank4]: output = self.forward(context=context, state=state, micro_batch=micro_batch, model=model)
281
+ [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/engine.py", line 44, in forward
282
+ [default4]:[rank4]: output = model(**micro_batch)
283
+ [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
284
+ [default4]:[rank4]: return self._call_impl(*args, **kwargs)
285
+ [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
286
+ [default4]:[rank4]: return forward_call(*args, **kwargs)
287
+ [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 891, in forward
288
+ [default4]:[rank4]: sharded_logits = self.model(
289
+ [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
290
+ [default4]:[rank4]: return self._call_impl(*args, **kwargs)
291
+ [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
292
+ [default4]:[rank4]: return forward_call(*args, **kwargs)
293
+ [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 764, in forward
294
+ [default4]:[rank4]: return self.forward_with_hidden_states(input_ids=input_ids, input_mask=input_mask)[0]
295
+ [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/models/llama.py", line 780, in forward_with_hidden_states
296
+ [default4]:[rank4]: hidden_encoder_states = encoder_block(**hidden_encoder_states)
297
+ [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
298
+ [default4]:[rank4]: return self._call_impl(*args, **kwargs)
299
+ [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
300
+ [default4]:[rank4]: return forward_call(*args, **kwargs)
301
+ [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/block.py", line 126, in forward
302
+ [default4]:[rank4]: new_kwargs[name] = recv_from_pipeline_state_buffer(
303
+ [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
304
+ [default4]:[rank4]: pipeline_state.run_communication()
305
+ [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 150, in run_communication
306
+ [default4]:[rank4]: recv_activation_tensor = recv_activation()
307
+ [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/state.py", line 31, in __call__
308
+ [default4]:[rank4]: return self.p2p.recv_tensors(num_tensors=1, from_rank=self.from_rank)[0]
309
+ [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 353, in recv_tensors
310
+ [default4]:[rank4]: buffers, futures = self.irecv_tensors(num_tensors=num_tensors, from_rank=from_rank, tag=tag)
311
+ [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 326, in irecv_tensors
312
+ [default4]:[rank4]: meta = self._recv_meta(from_rank=from_rank, tag=tag)
313
+ [default4]:[rank4]: File "/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/src/nanotron/parallel/pipeline_parallel/p2p.py", line 267, in _recv_meta
314
+ [default4]:[rank4]: self.second_metadata = torch.empty(second_metadata_num_elements, dtype=torch.long, device=self.device)
315
+ [default4]:[rank4]: torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate more than 1EB memory.
316
+ [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.
317
+ [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.
318
+ [default5]:[rank5]:[E ProcessGroupNCCL.cpp:583] [Rank 1] To avoid data inconsistency, we are taking the entire process down.
319
+ [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 600019 milliseconds before timing out.
320
+ [default5]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first):
321
+ [default5]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f6d59caf897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so)
322
+ [default5]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional<std::chrono::duration<long, std::ratio<1l, 1000l> > >) + 0x1d2 (0x7f6d5af88c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
323
+ [default5]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f6d5af8da80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
324
+ [default5]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f6d5af8edcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
325
+ [default5]:frame #4: <unknown function> + 0xd3e95 (0x7f6da6a27e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6)
326
+ [default5]:frame #5: <unknown function> + 0x8609 (0x7f6daba6e609 in /lib/x86_64-linux-gnu/libpthread.so.0)
327
+ [default5]:frame #6: clone + 0x43 (0x7f6dab839353 in /lib/x86_64-linux-gnu/libc.so.6)
328
+ [default5]:
329
+ [default5]:terminate called after throwing an instance of 'c10::DistBackendError'
330
+ [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 600019 milliseconds before timing out.
331
+ [default5]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first):
332
+ [default5]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f6d59caf897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so)
333
+ [default5]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional<std::chrono::duration<long, std::ratio<1l, 1000l> > >) + 0x1d2 (0x7f6d5af88c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
334
+ [default5]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7f6d5af8da80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
335
+ [default5]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7f6d5af8edcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
336
+ [default5]:frame #4: <unknown function> + 0xd3e95 (0x7f6da6a27e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6)
337
+ [default5]:frame #5: <unknown function> + 0x8609 (0x7f6daba6e609 in /lib/x86_64-linux-gnu/libpthread.so.0)
338
+ [default5]:frame #6: clone + 0x43 (0x7f6dab839353 in /lib/x86_64-linux-gnu/libc.so.6)
339
+ [default5]:
340
+ [default5]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first):
341
+ [default5]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7f6d59caf897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so)
342
+ [default5]:frame #1: <unknown function> + 0xe32119 (0x7f6d5ac12119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
343
+ [default5]:frame #2: <unknown function> + 0xd3e95 (0x7f6da6a27e95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6)
344
+ [default5]:frame #3: <unknown function> + 0x8609 (0x7f6daba6e609 in /lib/x86_64-linux-gnu/libpthread.so.0)
345
+ [default5]:frame #4: clone + 0x43 (0x7f6dab839353 in /lib/x86_64-linux-gnu/libc.so.6)
346
+ [default5]:
347
+ [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.
348
+ [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.
349
+ [default4]:[rank4]:[E ProcessGroupNCCL.cpp:583] [Rank 1] To avoid data inconsistency, we are taking the entire process down.
350
+ [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 600009 milliseconds before timing out.
351
+ [default4]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first):
352
+ [default4]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fb5966c7897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so)
353
+ [default4]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional<std::chrono::duration<long, std::ratio<1l, 1000l> > >) + 0x1d2 (0x7fb5979a0c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
354
+ [default4]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7fb5979a5a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
355
+ [default4]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7fb5979a6dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
356
+ [default4]:frame #4: <unknown function> + 0xd3e95 (0x7fb5e343fe95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6)
357
+ [default4]:frame #5: <unknown function> + 0x8609 (0x7fb5e8486609 in /lib/x86_64-linux-gnu/libpthread.so.0)
358
+ [default4]:frame #6: clone + 0x43 (0x7fb5e8251353 in /lib/x86_64-linux-gnu/libc.so.6)
359
+ [default4]:
360
+ [default4]:terminate called after throwing an instance of 'c10::DistBackendError'
361
+ [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 600009 milliseconds before timing out.
362
+ [default4]:Exception raised from checkTimeout at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:565 (most recent call first):
363
+ [default4]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fb5966c7897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so)
364
+ [default4]:frame #1: c10d::ProcessGroupNCCL::WorkNCCL::checkTimeout(std::optional<std::chrono::duration<long, std::ratio<1l, 1000l> > >) + 0x1d2 (0x7fb5979a0c62 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
365
+ [default4]:frame #2: c10d::ProcessGroupNCCL::watchdogHandler() + 0x1a0 (0x7fb5979a5a80 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
366
+ [default4]:frame #3: c10d::ProcessGroupNCCL::ncclCommWatchdog() + 0x10c (0x7fb5979a6dcc in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
367
+ [default4]:frame #4: <unknown function> + 0xd3e95 (0x7fb5e343fe95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6)
368
+ [default4]:frame #5: <unknown function> + 0x8609 (0x7fb5e8486609 in /lib/x86_64-linux-gnu/libpthread.so.0)
369
+ [default4]:frame #6: clone + 0x43 (0x7fb5e8251353 in /lib/x86_64-linux-gnu/libc.so.6)
370
+ [default4]:
371
+ [default4]:Exception raised from ncclCommWatchdog at ../torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1418 (most recent call first):
372
+ [default4]:frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fb5966c7897 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libc10.so)
373
+ [default4]:frame #1: <unknown function> + 0xe32119 (0x7fb59762a119 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/lib/libtorch_cuda.so)
374
+ [default4]:frame #2: <unknown function> + 0xd3e95 (0x7fb5e343fe95 in /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/../lib/libstdc++.so.6)
375
+ [default4]:frame #3: <unknown function> + 0x8609 (0x7fb5e8486609 in /lib/x86_64-linux-gnu/libpthread.so.0)
376
+ [default4]:frame #4: clone + 0x43 (0x7fb5e8251353 in /lib/x86_64-linux-gnu/libc.so.6)
377
+ [default4]:
378
+ W0704 00:00:10.677000 139684507649856 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 2654346 closing signal SIGTERM
379
+ W0704 00:00:10.677000 139684507649856 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 2654347 closing signal SIGTERM
380
+ W0704 00:00:10.677000 139684507649856 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 2654348 closing signal SIGTERM
381
+ W0704 00:00:10.678000 139684507649856 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 2654349 closing signal SIGTERM
382
+ W0704 00:00:10.681000 139684507649856 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 2654352 closing signal SIGTERM
383
+ W0704 00:00:10.682000 139684507649856 torch/distributed/elastic/multiprocessing/api.py:851] Sending process 2654353 closing signal SIGTERM
384
+ E0704 00:00:15.715000 139684507649856 torch/distributed/elastic/multiprocessing/api.py:826] failed (exitcode: -6) local_rank: 4 (pid: 2654350) of binary: /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/python3.10
385
+ Traceback (most recent call last):
386
+ File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/bin/torchrun", line 8, in <module>
387
+ sys.exit(main())
388
+ File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 347, in wrapper
389
+ return f(*args, **kwargs)
390
+ File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/run.py", line 879, in main
391
+ run(args)
392
+ File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/run.py", line 870, in run
393
+ elastic_launch(
394
+ File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 132, in __call__
395
+ return launch_agent(self._config, self._entrypoint, list(args))
396
+ File "/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 263, in launch_agent
397
+ raise ChildFailedError(
398
+ torch.distributed.elastic.multiprocessing.errors.ChildFailedError:
399
+ ============================================================
400
+ /fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron/run_train.py FAILED
401
+ ------------------------------------------------------------
402
+ Failures:
403
+ [1]:
404
+ time : 2024-07-04_00:00:10
405
+ host : ip-26-0-169-239.ec2.internal
406
+ rank : 5 (local_rank: 5)
407
+ exitcode : -6 (pid: 2654351)
408
+ error_file: <N/A>
409
+ traceback : Signal 6 (SIGABRT) received by PID 2654351
410
+ ------------------------------------------------------------
411
+ Root Cause (first observed failure):
412
+ [0]:
413
+ time : 2024-07-04_00:00:10
414
+ host : ip-26-0-169-239.ec2.internal
415
+ rank : 4 (local_rank: 4)
416
+ exitcode : -6 (pid: 2654350)
417
+ error_file: <N/A>
418
+ traceback : Signal 6 (SIGABRT) received by PID 2654350
419
+ ============================================================
420
+ srun: error: ip-26-0-169-239: task 0: Exited with exit code 1
421
+ 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.
422
+
llama-1B/8_GPUS/dp-2_tp-2_pp-2_mbz-2/profiler/ip-26-0-169-239_2654346.1720051209254005232.pt.trace.json.tmp ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c4bf0ed6460dcd283117685444e6e0e53d5c5529b991a877e8f554f44b739756
3
+ size 204976045
llama-1B/8_GPUS/dp-2_tp-2_pp-2_mbz-2/status.txt ADDED
@@ -0,0 +1 @@
 
 
1
+ oom