Hugging Face's logo

nanotron
/
old_bench

Model card Files Community
3outeille HF staff commited on
Commit
88a368c
•
1 Parent(s): 9f5a11e

Upload llama-1B/16_GPUS/dp-8_tp-1_pp-2_mbz-4

Browse files
Files changed (8) hide show
  1. .gitattributes +1 -0
  2. llama-1B/16_GPUS/dp-8_tp-1_pp-2_mbz-4/bench.slurm +111 -0
  3. llama-1B/16_GPUS/dp-8_tp-1_pp-2_mbz-4/config.yaml +90 -0
  4. llama-1B/16_GPUS/dp-8_tp-1_pp-2_mbz-4/log.out +320 -0
  5. llama-1B/16_GPUS/dp-8_tp-1_pp-2_mbz-4/log_metrics.csv +21 -0
  6. llama-1B/16_GPUS/dp-8_tp-1_pp-2_mbz-4/profiler.csv +2 -0
  7. llama-1B/16_GPUS/dp-8_tp-1_pp-2_mbz-4/profiler/ip-26-0-171-56_3064380.1719933802382173616.pt.trace.json +3 -0
  8. llama-1B/16_GPUS/dp-8_tp-1_pp-2_mbz-4/status.txt +1 -0
.gitattributes CHANGED
@@ -37,3 +37,4 @@ llama-1B/16_GPUS/dp-4_tp-4_pp-1_mbz-16/profiler/ip-26-0-161-178_137134.171992990
37
  llama-1B/16_GPUS/dp-1_tp-2_pp-8_mbz-8/profiler/ip-26-0-170-31_2724547.1719930375401132771.pt.trace.json filter=lfs diff=lfs merge=lfs -text
38
  llama-1B/16_GPUS/dp-1_tp-1_pp-16_mbz-1/profiler/ip-26-0-161-178_157959.1719931391546523089.pt.trace.json filter=lfs diff=lfs merge=lfs -text
39
  llama-1B/16_GPUS/dp-1_tp-2_pp-8_mbz-2/profiler/ip-26-0-163-134_1435141.1719931469116122110.pt.trace.json filter=lfs diff=lfs merge=lfs -text
 
 
37
  llama-1B/16_GPUS/dp-1_tp-2_pp-8_mbz-8/profiler/ip-26-0-170-31_2724547.1719930375401132771.pt.trace.json filter=lfs diff=lfs merge=lfs -text
38
  llama-1B/16_GPUS/dp-1_tp-1_pp-16_mbz-1/profiler/ip-26-0-161-178_157959.1719931391546523089.pt.trace.json filter=lfs diff=lfs merge=lfs -text
39
  llama-1B/16_GPUS/dp-1_tp-2_pp-8_mbz-2/profiler/ip-26-0-163-134_1435141.1719931469116122110.pt.trace.json filter=lfs diff=lfs merge=lfs -text
40
+ llama-1B/16_GPUS/dp-8_tp-1_pp-2_mbz-4/profiler/ip-26-0-171-56_3064380.1719933802382173616.pt.trace.json filter=lfs diff=lfs merge=lfs -text
llama-1B/16_GPUS/dp-8_tp-1_pp-2_mbz-4/bench.slurm ADDED
@@ -0,0 +1,111 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/bin/bash
2
+
3
+ #SBATCH --job-name=bench_cluster
4
+ #SBATCH --time=00:59:00
5
+ #SBATCH --partition=hopper-prod
6
+ #SBATCH --nodes=2
7
+ #SBATCH --gres=gpu:8
8
+ #SBATCH --qos=high
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/16_GPUS/dp-8_tp-1_pp-2_mbz-4/log.out
13
+ #SBATCH --error=/fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/16_GPUS/dp-8_tp-1_pp-2_mbz-4/log.out
14
+
15
+ # Function to update status based on squeue output
16
+ update_status() {
17
+ job_id=$1
18
+ status_file=$2
19
+ # For unknown reasons, it doenst update status for pending. It only works for running
20
+ while true; do
21
+ job_status=$(squeue --job $job_id --noheader --format=%T)
22
+ echo "Job status: $job_status"
23
+ if [ -z "$job_status" ]; then
24
+ # Job has finished or is not found
25
+ break
26
+ elif [ "$job_status" = "RUNNING" ]; then
27
+ printf "running" > $status_file
28
+ break
29
+ fi
30
+ sleep 10
31
+ done
32
+ }
33
+
34
+ # Misc initializations.
35
+ echo "========================"
36
+ echo "START TIME: $(date)"
37
+ source /fsx/ferdinandmom/miniforge3/etc/profile.d/conda.sh
38
+ conda activate /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster
39
+ echo python3 version = $(python3 --version)
40
+ echo "========================"
41
+
42
+ # Slurm stuff
43
+ export HOSTNAMES=$(scontrol show hostnames "$SLURM_JOB_NODELIST")
44
+ export MASTER_ADDR=$(scontrol show hostnames "$SLURM_JOB_NODELIST" | head -n 1)
45
+ export MASTER_PORT=$((1024 + RANDOM % 64511))
46
+
47
+ export TMPDIR=/scratch
48
+ export HF_DATASETS_CACHE="/admin/home/ferdinand_mom/.cache"
49
+ export CUBLAS_WORKSPACE_CONFIG=":4096:8"
50
+ export CUDA_DEVICE_MAX_CONNECTIONS="1"
51
+
52
+ huggingface-cli login --token $HUGGINGFACE_TOKEN
53
+
54
+
55
+ NANOTRON_REPO="/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron"
56
+ CMD="$NANOTRON_REPO/run_train.py --config-file /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/16_GPUS/dp-8_tp-1_pp-2_mbz-4/config.yaml"
57
+
58
+ LAUNCHER="torchrun \
59
+ --nproc_per_node 8 \
60
+ --nnodes 2 \
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/16_GPUS/dp-8_tp-1_pp-2_mbz-4/status.txt &
76
+
77
+ # Run the main command
78
+ srun -u $LAUNCHER $CMD
79
+ exit_status=$?
80
+
81
+ # Update status based on the exit status of `srun`
82
+ if [ $exit_status -eq 0 ]; then
83
+ printf "completed" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/16_GPUS/dp-8_tp-1_pp-2_mbz-4/status.txt
84
+ else
85
+ if grep -q "OutOfMemoryError" /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/16_GPUS/dp-8_tp-1_pp-2_mbz-4/log.out; then
86
+ printf "oom" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/16_GPUS/dp-8_tp-1_pp-2_mbz-4/status.txt
87
+ elif grep -q " CUDA error: an illegal memory access" /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/16_GPUS/dp-8_tp-1_pp-2_mbz-4/log.out; then
88
+ printf "oom" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/16_GPUS/dp-8_tp-1_pp-2_mbz-4/status.txt
89
+ elif grep -q "Timeout at NCCL" /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/16_GPUS/dp-8_tp-1_pp-2_mbz-4/log.out; then
90
+ printf "timeout" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/16_GPUS/dp-8_tp-1_pp-2_mbz-4/status.txt
91
+ else
92
+ printf "fail" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/16_GPUS/dp-8_tp-1_pp-2_mbz-4/status.txt
93
+ fi
94
+ fi
95
+
96
+ # Run the report script if the job completed successfully
97
+ if [ $exit_status -eq 0 ]; then
98
+ python /fsx/ferdinandmom/ferdinand-hf/bench_cluster/main.py report --inp_dir /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/16_GPUS/dp-8_tp-1_pp-2_mbz-4 --is_logs
99
+ python /fsx/ferdinandmom/ferdinand-hf/bench_cluster/main.py report --inp_dir /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/16_GPUS/dp-8_tp-1_pp-2_mbz-4 --is_profiler
100
+ fi
101
+
102
+
103
+ # Push to hub the folder using huggingface_cli
104
+ huggingface-cli upload nanotron/bench_cluster /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/16_GPUS/dp-8_tp-1_pp-2_mbz-4 llama-1B/16_GPUS/dp-8_tp-1_pp-2_mbz-4 --commit-message "Upload llama-1B/16_GPUS/dp-8_tp-1_pp-2_mbz-4"
105
+
106
+ # Verify the upload
107
+ if [ $? -eq 0 ]; then
108
+ echo "Uploading to Huggingface Hub successful"
109
+ else
110
+ echo "Failed to upload to Huggingface Hub"
111
+ fi
llama-1B/16_GPUS/dp-8_tp-1_pp-2_mbz-4/config.yaml ADDED
@@ -0,0 +1,90 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ general:
2
+ project: bench_cluster
3
+ seed: 42
4
+ model:
5
+ ddp_bucket_cap_mb: 25
6
+ dtype: bfloat16
7
+ init_method:
8
+ std: 0.025
9
+ make_vocab_size_divisible_by: 1
10
+ model_config:
11
+ bos_token_id: 1
12
+ eos_token_id: 2
13
+ hidden_act: silu
14
+ hidden_size: 2048
15
+ initializer_range: 0.02
16
+ intermediate_size: 4096
17
+ is_llama_config: true
18
+ max_position_embeddings: 4096
19
+ num_attention_heads: 32
20
+ num_hidden_layers: 24
21
+ num_key_value_heads: 32
22
+ pad_token_id: null
23
+ pretraining_tp: 1
24
+ rms_norm_eps: 1.0e-05
25
+ rope_scaling: null
26
+ rope_theta: 10000.0
27
+ tie_word_embeddings: true
28
+ use_cache: true
29
+ vocab_size: 50257
30
+ optimizer:
31
+ accumulate_grad_in_fp32: true
32
+ clip_grad: 1.0
33
+ learning_rate_scheduler:
34
+ learning_rate: 0.0001
35
+ lr_decay_style: linear
36
+ lr_warmup_style: linear
37
+ lr_warmup_steps: 1
38
+ min_decay_lr: 1.0e-05
39
+ optimizer_factory:
40
+ adam_beta1: 0.9
41
+ adam_beta2: 0.95
42
+ adam_eps: 1.0e-08
43
+ name: adamW
44
+ torch_adam_is_fused: true
45
+ weight_decay: 0.01
46
+ zero_stage: 1
47
+ parallelism:
48
+ dp: 8
49
+ expert_parallel_size: 1
50
+ pp: 2
51
+ pp_engine: 1f1b
52
+ tp: 1
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/16_GPUS/dp-8_tp-1_pp-2_mbz-4
57
+ tokenizer:
58
+ tokenizer_max_length: null
59
+ tokenizer_name_or_path: openai-community/gpt2
60
+ tokenizer_revision: null
61
+ data_stages:
62
+ - name: Training Stage
63
+ start_training_step: 1
64
+ data:
65
+ dataset:
66
+ dataset_overwrite_cache: false
67
+ dataset_processing_num_proc_per_process: 64
68
+ hf_dataset_config_name: null
69
+ hf_dataset_or_datasets: roneneldan/TinyStories
70
+ hf_dataset_splits: train
71
+ text_column_name: text
72
+ num_loading_workers: 32
73
+ seed: 42
74
+ lighteval: null
75
+ tokens:
76
+ train_steps: 20
77
+ val_check_interval: -1
78
+ batch_accumulation_per_replica: 32
79
+ limit_test_batches: 0
80
+ limit_val_batches: 0
81
+ micro_batch_size: 4
82
+ sequence_length: 4096
83
+ logging:
84
+ iteration_step_info_interval: 1
85
+ log_level: info
86
+ log_level_replica: info
87
+ checkpoints:
88
+ checkpoint_interval: 100000
89
+ checkpoints_path: /dev/null
90
+ resume_checkpoint_path: null
llama-1B/16_GPUS/dp-8_tp-1_pp-2_mbz-4/log.out ADDED
@@ -0,0 +1,320 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ========================
2
+ START TIME: Tue Jul 2 15:19:45 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
+ W0702 15:19:47.592000 140244399671104 torch/distributed/run.py:757]
18
+ W0702 15:19:47.592000 140244399671104 torch/distributed/run.py:757] *****************************************
19
+ W0702 15:19:47.592000 140244399671104 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
+ W0702 15:19:47.592000 140244399671104 torch/distributed/run.py:757] *****************************************
21
+ W0702 15:19:47.605000 140705257113408 torch/distributed/run.py:757]
22
+ W0702 15:19:47.605000 140705257113408 torch/distributed/run.py:757] *****************************************
23
+ W0702 15:19:47.605000 140705257113408 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.
24
+ W0702 15:19:47.605000 140705257113408 torch/distributed/run.py:757] *****************************************
25
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: Config:
26
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: Config(general=GeneralArgs(project='bench_cluster',
27
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: run='%date_%jobid',
28
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: seed=42,
29
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: step=None,
30
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: consumed_train_samples=None,
31
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: benchmark_csv_path=None,
32
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: ignore_sanity_checks=True),
33
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: parallelism=ParallelismArgs(dp=8,
34
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: pp=2,
35
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: tp=1,
36
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: pp_engine=<nanotron.parallel.pipeline_parallel.engine.OneForwardOneBackwardPipelineEngine object at 0x7f08f97dc910>,
37
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: tp_mode=<TensorParallelLinearMode.REDUCE_SCATTER: 2>,
38
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: tp_linear_async_communication=False,
39
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: expert_parallel_size=1),
40
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: model=ModelArgs(model_config=LlamaConfig(bos_token_id=1,
41
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: eos_token_id=2,
42
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: hidden_act='silu',
43
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: hidden_size=2048,
44
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: initializer_range=0.02,
45
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: intermediate_size=4096,
46
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: is_llama_config=True,
47
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: max_position_embeddings=4096,
48
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: num_attention_heads=32,
49
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: num_hidden_layers=24,
50
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: num_key_value_heads=32,
51
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: pad_token_id=None,
52
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: pretraining_tp=1,
53
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: rms_norm_eps=1e-05,
54
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: rope_scaling=None,
55
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: rope_theta=10000.0,
56
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: tie_word_embeddings=True,
57
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: use_cache=True,
58
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: vocab_size=50257),
59
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: init_method=RandomInit(std=0.025),
60
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: dtype=torch.bfloat16,
61
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: make_vocab_size_divisible_by=1,
62
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: ddp_bucket_cap_mb=25),
63
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: tokenizer=TokenizerArgs(tokenizer_name_or_path='openai-community/gpt2',
64
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: tokenizer_revision=None,
65
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: tokenizer_max_length=None),
66
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: checkpoints=CheckpointsArgs(checkpoints_path=Path('/dev/null'),
67
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: checkpoint_interval=100000,
68
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: save_initial_state=False,
69
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: resume_checkpoint_path=None,
70
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: checkpoints_path_is_shared_file_system=False),
71
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: logging=LoggingArgs(log_level='info',
72
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: log_level_replica='info',
73
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: iteration_step_info_interval=1),
74
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: tokens=TokensArgs(sequence_length=4096,
75
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: train_steps=20,
76
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: micro_batch_size=4,
77
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: batch_accumulation_per_replica=32,
78
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: val_check_interval=-1,
79
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: limit_val_batches=0,
80
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: limit_test_batches=0),
81
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: optimizer=OptimizerArgs(optimizer_factory=AdamWOptimizerArgs(adam_eps=1e-08,
82
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: adam_beta1=0.9,
83
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: adam_beta2=0.95,
84
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: torch_adam_is_fused=True,
85
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: name='adamW'),
86
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: zero_stage=1,
87
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: weight_decay=0.01,
88
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: clip_grad=1.0,
89
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: accumulate_grad_in_fp32=True,
90
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: learning_rate_scheduler=LRSchedulerArgs(learning_rate=0.0001,
91
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: lr_warmup_steps=1,
92
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: lr_warmup_style='linear',
93
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: lr_decay_style='linear',
94
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: lr_decay_steps=19,
95
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: lr_decay_starting_step=None,
96
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: min_decay_lr=1e-05)),
97
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: data_stages=[DatasetStageArgs(name='Training Stage',
98
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: start_training_step=1,
99
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: data=DataArgs(dataset=PretrainDatasetsArgs(hf_dataset_or_datasets='roneneldan/TinyStories',
100
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: hf_dataset_splits='train',
101
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: hf_dataset_config_name=None,
102
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: dataset_processing_num_proc_per_process=64,
103
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: dataset_overwrite_cache=False,
104
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: text_column_name='text'),
105
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: seed=42,
106
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: num_loading_workers=32))],
107
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: profiler=ProfilerArgs(profiler_export_path=Path('/fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/16_GPUS/dp-8_tp-1_pp-2_mbz-4')),
108
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: lighteval=None)
109
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: Model Config:
110
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: LlamaConfig(bos_token_id=1,
111
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: eos_token_id=2,
112
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: hidden_act='silu',
113
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: hidden_size=2048,
114
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: initializer_range=0.02,
115
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: intermediate_size=4096,
116
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: is_llama_config=True,
117
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: max_position_embeddings=4096,
118
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: num_attention_heads=32,
119
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: num_hidden_layers=24,
120
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: num_key_value_heads=32,
121
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: pad_token_id=None,
122
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: pretraining_tp=1,
123
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: rms_norm_eps=1e-05,
124
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: rope_scaling=None,
125
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: rope_theta=10000.0,
126
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: tie_word_embeddings=True,
127
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: use_cache=True,
128
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: vocab_size=50257)
129
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: Building model..
130
+ [default0]:07/02/2024 15:20:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: Setting PP block ranks...
131
+ [default2]:07/02/2024 15:20:16 [INFO|DP=2|PP=0|TP=0|ip-26-0-171-56]: No checkpoint path provided.
132
+ [default0]:07/02/2024 15:20:16 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: Total number of parameters: 1.21G (2312.82MiB)
133
+ [default0]:07/02/2024 15:20:16 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: Local number of parameters: 690M (1316.43MiB)
134
+ [default0]:07/02/2024 15:20:16 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: [After model building] Memory usage: 1330.44MiB. Peak allocated: 1332.47MiB Peak reserved: 1364.00MiB
135
+ [default0]:07/02/2024 15:20:16 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: No checkpoint path provided.
136
+ [default0]:07/02/2024 15:20:16 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: Parametrizing model parameters using StandardParametrizator
137
+ [default0]:07/02/2024 15:20:16 [INFO|DP=0|PP=1|TP=0|ip-26-0-175-132]: Local number of parameters: 522M (996.40MiB)
138
+ [default0]:07/02/2024 15:20:16 [INFO|DP=0|PP=1|TP=0|ip-26-0-175-132]: [After model building] Memory usage: 1006.41MiB. Peak allocated: 1008.44MiB Peak reserved: 1032.00MiB
139
+ [default0]:07/02/2024 15:20:16 [INFO|DP=0|PP=1|TP=0|ip-26-0-175-132]: No checkpoint path provided.
140
+ [default2]:07/02/2024 15:20:16 [INFO|DP=2|PP=1|TP=0|ip-26-0-175-132]: No checkpoint path provided.
141
+ [default3]:07/02/2024 15:20:16 [INFO|DP=3|PP=1|TP=0|ip-26-0-175-132]: No checkpoint path provided.
142
+ [default3]:07/02/2024 15:20:16 [INFO|DP=3|PP=0|TP=0|ip-26-0-171-56]: No checkpoint path provided.
143
+ [default1]:07/02/2024 15:20:16 [INFO|DP=1|PP=1|TP=0|ip-26-0-175-132]: No checkpoint path provided.
144
+ [default1]:07/02/2024 15:20:16 [INFO|DP=1|PP=0|TP=0|ip-26-0-171-56]: No checkpoint path provided.
145
+ [default6]:07/02/2024 15:20:16 [INFO|DP=6|PP=0|TP=0|ip-26-0-171-56]: No checkpoint path provided.
146
+ [default4]:07/02/2024 15:20:16 [INFO|DP=4|PP=0|TP=0|ip-26-0-171-56]: No checkpoint path provided.
147
+ [default5]:07/02/2024 15:20:16 [INFO|DP=5|PP=0|TP=0|ip-26-0-171-56]: No checkpoint path provided.
148
+ [default7]:07/02/2024 15:20:16 [INFO|DP=7|PP=0|TP=0|ip-26-0-171-56]: No checkpoint path provided.
149
+ [default7]:07/02/2024 15:20:16 [INFO|DP=7|PP=1|TP=0|ip-26-0-175-132]: No checkpoint path provided.
150
+ [default5]:07/02/2024 15:20:16 [INFO|DP=5|PP=1|TP=0|ip-26-0-175-132]: No checkpoint path provided.
151
+ [default4]:07/02/2024 15:20:16 [INFO|DP=4|PP=1|TP=0|ip-26-0-175-132]: No checkpoint path provided.
152
+ [default6]:07/02/2024 15:20:16 [INFO|DP=6|PP=1|TP=0|ip-26-0-175-132]: No checkpoint path provided.
153
+ [default0]:07/02/2024 15:20:23 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: [Optimizer Building] Using LearningRateForSP as learning rate
154
+ [default0]:07/02/2024 15:20:23 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: [ZeRO sharding] Size of optimizer params per rank:
155
+ [default0]:07/02/2024 15:20:23 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: [ZeRO sharding] DP Rank 0 has 86.3M out of 690M (12.50%) params' optimizer states
156
+ [default0]:07/02/2024 15:20:23 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: [ZeRO sharding] DP Rank 1 has 86.3M out of 690M (12.50%) params' optimizer states
157
+ [default0]:07/02/2024 15:20:23 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: [ZeRO sharding] DP Rank 2 has 86.3M out of 690M (12.50%) params' optimizer states
158
+ [default0]:07/02/2024 15:20:23 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: [ZeRO sharding] DP Rank 3 has 86.3M out of 690M (12.50%) params' optimizer states
159
+ [default0]:07/02/2024 15:20:23 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: [ZeRO sharding] DP Rank 4 has 86.3M out of 690M (12.50%) params' optimizer states
160
+ [default0]:07/02/2024 15:20:23 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: [ZeRO sharding] DP Rank 5 has 86.3M out of 690M (12.50%) params' optimizer states
161
+ [default0]:07/02/2024 15:20:23 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: [ZeRO sharding] DP Rank 6 has 86.3M out of 690M (12.50%) params' optimizer states
162
+ [default0]:07/02/2024 15:20:23 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: [ZeRO sharding] DP Rank 7 has 86.3M out of 690M (12.50%) params' optimizer states
163
+ [default0]:07/02/2024 15:20:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: [Training Plan] Stage Training Stage has 19 remaining training steps and has consumed 0 samples
164
+ [default0]:07/02/2024 15:20:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: Using `datasets` library
165
+ [default0]:07/02/2024 15:20:25 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: Loading tokenizer from openai-community/gpt2 and transformers/hf_hub versions ('4.41.2', '0.23.4')
166
+ [default0]:Repo card metadata block was not found. Setting CardData to empty.
167
+ [default0]:07/02/2024 15:20:25 [WARNING|DP=0|PP=0|TP=0|ip-26-0-171-56]: Repo card metadata block was not found. Setting CardData to empty.
168
+ [default0]:07/02/2024 15:20:26 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: [Training Plan] There are 1 training stages
169
+ [default0]:07/02/2024 15:20:26 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: [Stage Training Stage] start from step 1
170
+ [default0]:07/02/2024 15:20:26 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]:
171
+ [default0]:07/02/2024 15:20:26 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: [Start training] datetime: 2024-07-02 15:20:26.536278 | mbs: 4 | grad_accum: 32 | global_batch_size: 1024 | sequence_length: 4096 | train_steps: 20 | start_iteration_step: 0 | consumed_train_samples: 0
172
+ [default0]:07/02/2024 15:20:26 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: Resuming training from stage Training Stage, it has trained for 0 samples and has 19 remaining train steps
173
+ [default0]:07/02/2024 15:20:26 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: Memory usage: 4293.29MiB. Peak allocated 4293.29MiB. Peak reserved: 4328.00MiB
174
+ [default5]:07/02/2024 15:20:26 [WARNING|DP=5|PP=0|TP=0|ip-26-0-171-56]: Repo card metadata block was not found. Setting CardData to empty.
175
+ [default4]:07/02/2024 15:20:26 [WARNING|DP=4|PP=0|TP=0|ip-26-0-171-56]: Repo card metadata block was not found. Setting CardData to empty.
176
+ [default6]:07/02/2024 15:20:26 [WARNING|DP=6|PP=0|TP=0|ip-26-0-171-56]: Repo card metadata block was not found. Setting CardData to empty.
177
+ [default1]:Repo card metadata block was not found. Setting CardData to empty.
178
+ [default5]:Repo card metadata block was not found. Setting CardData to empty.
179
+ [default0]:07/02/2024 15:20:26 [WARNING|DP=0|PP=1|TP=0|ip-26-0-175-132]: Repo card metadata block was not found. Setting CardData to empty.
180
+ [default4]:Repo card metadata block was not found. Setting CardData to empty.
181
+ [default1]:07/02/2024 15:20:26 [WARNING|DP=1|PP=0|TP=0|ip-26-0-171-56]: Repo card metadata block was not found. Setting CardData to empty.
182
+ [default6]:07/02/2024 15:20:26 [WARNING|DP=6|PP=1|TP=0|ip-26-0-175-132]: Repo card metadata block was not found. Setting CardData to empty.
183
+ [default6]:Repo card metadata block was not found. Setting CardData to empty.
184
+ [default7]:07/02/2024 15:20:26 [WARNING|DP=7|PP=1|TP=0|ip-26-0-175-132]: Repo card metadata block was not found. Setting CardData to empty.
185
+ [default2]:07/02/2024 15:20:26 [WARNING|DP=2|PP=1|TP=0|ip-26-0-175-132]: Repo card metadata block was not found. Setting CardData to empty.
186
+ [default1]:Repo card metadata block was not found. Setting CardData to empty.
187
+ [default5]:07/02/2024 15:20:26 [WARNING|DP=5|PP=1|TP=0|ip-26-0-175-132]: Repo card metadata block was not found. Setting CardData to empty.
188
+ [default4]:07/02/2024 15:20:26 [WARNING|DP=4|PP=1|TP=0|ip-26-0-175-132]: Repo card metadata block was not found. Setting CardData to empty.
189
+ [default7]:Repo card metadata block was not found. Setting CardData to empty.
190
+ [default1]:07/02/2024 15:20:26 [WARNING|DP=1|PP=1|TP=0|ip-26-0-175-132]: Repo card metadata block was not found. Setting CardData to empty.
191
+ [default0]:Repo card metadata block was not found. Setting CardData to empty.
192
+ [default3]:07/02/2024 15:20:26 [WARNING|DP=3|PP=1|TP=0|ip-26-0-175-132]: Repo card metadata block was not found. Setting CardData to empty.
193
+ [default6]:Repo card metadata block was not found. Setting CardData to empty.
194
+ [default3]:Repo card metadata block was not found. Setting CardData to empty.
195
+ [default4]:Repo card metadata block was not found. Setting CardData to empty.
196
+ [default2]:Repo card metadata block was not found. Setting CardData to empty.
197
+ [default5]:Repo card metadata block was not found. Setting CardData to empty.
198
+ [default2]:07/02/2024 15:20:26 [WARNING|DP=2|PP=0|TP=0|ip-26-0-171-56]: Repo card metadata block was not found. Setting CardData to empty.
199
+ [default3]:07/02/2024 15:20:26 [WARNING|DP=3|PP=0|TP=0|ip-26-0-171-56]: Repo card metadata block was not found. Setting CardData to empty.
200
+ [default7]:Repo card metadata block was not found. Setting CardData to empty.
201
+ [default7]:07/02/2024 15:20:26 [WARNING|DP=7|PP=0|TP=0|ip-26-0-171-56]: Repo card metadata block was not found. Setting CardData to empty.
202
+ [default3]:Repo card metadata block was not found. Setting CardData to empty.
203
+ [default2]:Repo card metadata block was not found. Setting CardData to empty.
204
+ [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.)
205
+ [default0]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
206
+ [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.)
207
+ [default4]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
208
+ [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.)
209
+ [default6]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
210
+ [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.)
211
+ [default1]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
212
+ [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.)
213
+ [default3]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
214
+ [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.)
215
+ [default2]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
216
+ [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.)
217
+ [default5]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
218
+ [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.)
219
+ [default7]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
220
+ [default0]:/fsx/ferdinandmom/miniforge3/envs/env-bench-cluster/lib/python3.10/site-packages/torch/autograd/graph.py:744: UserWarning: Attempting to run cuBLAS, but there was no current CUDA context! Attempting to set the primary context... (Triggered internally at ../aten/src/ATen/cuda/CublasHandlePool.cpp:135.)
221
+ [default0]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
222
+ [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.)
223
+ [default0]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
224
+ [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.)
225
+ [default6]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
226
+ [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.)
227
+ [default3]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
228
+ [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.)
229
+ [default4]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
230
+ [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.)
231
+ [default1]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
232
+ [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.)
233
+ [default5]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
234
+ [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.)
235
+ [default2]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
236
+ [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.)
237
+ [default7]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
238
+ [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
239
+ [default6]: warnings.warn(
240
+ [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
241
+ [default6]: warnings.warn(
242
+ [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
243
+ [default4]: warnings.warn(
244
+ [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
245
+ [default4]: warnings.warn(
246
+ [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
247
+ [default0]: warnings.warn(
248
+ [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
249
+ [default0]: warnings.warn(
250
+ [default0]:07/02/2024 15:20:51 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: Memory usage: 4364.43MiB. Peak allocated 38808.12MiB. Peak reserved: 39078.00MiB
251
+ [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
252
+ [default1]: warnings.warn(
253
+ [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
254
+ [default1]: warnings.warn(
255
+ [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
256
+ [default5]: warnings.warn(
257
+ [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
258
+ [default5]: warnings.warn(
259
+ [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
260
+ [default7]: warnings.warn(
261
+ [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
262
+ [default7]: warnings.warn(
263
+ [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
264
+ [default3]: warnings.warn(
265
+ [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
266
+ [default3]: warnings.warn(
267
+ [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
268
+ [default2]: warnings.warn(
269
+ [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
270
+ [default2]: warnings.warn(
271
+ [default0]:07/02/2024 15:20:54 [INFO|DP=0|PP=1|TP=0|ip-26-0-175-132]: iteration: 1 / 20 | consumed_tokens: 4.19M | elapsed_time_per_iteration_ms: 22.1K | tokens_per_sec: 190K | tokens_per_sec_per_gpu: 11.9K | global_batch_size: 1.02K | lm_loss: 11.1 | lr: 0.0001 | model_tflops_per_gpu: 108 | hardware_tflops_per_gpu: 108 | grad_norm: 24.6 | cuda_memory_allocated: 4G | cuda_max_memory_reserved: 24.7G | hd_total_memory_tb: 312G | hd_used_memory_tb: 65.5G | hd_free_memory_tb: 247G
272
+ [default0]:07/02/2024 15:20:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: Memory usage: 5024.53MiB. Peak allocated 7821.94MiB. Peak reserved: 40396.00MiB
273
+ [default0]:07/02/2024 15:21:07 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: Memory usage: 5024.53MiB. Peak allocated 39468.21MiB. Peak reserved: 40396.00MiB
274
+ [default0]:07/02/2024 15:21:08 [INFO|DP=0|PP=1|TP=0|ip-26-0-175-132]: iteration: 2 / 20 | consumed_tokens: 8.39M | elapsed_time_per_iteration_ms: 13.9K | tokens_per_sec: 302K | tokens_per_sec_per_gpu: 18.9K | global_batch_size: 1.02K | lm_loss: 11.1 | lr: 9.53e-05 | model_tflops_per_gpu: 172 | hardware_tflops_per_gpu: 172 | grad_norm: 24.8 | cuda_memory_allocated: 4G | cuda_max_memory_reserved: 24.7G | hd_total_memory_tb: 312G | hd_used_memory_tb: 65.5G | hd_free_memory_tb: 247G
275
+ [default0]:07/02/2024 15:21:08 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: Memory usage: 5024.53MiB. Peak allocated 7821.94MiB. Peak reserved: 40396.00MiB
276
+ [default0]:07/02/2024 15:21:20 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: Memory usage: 5024.53MiB. Peak allocated 39468.21MiB. Peak reserved: 40396.00MiB
277
+ [default0]:07/02/2024 15:21:21 [INFO|DP=0|PP=1|TP=0|ip-26-0-175-132]: iteration: 3 / 20 | consumed_tokens: 12.6M | elapsed_time_per_iteration_ms: 13.6K | tokens_per_sec: 309K | tokens_per_sec_per_gpu: 19.3K | global_batch_size: 1.02K | lm_loss: 10.5 | lr: 9.05e-05 | model_tflops_per_gpu: 175 | hardware_tflops_per_gpu: 175 | grad_norm: 197 | cuda_memory_allocated: 4G | cuda_max_memory_reserved: 24.7G | hd_total_memory_tb: 312G | hd_used_memory_tb: 65.5G | hd_free_memory_tb: 247G
278
+ [default0]:07/02/2024 15:21:21 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: Memory usage: 5024.53MiB. Peak allocated 7821.94MiB. Peak reserved: 40396.00MiB
279
+ [default0]:STAGE:2024-07-02 15:21:21 3064380:3064380 ActivityProfilerController.cpp:314] Completed Stage: Warm Up
280
+ [default0]:07/02/2024 15:21:34 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: Memory usage: 5024.53MiB. Peak allocated 39468.21MiB. Peak reserved: 40396.00MiB
281
+ [default0]:07/02/2024 15:21:36 [INFO|DP=0|PP=1|TP=0|ip-26-0-175-132]: iteration: 4 / 20 | consumed_tokens: 16.8M | elapsed_time_per_iteration_ms: 14.1K | tokens_per_sec: 297K | tokens_per_sec_per_gpu: 18.5K | global_batch_size: 1.02K | lm_loss: 13.9 | lr: 8.58e-05 | model_tflops_per_gpu: 168 | hardware_tflops_per_gpu: 168 | grad_norm: 17.9 | cuda_memory_allocated: 4G | cuda_max_memory_reserved: 24.7G | hd_total_memory_tb: 312G | hd_used_memory_tb: 65.5G | hd_free_memory_tb: 247G
282
+ [default0]:07/02/2024 15:21:36 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: Memory usage: 5024.53MiB. Peak allocated 7821.94MiB. Peak reserved: 40396.00MiB
283
+ [default0]:07/02/2024 15:21:50 [INFO|DP=0|PP=1|TP=0|ip-26-0-175-132]: iteration: 5 / 20 | consumed_tokens: 21M | elapsed_time_per_iteration_ms: 14.7K | tokens_per_sec: 285K | tokens_per_sec_per_gpu: 17.8K | global_batch_size: 1.02K | lm_loss: 9.71 | lr: 8.11e-05 | model_tflops_per_gpu: 162 | hardware_tflops_per_gpu: 162 | grad_norm: 20.2
284
+ [default0]:07/02/2024 15:21:50 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: Memory usage: 5024.53MiB. Peak allocated 39468.21MiB. Peak reserved: 40396.00MiB
285
+ [default0]:07/02/2024 15:22:05 [INFO|DP=0|PP=1|TP=0|ip-26-0-175-132]: iteration: 6 / 20 | consumed_tokens: 25.2M | elapsed_time_per_iteration_ms: 14.3K | tokens_per_sec: 293K | tokens_per_sec_per_gpu: 18.3K | global_batch_size: 1.02K | lm_loss: 13.7 | lr: 7.63e-05 | model_tflops_per_gpu: 166 | hardware_tflops_per_gpu: 166 | grad_norm: 98.3
286
+ [default0]:STAGE:2024-07-02 15:22:17 3064380:3064380 ActivityProfilerController.cpp:320] Completed Stage: Collection
287
+ [default0]:STAGE:2024-07-02 15:22:18 3064380:3064380 ActivityProfilerController.cpp:324] Completed Stage: Post Processing
288
+ [default0]:07/02/2024 15:23:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: Memory usage: 5024.53MiB. Peak allocated 39468.21MiB. Peak reserved: 40396.00MiB
289
+ [default0]:07/02/2024 15:23:54 [INFO|DP=0|PP=1|TP=0|ip-26-0-175-132]: iteration: 7 / 20 | consumed_tokens: 29.4M | elapsed_time_per_iteration_ms: 110K | tokens_per_sec: 38.3K | tokens_per_sec_per_gpu: 2.39K | global_batch_size: 1.02K | lm_loss: 9.73 | lr: 7.16e-05 | model_tflops_per_gpu: 21.7 | hardware_tflops_per_gpu: 21.7 | grad_norm: 12.7
290
+ [default0]:07/02/2024 15:23:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: Memory usage: 5024.53MiB. Peak allocated 39468.21MiB. Peak reserved: 40396.00MiB
291
+ [default0]:07/02/2024 15:24:09 [INFO|DP=0|PP=1|TP=0|ip-26-0-175-132]: iteration: 8 / 20 | consumed_tokens: 33.6M | elapsed_time_per_iteration_ms: 15K | tokens_per_sec: 279K | tokens_per_sec_per_gpu: 17.4K | global_batch_size: 1.02K | lm_loss: 8.94 | lr: 6.68e-05 | model_tflops_per_gpu: 158 | hardware_tflops_per_gpu: 158 | grad_norm: 11.9
292
+ [default0]:07/02/2024 15:24:09 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: Memory usage: 5024.53MiB. Peak allocated 39468.21MiB. Peak reserved: 40396.00MiB
293
+ [default0]:07/02/2024 15:24:23 [INFO|DP=0|PP=1|TP=0|ip-26-0-175-132]: iteration: 9 / 20 | consumed_tokens: 37.7M | elapsed_time_per_iteration_ms: 14.2K | tokens_per_sec: 295K | tokens_per_sec_per_gpu: 18.4K | global_batch_size: 1.02K | lm_loss: 8.51 | lr: 6.21e-05 | model_tflops_per_gpu: 167 | hardware_tflops_per_gpu: 167 | grad_norm: 5.69
294
+ [default0]:07/02/2024 15:24:23 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: Memory usage: 5024.53MiB. Peak allocated 39468.21MiB. Peak reserved: 40396.00MiB
295
+ [default0]:07/02/2024 15:24:37 [INFO|DP=0|PP=1|TP=0|ip-26-0-175-132]: iteration: 10 / 20 | consumed_tokens: 41.9M | elapsed_time_per_iteration_ms: 14K | tokens_per_sec: 299K | tokens_per_sec_per_gpu: 18.7K | global_batch_size: 1.02K | lm_loss: 8.14 | lr: 5.74e-05 | model_tflops_per_gpu: 170 | hardware_tflops_per_gpu: 170 | grad_norm: 4.25
296
+ [default0]:07/02/2024 15:24:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: Memory usage: 5024.53MiB. Peak allocated 39468.21MiB. Peak reserved: 40396.00MiB
297
+ [default0]:07/02/2024 15:24:52 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: Memory usage: 5024.53MiB. Peak allocated 39468.21MiB. Peak reserved: 40396.00MiB
298
+ [default0]:07/02/2024 15:24:52 [INFO|DP=0|PP=1|TP=0|ip-26-0-175-132]: iteration: 11 / 20 | consumed_tokens: 46.1M | elapsed_time_per_iteration_ms: 14.3K | tokens_per_sec: 292K | tokens_per_sec_per_gpu: 18.3K | global_batch_size: 1.02K | lm_loss: 8.01 | lr: 5.26e-05 | model_tflops_per_gpu: 166 | hardware_tflops_per_gpu: 166 | grad_norm: 7.09
299
+ [default0]:07/02/2024 15:25:07 [INFO|DP=0|PP=1|TP=0|ip-26-0-175-132]: iteration: 12 / 20 | consumed_tokens: 50.3M | elapsed_time_per_iteration_ms: 15.4K | tokens_per_sec: 272K | tokens_per_sec_per_gpu: 17K | global_batch_size: 1.02K | lm_loss: 7.81 | lr: 4.79e-05 | model_tflops_per_gpu: 155 | hardware_tflops_per_gpu: 155 | grad_norm: 6.25
300
+ [default0]:07/02/2024 15:25:07 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: Memory usage: 5024.53MiB. Peak allocated 39468.21MiB. Peak reserved: 40396.00MiB
301
+ [default0]:07/02/2024 15:25:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: Memory usage: 5024.53MiB. Peak allocated 39468.21MiB. Peak reserved: 40396.00MiB
302
+ [default0]:07/02/2024 15:25:22 [INFO|DP=0|PP=1|TP=0|ip-26-0-175-132]: iteration: 13 / 20 | consumed_tokens: 54.5M | elapsed_time_per_iteration_ms: 14.6K | tokens_per_sec: 287K | tokens_per_sec_per_gpu: 18K | global_batch_size: 1.02K | lm_loss: 7.6 | lr: 4.32e-05 | model_tflops_per_gpu: 163 | hardware_tflops_per_gpu: 163 | grad_norm: 2.65
303
+ [default0]:07/02/2024 15:25:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: Memory usage: 5024.53MiB. Peak allocated 39468.21MiB. Peak reserved: 40396.00MiB
304
+ [default0]:07/02/2024 15:25:37 [INFO|DP=0|PP=1|TP=0|ip-26-0-175-132]: iteration: 14 / 20 | consumed_tokens: 58.7M | elapsed_time_per_iteration_ms: 14.8K | tokens_per_sec: 283K | tokens_per_sec_per_gpu: 17.7K | global_batch_size: 1.02K | lm_loss: 7.53 | lr: 3.84e-05 | model_tflops_per_gpu: 161 | hardware_tflops_per_gpu: 161 | grad_norm: 3.82
305
+ [default0]:07/02/2024 15:25:50 [INFO|DP=0|PP=1|TP=0|ip-26-0-175-132]: iteration: 15 / 20 | consumed_tokens: 62.9M | elapsed_time_per_iteration_ms: 13.8K | tokens_per_sec: 305K | tokens_per_sec_per_gpu: 19.1K | global_batch_size: 1.02K | lm_loss: 7.42 | lr: 3.37e-05 | model_tflops_per_gpu: 173 | hardware_tflops_per_gpu: 173 | grad_norm: 3.42
306
+ [default0]:07/02/2024 15:25:50 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: Memory usage: 5024.53MiB. Peak allocated 39468.21MiB. Peak reserved: 40396.00MiB
307
+ [default0]:07/02/2024 15:26:06 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: Memory usage: 5024.53MiB. Peak allocated 39468.21MiB. Peak reserved: 40396.00MiB
308
+ [default0]:07/02/2024 15:26:06 [INFO|DP=0|PP=1|TP=0|ip-26-0-175-132]: iteration: 16 / 20 | consumed_tokens: 67.1M | elapsed_time_per_iteration_ms: 15.5K | tokens_per_sec: 271K | tokens_per_sec_per_gpu: 17K | global_batch_size: 1.02K | lm_loss: 7.31 | lr: 2.89e-05 | model_tflops_per_gpu: 154 | hardware_tflops_per_gpu: 154 | grad_norm: 2.89
309
+ [default0]:07/02/2024 15:26:20 [INFO|DP=0|PP=1|TP=0|ip-26-0-175-132]: iteration: 17 / 20 | consumed_tokens: 71.3M | elapsed_time_per_iteration_ms: 14.3K | tokens_per_sec: 294K | tokens_per_sec_per_gpu: 18.4K | global_batch_size: 1.02K | lm_loss: 7.21 | lr: 2.42e-05 | model_tflops_per_gpu: 167 | hardware_tflops_per_gpu: 167 | grad_norm: 2.21
310
+ [default0]:07/02/2024 15:26:20 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: Memory usage: 5024.53MiB. Peak allocated 39468.21MiB. Peak reserved: 40396.00MiB
311
+ [default0]:07/02/2024 15:26:36 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: Memory usage: 5024.53MiB. Peak allocated 39468.21MiB. Peak reserved: 40396.00MiB
312
+ [default0]:07/02/2024 15:26:36 [INFO|DP=0|PP=1|TP=0|ip-26-0-175-132]: iteration: 18 / 20 | consumed_tokens: 75.5M | elapsed_time_per_iteration_ms: 15.7K | tokens_per_sec: 267K | tokens_per_sec_per_gpu: 16.7K | global_batch_size: 1.02K | lm_loss: 7.15 | lr: 1.95e-05 | model_tflops_per_gpu: 152 | hardware_tflops_per_gpu: 152 | grad_norm: 2.57
313
+ [default0]:07/02/2024 15:26:51 [INFO|DP=0|PP=0|TP=0|ip-26-0-171-56]: Memory usage: 5024.53MiB. Peak allocated 39468.21MiB. Peak reserved: 40396.00MiB
314
+ [default0]:07/02/2024 15:26:51 [INFO|DP=0|PP=1|TP=0|ip-26-0-175-132]: iteration: 19 / 20 | consumed_tokens: 79.7M | elapsed_time_per_iteration_ms: 14.9K | tokens_per_sec: 281K | tokens_per_sec_per_gpu: 17.6K | global_batch_size: 1.02K | lm_loss: 7.09 | lr: 1.47e-05 | model_tflops_per_gpu: 160 | hardware_tflops_per_gpu: 160 | grad_norm: 2.47
315
+ [default0]:07/02/2024 15:27:05 [INFO|DP=0|PP=1|TP=0|ip-26-0-175-132]: iteration: 20 / 20 | consumed_tokens: 83.9M | elapsed_time_per_iteration_ms: 14.7K | tokens_per_sec: 285K | tokens_per_sec_per_gpu: 17.8K | global_batch_size: 1.02K | lm_loss: 7.03 | lr: 1e-05 | model_tflops_per_gpu: 162 | hardware_tflops_per_gpu: 162 | grad_norm: 2.02
316
+ Saved 1 csv files over 1 completed logs
317
+ Processing file: /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/16_GPUS/dp-8_tp-1_pp-2_mbz-4/profiler/ip-26-0-171-56_3064380.1719933802382173616.pt.trace.json
318
+ Results written to /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/16_GPUS/dp-8_tp-1_pp-2_mbz-4/profiler.csv
319
+ 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.
320
+
llama-1B/16_GPUS/dp-8_tp-1_pp-2_mbz-4/log_metrics.csv ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ iteration,consumed_tokens,elapsed_time_per_iteration_ms,tokens_per_sec,tokens_per_sec_per_gpu,global_batch_size,lm_loss,lr,model_tflops_per_gpu,hardware_tflops_per_gpu,grad_norm,memory_usage_MiB,peak_allocated_MiB,peak_reserved_MiB
2
+ 1,4190000.0000000005,22100.0,190000.0,11900.0,1020.0,11.1,0.0001,108.0,108.0,24.6,5024.53,39468.21,40396.0
3
+ 2,8390000.0,13900.0,302000.0,18900.0,1020.0,11.1,9.53e-05,172.0,172.0,24.8,5024.53,39468.21,40396.0
4
+ 3,12600000.0,13600.0,309000.0,19300.0,1020.0,10.5,9.05e-05,175.0,175.0,197.0,5024.53,39468.21,40396.0
5
+ 4,16800000.0,14100.0,297000.0,18500.0,1020.0,13.9,8.58e-05,168.0,168.0,17.9,5024.53,7821.94,40396.0
6
+ 5,21000000.0,14700.0,285000.0,17800.0,1020.0,9.71,8.11e-05,162.0,162.0,20.2,5024.53,39468.21,40396.0
7
+ 6,25200000.0,14300.0,293000.0,18300.0,1020.0,13.7,7.63e-05,166.0,166.0,98.3,5024.53,39468.21,40396.0
8
+ 7,29400000.0,110000.0,38300.0,2390.0,1020.0,9.73,7.16e-05,21.7,21.7,12.7,5024.53,39468.21,40396.0
9
+ 8,33600000.0,15000.0,279000.0,17400.0,1020.0,8.94,6.68e-05,158.0,158.0,11.9,5024.53,39468.21,40396.0
10
+ 9,37700000.0,14200.0,295000.0,18400.0,1020.0,8.51,6.21e-05,167.0,167.0,5.69,5024.53,39468.21,40396.0
11
+ 10,41900000.0,14000.0,299000.0,18700.0,1020.0,8.14,5.74e-05,170.0,170.0,4.25,5024.53,39468.21,40396.0
12
+ 11,46100000.0,14300.0,292000.0,18300.0,1020.0,8.01,5.26e-05,166.0,166.0,7.09,,,
13
+ 12,50300000.0,15400.0,272000.0,17000.0,1020.0,7.81,4.79e-05,155.0,155.0,6.25,5024.53,39468.21,40396.0
14
+ 13,54500000.0,14600.0,287000.0,18000.0,1020.0,7.6,4.32e-05,163.0,163.0,2.65,5024.53,39468.21,40396.0
15
+ 14,58700000.0,14800.0,283000.0,17700.0,1020.0,7.53,3.84e-05,161.0,161.0,3.82,,,
16
+ 15,62900000.0,13800.0,305000.0,19100.0,1020.0,7.42,3.37e-05,173.0,173.0,3.42,5024.53,39468.21,40396.0
17
+ 16,67099999.99999999,15500.0,271000.0,17000.0,1020.0,7.31,2.89e-05,154.0,154.0,2.89,,,
18
+ 17,71300000.0,14300.0,294000.0,18400.0,1020.0,7.21,2.42e-05,167.0,167.0,2.21,5024.53,39468.21,40396.0
19
+ 18,75500000.0,15700.0,267000.0,16700.0,1020.0,7.15,1.95e-05,152.0,152.0,2.57,5024.53,39468.21,40396.0
20
+ 19,79700000.0,14900.0,281000.0,17600.0,1020.0,7.09,1.47e-05,160.0,160.0,2.47,,,
21
+ 20,83900000.0,14700.0,285000.0,17800.0,1020.0,7.03,1e-05,162.0,162.0,2.02,,,
llama-1B/16_GPUS/dp-8_tp-1_pp-2_mbz-4/profiler.csv ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ forward,backward
2
+ 0ms 879μs,1ms 268μs
llama-1B/16_GPUS/dp-8_tp-1_pp-2_mbz-4/profiler/ip-26-0-171-56_3064380.1719933802382173616.pt.trace.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:752d1590b09e01706dd48e02060857925d63228d97b7bfe33f87b318216c311c
3
+ size 2540906029
llama-1B/16_GPUS/dp-8_tp-1_pp-2_mbz-4/status.txt ADDED
@@ -0,0 +1 @@
 
 
1
+ completed