Hugging Face's logo

nanotron
/
old_bench

Model card Files Community
3outeille HF staff commited on
Commit
8006f18
·
verified ·
1 Parent(s): 996fa8d

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

Browse files
Files changed (8) hide show
  1. .gitattributes +1 -0
  2. llama-1B/8_GPUS/dp-8_tp-1_pp-1_mbz-4/bench.slurm +111 -0
  3. llama-1B/8_GPUS/dp-8_tp-1_pp-1_mbz-4/config.yaml +90 -0
  4. llama-1B/8_GPUS/dp-8_tp-1_pp-1_mbz-4/log.out +240 -0
  5. llama-1B/8_GPUS/dp-8_tp-1_pp-1_mbz-4/log_metrics.csv +21 -0
  6. llama-1B/8_GPUS/dp-8_tp-1_pp-1_mbz-4/profiler.csv +2 -0
  7. llama-1B/8_GPUS/dp-8_tp-1_pp-1_mbz-4/profiler/ip-26-0-160-225_339725.1720044373966110627.pt.trace.json +3 -0
  8. llama-1B/8_GPUS/dp-8_tp-1_pp-1_mbz-4/status.txt +1 -0
.gitattributes CHANGED
@@ -107,3 +107,4 @@ llama-1B/8_GPUS/dp-4_tp-2_pp-1_mbz-8/profiler/ip-26-0-163-220_391612.17200425206
107
  llama-1B/8_GPUS/dp-1_tp-1_pp-8_mbz-4/profiler/ip-26-0-160-225_333932.1720042364135456939.pt.trace.json filter=lfs diff=lfs merge=lfs -text
108
  llama-1B/8_GPUS/dp-4_tp-2_pp-1_mbz-2/profiler/ip-26-0-174-36_223115.1720042846918455240.pt.trace.json.tmp filter=lfs diff=lfs merge=lfs -text
109
  llama-1B/8_GPUS/dp-2_tp-2_pp-2_mbz-4/profiler/ip-26-0-164-207_594258.1720042744576660800.pt.trace.json filter=lfs diff=lfs merge=lfs -text
 
 
107
  llama-1B/8_GPUS/dp-1_tp-1_pp-8_mbz-4/profiler/ip-26-0-160-225_333932.1720042364135456939.pt.trace.json filter=lfs diff=lfs merge=lfs -text
108
  llama-1B/8_GPUS/dp-4_tp-2_pp-1_mbz-2/profiler/ip-26-0-174-36_223115.1720042846918455240.pt.trace.json.tmp filter=lfs diff=lfs merge=lfs -text
109
  llama-1B/8_GPUS/dp-2_tp-2_pp-2_mbz-4/profiler/ip-26-0-164-207_594258.1720042744576660800.pt.trace.json filter=lfs diff=lfs merge=lfs -text
110
+ llama-1B/8_GPUS/dp-8_tp-1_pp-1_mbz-4/profiler/ip-26-0-160-225_339725.1720044373966110627.pt.trace.json filter=lfs diff=lfs merge=lfs -text
llama-1B/8_GPUS/dp-8_tp-1_pp-1_mbz-4/bench.slurm ADDED
@@ -0,0 +1,111 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/bin/bash
2
+
3
+ #SBATCH --job-name=bench_cluster
4
+ #SBATCH --time=02:00:00
5
+ #SBATCH --partition=hopper-prod
6
+ #SBATCH --nodes=1
7
+ #SBATCH --gres=gpu:8
8
+ #SBATCH --qos=normal
9
+ #SBATCH --ntasks-per-node=1
10
+ #SBATCH --cpus-per-task=96
11
+ #SBATCH --exclusive
12
+ #SBATCH --output=/fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-8_tp-1_pp-1_mbz-4/log.out
13
+ #SBATCH --error=/fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-8_tp-1_pp-1_mbz-4/log.out
14
+
15
+ # Function to update status based on squeue output
16
+ update_status() {
17
+ job_id=$1
18
+ status_file=$2
19
+ # For unknown reasons, it doenst update status for pending. It only works for running
20
+ while true; do
21
+ job_status=$(squeue --job $job_id --noheader --format=%T)
22
+ echo "Job status: $job_status"
23
+ if [ -z "$job_status" ]; then
24
+ # Job has finished or is not found
25
+ break
26
+ elif [ "$job_status" = "RUNNING" ]; then
27
+ printf "running" > $status_file
28
+ break
29
+ fi
30
+ sleep 10
31
+ done
32
+ }
33
+
34
+ # Misc initializations.
35
+ echo "========================"
36
+ echo "START TIME: $(date)"
37
+ source /fsx/ferdinandmom/miniforge3/etc/profile.d/conda.sh
38
+ conda activate /fsx/ferdinandmom/miniforge3/envs/env-bench-cluster
39
+ echo python3 version = $(python3 --version)
40
+ echo "========================"
41
+
42
+ # Slurm stuff
43
+ export HOSTNAMES=$(scontrol show hostnames "$SLURM_JOB_NODELIST")
44
+ export MASTER_ADDR=$(scontrol show hostnames "$SLURM_JOB_NODELIST" | head -n 1)
45
+ export MASTER_PORT=$((1024 + RANDOM % 64511))
46
+
47
+ export TMPDIR=/scratch
48
+ export HF_DATASETS_CACHE="/admin/home/ferdinand_mom/.cache"
49
+ export CUBLAS_WORKSPACE_CONFIG=":4096:8"
50
+ export CUDA_DEVICE_MAX_CONNECTIONS="1"
51
+
52
+ huggingface-cli login --token $HUGGINGFACE_TOKEN
53
+
54
+
55
+ NANOTRON_REPO="/fsx/ferdinandmom/ferdinand-hf/bench_cluster/nanotron"
56
+ CMD="$NANOTRON_REPO/run_train.py --config-file /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-8_tp-1_pp-1_mbz-4/config.yaml"
57
+
58
+ LAUNCHER="torchrun \
59
+ --nproc_per_node 8 \
60
+ --nnodes 1 \
61
+ --rdzv_endpoint ${MASTER_ADDR}:${MASTER_PORT} \
62
+ --rdzv_backend c10d \
63
+ --max_restarts 0 \
64
+ --tee 3 \
65
+ --node_rank ${SLURM_PROCID}"
66
+
67
+ # Checkout the bench_cluster branch
68
+ cd $NANOTRON_REPO
69
+ git checkout bench_cluster
70
+ cd ..
71
+ # Get the current job ID
72
+ job_id=${SLURM_JOB_ID}
73
+
74
+ # Update status to "pending" or "running" in the background
75
+ update_status $job_id /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-8_tp-1_pp-1_mbz-4/status.txt &
76
+
77
+ # Run the main command
78
+ srun -u $LAUNCHER $CMD
79
+ exit_status=$?
80
+
81
+ # Update status based on the exit status of `srun`
82
+ if [ $exit_status -eq 0 ]; then
83
+ printf "completed" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-8_tp-1_pp-1_mbz-4/status.txt
84
+ else
85
+ if grep -q "OutOfMemoryError" /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-8_tp-1_pp-1_mbz-4/log.out; then
86
+ printf "oom" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-8_tp-1_pp-1_mbz-4/status.txt
87
+ elif grep -q " CUDA error: an illegal memory access" /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-8_tp-1_pp-1_mbz-4/log.out; then
88
+ printf "oom" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-8_tp-1_pp-1_mbz-4/status.txt
89
+ elif grep -q "Timeout at NCCL" /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-8_tp-1_pp-1_mbz-4/log.out; then
90
+ printf "timeout" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-8_tp-1_pp-1_mbz-4/status.txt
91
+ else
92
+ printf "fail" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-8_tp-1_pp-1_mbz-4/status.txt
93
+ fi
94
+ fi
95
+
96
+ # Run the report script if the job completed successfully
97
+ if [ $exit_status -eq 0 ]; then
98
+ python /fsx/ferdinandmom/ferdinand-hf/bench_cluster/main.py report --inp_dir /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-8_tp-1_pp-1_mbz-4 --is_logs
99
+ python /fsx/ferdinandmom/ferdinand-hf/bench_cluster/main.py report --inp_dir /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-8_tp-1_pp-1_mbz-4 --is_profiler
100
+ fi
101
+
102
+
103
+ # Push to hub the folder using huggingface_cli
104
+ huggingface-cli upload nanotron/bench_cluster /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-8_tp-1_pp-1_mbz-4 llama-1B/8_GPUS/dp-8_tp-1_pp-1_mbz-4 --commit-message "Upload llama-1B/8_GPUS/dp-8_tp-1_pp-1_mbz-4"
105
+
106
+ # Verify the upload
107
+ if [ $? -eq 0 ]; then
108
+ echo "Uploading to Huggingface Hub successful"
109
+ else
110
+ echo "Failed to upload to Huggingface Hub"
111
+ fi
llama-1B/8_GPUS/dp-8_tp-1_pp-1_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: 1
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/8_GPUS/dp-8_tp-1_pp-1_mbz-4
57
+ tokenizer:
58
+ tokenizer_max_length: null
59
+ tokenizer_name_or_path: openai-community/gpt2
60
+ tokenizer_revision: null
61
+ data_stages:
62
+ - name: Training Stage
63
+ start_training_step: 1
64
+ data:
65
+ dataset:
66
+ dataset_overwrite_cache: false
67
+ dataset_processing_num_proc_per_process: 64
68
+ hf_dataset_config_name: null
69
+ hf_dataset_or_datasets: roneneldan/TinyStories
70
+ hf_dataset_splits: train
71
+ text_column_name: text
72
+ num_loading_workers: 0
73
+ seed: 42
74
+ lighteval: null
75
+ tokens:
76
+ train_steps: 20
77
+ val_check_interval: -1
78
+ batch_accumulation_per_replica: 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/8_GPUS/dp-8_tp-1_pp-1_mbz-4/log.out ADDED
@@ -0,0 +1,240 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ========================
2
+ START TIME: Wed Jul 3 22:01:56 UTC 2024
3
+ python3 version = Python 3.10.14
4
+ ========================
5
+ The token has not been saved to the git credentials helper. Pass `add_to_git_credential=True` in this function directly or `--add-to-git-credential` if using via `huggingface-cli` if you want to set the git credential as well.
6
+ Token is valid (permission: write).
7
+ Your token has been saved to /admin/home/ferdinand_mom/.cache/huggingface/token
8
+ Login successful
9
+ Already on 'bench_cluster'
10
+ M examples/config_tiny_llama.py
11
+ M examples/config_tiny_llama.yaml
12
+ M examples/train_tiny_llama.sh
13
+ M src/nanotron/models/llama.py
14
+ M src/nanotron/trainer.py
15
+ Your branch is up to date with 'origin/bench_cluster'.
16
+ Job status: RUNNING
17
+ W0703 22:02:02.875000 139751353034560 torch/distributed/run.py:757]
18
+ W0703 22:02:02.875000 139751353034560 torch/distributed/run.py:757] *****************************************
19
+ W0703 22:02:02.875000 139751353034560 torch/distributed/run.py:757] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed.
20
+ W0703 22:02:02.875000 139751353034560 torch/distributed/run.py:757] *****************************************
21
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Config:
22
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Config(general=GeneralArgs(project='bench_cluster',
23
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: run='%date_%jobid',
24
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: seed=42,
25
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: step=None,
26
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: consumed_train_samples=None,
27
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: benchmark_csv_path=None,
28
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: ignore_sanity_checks=True),
29
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: parallelism=ParallelismArgs(dp=8,
30
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: pp=1,
31
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: tp=1,
32
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: pp_engine=<nanotron.parallel.pipeline_parallel.engine.OneForwardOneBackwardPipelineEngine object at 0x7f1fbb0888e0>,
33
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: tp_mode=<TensorParallelLinearMode.REDUCE_SCATTER: 2>,
34
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: tp_linear_async_communication=False,
35
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: expert_parallel_size=1),
36
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: model=ModelArgs(model_config=LlamaConfig(bos_token_id=1,
37
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: eos_token_id=2,
38
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: hidden_act='silu',
39
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: hidden_size=2048,
40
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: initializer_range=0.02,
41
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: intermediate_size=4096,
42
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: is_llama_config=True,
43
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: max_position_embeddings=4096,
44
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: num_attention_heads=32,
45
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: num_hidden_layers=24,
46
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: num_key_value_heads=32,
47
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: pad_token_id=None,
48
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: pretraining_tp=1,
49
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: rms_norm_eps=1e-05,
50
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: rope_scaling=None,
51
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: rope_theta=10000.0,
52
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: tie_word_embeddings=True,
53
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: use_cache=True,
54
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: vocab_size=50257),
55
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: init_method=RandomInit(std=0.025),
56
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: dtype=torch.bfloat16,
57
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: make_vocab_size_divisible_by=1,
58
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: ddp_bucket_cap_mb=25),
59
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: tokenizer=TokenizerArgs(tokenizer_name_or_path='openai-community/gpt2',
60
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: tokenizer_revision=None,
61
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: tokenizer_max_length=None),
62
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: checkpoints=CheckpointsArgs(checkpoints_path=Path('/dev/null'),
63
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: checkpoint_interval=100000,
64
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: save_initial_state=False,
65
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: resume_checkpoint_path=None,
66
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: checkpoints_path_is_shared_file_system=False),
67
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: logging=LoggingArgs(log_level='info',
68
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: log_level_replica='info',
69
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: iteration_step_info_interval=1),
70
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: tokens=TokensArgs(sequence_length=4096,
71
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: train_steps=20,
72
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: micro_batch_size=4,
73
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: batch_accumulation_per_replica=32,
74
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: val_check_interval=-1,
75
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: limit_val_batches=0,
76
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: limit_test_batches=0),
77
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: optimizer=OptimizerArgs(optimizer_factory=AdamWOptimizerArgs(adam_eps=1e-08,
78
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: adam_beta1=0.9,
79
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: adam_beta2=0.95,
80
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: torch_adam_is_fused=True,
81
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: name='adamW'),
82
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: zero_stage=1,
83
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: weight_decay=0.01,
84
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: clip_grad=1.0,
85
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: accumulate_grad_in_fp32=True,
86
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: learning_rate_scheduler=LRSchedulerArgs(learning_rate=0.0001,
87
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: lr_warmup_steps=1,
88
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: lr_warmup_style='linear',
89
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: lr_decay_style='linear',
90
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: lr_decay_steps=19,
91
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: lr_decay_starting_step=None,
92
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: min_decay_lr=1e-05)),
93
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: data_stages=[DatasetStageArgs(name='Training Stage',
94
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: start_training_step=1,
95
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: data=DataArgs(dataset=PretrainDatasetsArgs(hf_dataset_or_datasets='roneneldan/TinyStories',
96
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: hf_dataset_splits='train',
97
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: hf_dataset_config_name=None,
98
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: dataset_processing_num_proc_per_process=64,
99
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: dataset_overwrite_cache=False,
100
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: text_column_name='text'),
101
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: seed=42,
102
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: num_loading_workers=0))],
103
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: profiler=ProfilerArgs(profiler_export_path=Path('/fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-8_tp-1_pp-1_mbz-4')),
104
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: lighteval=None)
105
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Model Config:
106
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: LlamaConfig(bos_token_id=1,
107
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: eos_token_id=2,
108
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: hidden_act='silu',
109
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: hidden_size=2048,
110
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: initializer_range=0.02,
111
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: intermediate_size=4096,
112
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: is_llama_config=True,
113
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: max_position_embeddings=4096,
114
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: num_attention_heads=32,
115
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: num_hidden_layers=24,
116
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: num_key_value_heads=32,
117
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: pad_token_id=None,
118
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: pretraining_tp=1,
119
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: rms_norm_eps=1e-05,
120
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: rope_scaling=None,
121
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: rope_theta=10000.0,
122
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: tie_word_embeddings=True,
123
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: use_cache=True,
124
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: vocab_size=50257)
125
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Building model..
126
+ [default0]:07/03/2024 22:02:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Setting PP block ranks...
127
+ [default0]:07/03/2024 22:02:33 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Total number of parameters: 1.11G (2116.51MiB)
128
+ [default0]:07/03/2024 22:02:33 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Local number of parameters: 1.11G (2116.51MiB)
129
+ [default0]:07/03/2024 22:02:33 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: [After model building] Memory usage: 2140.53MiB. Peak allocated: 2338.88MiB Peak reserved: 2392.00MiB
130
+ [default0]:07/03/2024 22:02:33 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: No checkpoint path provided.
131
+ [default0]:07/03/2024 22:02:33 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Parametrizing model parameters using StandardParametrizator
132
+ [default6]:07/03/2024 22:02:33 [INFO|DP=6|PP=0|TP=0|ip-26-0-160-225]: No checkpoint path provided.
133
+ [default7]:07/03/2024 22:02:33 [INFO|DP=7|PP=0|TP=0|ip-26-0-160-225]: No checkpoint path provided.
134
+ [default2]:07/03/2024 22:02:33 [INFO|DP=2|PP=0|TP=0|ip-26-0-160-225]: No checkpoint path provided.
135
+ [default5]:07/03/2024 22:02:33 [INFO|DP=5|PP=0|TP=0|ip-26-0-160-225]: No checkpoint path provided.
136
+ [default3]:07/03/2024 22:02:33 [INFO|DP=3|PP=0|TP=0|ip-26-0-160-225]: No checkpoint path provided.
137
+ [default4]:07/03/2024 22:02:33 [INFO|DP=4|PP=0|TP=0|ip-26-0-160-225]: No checkpoint path provided.
138
+ [default1]:07/03/2024 22:02:33 [INFO|DP=1|PP=0|TP=0|ip-26-0-160-225]: No checkpoint path provided.
139
+ [default0]:07/03/2024 22:02:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: [Optimizer Building] Using LearningRateForSP as learning rate
140
+ [default0]:07/03/2024 22:02:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: [ZeRO sharding] Size of optimizer params per rank:
141
+ [default0]:07/03/2024 22:02:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: [ZeRO sharding] DP Rank 0 has 139M out of 1.11G (12.50%) params' optimizer states
142
+ [default0]:07/03/2024 22:02:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: [ZeRO sharding] DP Rank 1 has 139M out of 1.11G (12.50%) params' optimizer states
143
+ [default0]:07/03/2024 22:02:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: [ZeRO sharding] DP Rank 2 has 139M out of 1.11G (12.50%) params' optimizer states
144
+ [default0]:07/03/2024 22:02:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: [ZeRO sharding] DP Rank 3 has 139M out of 1.11G (12.50%) params' optimizer states
145
+ [default0]:07/03/2024 22:02:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: [ZeRO sharding] DP Rank 4 has 139M out of 1.11G (12.50%) params' optimizer states
146
+ [default0]:07/03/2024 22:02:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: [ZeRO sharding] DP Rank 5 has 139M out of 1.11G (12.50%) params' optimizer states
147
+ [default0]:07/03/2024 22:02:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: [ZeRO sharding] DP Rank 6 has 139M out of 1.11G (12.50%) params' optimizer states
148
+ [default0]:07/03/2024 22:02:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: [ZeRO sharding] DP Rank 7 has 139M out of 1.11G (12.50%) params' optimizer states
149
+ [default0]:07/03/2024 22:02:42 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: [Training Plan] Stage Training Stage has 19 remaining training steps and has consumed 0 samples
150
+ [default0]:07/03/2024 22:02:42 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Using `datasets` library
151
+ [default0]:07/03/2024 22:02:42 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Loading tokenizer from openai-community/gpt2 and transformers/hf_hub versions ('4.41.2', '0.23.4')
152
+ [default0]:07/03/2024 22:02:42 [WARNING|DP=0|PP=0|TP=0|ip-26-0-160-225]: Repo card metadata block was not found. Setting CardData to empty.
153
+ [default0]:Repo card metadata block was not found. Setting CardData to empty.
154
+ [default0]:07/03/2024 22:02:44 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: [Training Plan] There are 1 training stages
155
+ [default0]:07/03/2024 22:02:44 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: [Stage Training Stage] start from step 1
156
+ [default0]:07/03/2024 22:02:44 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]:
157
+ [default0]:07/03/2024 22:02:44 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: [Start training] datetime: 2024-07-03 22:02:44.942292 | mbs: 4 | grad_accum: 32 | global_batch_size: 1024 | sequence_length: 4096 | train_steps: 20 | start_iteration_step: 0 | consumed_train_samples: 0
158
+ [default0]:07/03/2024 22:02:44 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Resuming training from stage Training Stage, it has trained for 0 samples and has 19 remaining train steps
159
+ [default0]:07/03/2024 22:02:44 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Memory usage: 6904.53MiB. Peak allocated 6904.53MiB. Peak reserved: 7156.00MiB
160
+ [default3]:Repo card metadata block was not found. Setting CardData to empty.
161
+ [default5]:07/03/2024 22:02:45 [WARNING|DP=5|PP=0|TP=0|ip-26-0-160-225]: Repo card metadata block was not found. Setting CardData to empty.
162
+ [default3]:07/03/2024 22:02:45 [WARNING|DP=3|PP=0|TP=0|ip-26-0-160-225]: Repo card metadata block was not found. Setting CardData to empty.
163
+ [default6]:07/03/2024 22:02:45 [WARNING|DP=6|PP=0|TP=0|ip-26-0-160-225]: Repo card metadata block was not found. Setting CardData to empty.
164
+ [default7]:07/03/2024 22:02:45 [WARNING|DP=7|PP=0|TP=0|ip-26-0-160-225]: 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
+ [default7]: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
+ [default4]:07/03/2024 22:02:45 [WARNING|DP=4|PP=0|TP=0|ip-26-0-160-225]: Repo card metadata block was not found. Setting CardData to empty.
169
+ [default4]:Repo card metadata block was not found. Setting CardData to empty.
170
+ [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.)
171
+ [default0]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
172
+ [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.)
173
+ [default6]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
174
+ [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.)
175
+ [default4]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
176
+ [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.)
177
+ [default5]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
178
+ [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.)
179
+ [default7]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
180
+ [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.)
181
+ [default3]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
182
+ [default2]:07/03/2024 22:02:50 [WARNING|DP=2|PP=0|TP=0|ip-26-0-160-225]: Repo card metadata block was not found. Setting CardData to empty.
183
+ [default2]:Repo card metadata block was not found. Setting CardData to empty.
184
+ [default1]:07/03/2024 22:02:50 [WARNING|DP=1|PP=0|TP=0|ip-26-0-160-225]: Repo card metadata block was not found. Setting CardData to empty.
185
+ [default1]:Repo card metadata block was not found. Setting CardData to empty.
186
+ [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.)
187
+ [default1]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
188
+ [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.)
189
+ [default2]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
190
+ [default0]:07/03/2024 22:03:02 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Memory usage: 6980.83MiB. Peak allocated 42628.63MiB. Peak reserved: 44258.00MiB
191
+ [default0]:07/03/2024 22:03:08 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: iteration: 1 / 20 | consumed_tokens: 4.19M | elapsed_time_per_iteration_ms: 23.7K | tokens_per_sec: 177K | tokens_per_sec_per_gpu: 22.1K | global_batch_size: 1.02K | lm_loss: 11.3 | lr: 0.0001 | model_tflops_per_gpu: 201 | hardware_tflops_per_gpu: 201 | grad_norm: 33.1 | cuda_memory_allocated: 8.43G | cuda_max_memory_reserved: 46.4G | hd_total_memory_tb: 312G | hd_used_memory_tb: 66.5G | hd_free_memory_tb: 246G
192
+ [default0]:07/03/2024 22:03:08 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Memory usage: 8040.99MiB. Peak allocated 12538.57MiB. Peak reserved: 44258.00MiB
193
+ [default0]:07/03/2024 22:03:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Memory usage: 8041.01MiB. Peak allocated 43688.81MiB. Peak reserved: 44258.00MiB
194
+ [default0]:07/03/2024 22:03:23 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: iteration: 2 / 20 | consumed_tokens: 8.39M | elapsed_time_per_iteration_ms: 14.5K | tokens_per_sec: 290K | tokens_per_sec_per_gpu: 36.3K | global_batch_size: 1.02K | lm_loss: 11.3 | lr: 9.53e-05 | model_tflops_per_gpu: 329 | hardware_tflops_per_gpu: 329 | grad_norm: 33.3 | cuda_memory_allocated: 8.43G | cuda_max_memory_reserved: 46.4G | hd_total_memory_tb: 312G | hd_used_memory_tb: 66.5G | hd_free_memory_tb: 246G
195
+ [default0]:07/03/2024 22:03:23 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Memory usage: 8040.99MiB. Peak allocated 12538.59MiB. Peak reserved: 44258.00MiB
196
+ [default0]:07/03/2024 22:03:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Memory usage: 8041.01MiB. Peak allocated 43688.81MiB. Peak reserved: 44258.00MiB
197
+ [default0]:07/03/2024 22:03:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: iteration: 3 / 20 | consumed_tokens: 12.6M | elapsed_time_per_iteration_ms: 14.4K | tokens_per_sec: 292K | tokens_per_sec_per_gpu: 36.5K | global_batch_size: 1.02K | lm_loss: 16 | lr: 9.05e-05 | model_tflops_per_gpu: 331 | hardware_tflops_per_gpu: 331 | grad_norm: 249 | cuda_memory_allocated: 8.43G | cuda_max_memory_reserved: 46.4G | hd_total_memory_tb: 312G | hd_used_memory_tb: 66.5G | hd_free_memory_tb: 246G
198
+ [default0]:STAGE:2024-07-03 22:03:37 339725:339725 ActivityProfilerController.cpp:314] Completed Stage: Warm Up
199
+ [default0]:07/03/2024 22:03:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Memory usage: 8040.99MiB. Peak allocated 12538.59MiB. Peak reserved: 44258.00MiB
200
+ [default0]:07/03/2024 22:03:51 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Memory usage: 8041.01MiB. Peak allocated 43688.81MiB. Peak reserved: 44258.00MiB
201
+ [default0]:07/03/2024 22:03:51 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: iteration: 4 / 20 | consumed_tokens: 16.8M | elapsed_time_per_iteration_ms: 14.4K | tokens_per_sec: 290K | tokens_per_sec_per_gpu: 36.3K | global_batch_size: 1.02K | lm_loss: 15.1 | lr: 8.58e-05 | model_tflops_per_gpu: 329 | hardware_tflops_per_gpu: 329 | grad_norm: 41.6 | cuda_memory_allocated: 8.43G | cuda_max_memory_reserved: 46.4G | hd_total_memory_tb: 312G | hd_used_memory_tb: 66.5G | hd_free_memory_tb: 246G
202
+ [default0]:07/03/2024 22:03:51 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Memory usage: 8040.99MiB. Peak allocated 12538.59MiB. Peak reserved: 44258.00MiB
203
+ [default0]:07/03/2024 22:04:06 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: iteration: 5 / 20 | consumed_tokens: 21M | elapsed_time_per_iteration_ms: 14.4K | tokens_per_sec: 291K | tokens_per_sec_per_gpu: 36.3K | global_batch_size: 1.02K | lm_loss: 10.8 | lr: 8.11e-05 | model_tflops_per_gpu: 330 | hardware_tflops_per_gpu: 330 | grad_norm: 26
204
+ [default0]:07/03/2024 22:04:06 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Memory usage: 8040.99MiB. Peak allocated 43688.81MiB. Peak reserved: 44258.00MiB
205
+ [default0]:07/03/2024 22:04:20 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: iteration: 6 / 20 | consumed_tokens: 25.2M | elapsed_time_per_iteration_ms: 14.4K | tokens_per_sec: 292K | tokens_per_sec_per_gpu: 36.5K | global_batch_size: 1.02K | lm_loss: 10.8 | lr: 7.63e-05 | model_tflops_per_gpu: 331 | hardware_tflops_per_gpu: 331 | grad_norm: 18.9
206
+ [default0]:STAGE:2024-07-03 22:04:38 339725:339725 ActivityProfilerController.cpp:320] Completed Stage: Collection
207
+ [default0]:STAGE:2024-07-03 22:04:39 339725:339725 ActivityProfilerController.cpp:324] Completed Stage: Post Processing
208
+ [default0]:07/03/2024 22:06:42 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Memory usage: 8040.99MiB. Peak allocated 43688.81MiB. Peak reserved: 44258.00MiB
209
+ [default0]:07/03/2024 22:06:56 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: iteration: 7 / 20 | consumed_tokens: 29.4M | elapsed_time_per_iteration_ms: 14.1K | tokens_per_sec: 297K | tokens_per_sec_per_gpu: 37.1K | global_batch_size: 1.02K | lm_loss: 10.2 | lr: 7.16e-05 | model_tflops_per_gpu: 336 | hardware_tflops_per_gpu: 336 | grad_norm: 7.97
210
+ [default0]:07/03/2024 22:06:56 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Memory usage: 8040.99MiB. Peak allocated 43688.81MiB. Peak reserved: 44258.00MiB
211
+ [default0]:07/03/2024 22:07:10 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: iteration: 8 / 20 | consumed_tokens: 33.6M | elapsed_time_per_iteration_ms: 14.3K | tokens_per_sec: 293K | tokens_per_sec_per_gpu: 36.6K | global_batch_size: 1.02K | lm_loss: 9.15 | lr: 6.68e-05 | model_tflops_per_gpu: 332 | hardware_tflops_per_gpu: 332 | grad_norm: 6.46
212
+ [default0]:07/03/2024 22:07:10 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Memory usage: 8040.99MiB. Peak allocated 43688.81MiB. Peak reserved: 44258.00MiB
213
+ [default0]:07/03/2024 22:07:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: iteration: 9 / 20 | consumed_tokens: 37.7M | elapsed_time_per_iteration_ms: 14.4K | tokens_per_sec: 292K | tokens_per_sec_per_gpu: 36.5K | global_batch_size: 1.02K | lm_loss: 11.2 | lr: 6.21e-05 | model_tflops_per_gpu: 331 | hardware_tflops_per_gpu: 331 | grad_norm: 59.7
214
+ [default0]:07/03/2024 22:07:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Memory usage: 8040.99MiB. Peak allocated 43688.81MiB. Peak reserved: 44258.00MiB
215
+ [default0]:07/03/2024 22:07:39 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: iteration: 10 / 20 | consumed_tokens: 41.9M | elapsed_time_per_iteration_ms: 14.4K | tokens_per_sec: 291K | tokens_per_sec_per_gpu: 36.3K | global_batch_size: 1.02K | lm_loss: 9.59 | lr: 5.74e-05 | model_tflops_per_gpu: 330 | hardware_tflops_per_gpu: 330 | grad_norm: 44.1
216
+ [default0]:07/03/2024 22:07:39 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Memory usage: 8040.99MiB. Peak allocated 43688.81MiB. Peak reserved: 44258.00MiB
217
+ [default0]:07/03/2024 22:07:53 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: iteration: 11 / 20 | consumed_tokens: 46.1M | elapsed_time_per_iteration_ms: 14.4K | tokens_per_sec: 290K | tokens_per_sec_per_gpu: 36.3K | global_batch_size: 1.02K | lm_loss: 8.08 | lr: 5.26e-05 | model_tflops_per_gpu: 329 | hardware_tflops_per_gpu: 329 | grad_norm: 8.4
218
+ [default0]:07/03/2024 22:07:53 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Memory usage: 8040.99MiB. Peak allocated 43688.81MiB. Peak reserved: 44258.00MiB
219
+ [default0]:07/03/2024 22:08:08 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: iteration: 12 / 20 | consumed_tokens: 50.3M | elapsed_time_per_iteration_ms: 14.4K | tokens_per_sec: 292K | tokens_per_sec_per_gpu: 36.4K | global_batch_size: 1.02K | lm_loss: 7.86 | lr: 4.79e-05 | model_tflops_per_gpu: 331 | hardware_tflops_per_gpu: 331 | grad_norm: 5.08
220
+ [default0]:07/03/2024 22:08:08 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Memory usage: 8040.99MiB. Peak allocated 43688.81MiB. Peak reserved: 44258.00MiB
221
+ [default0]:07/03/2024 22:08:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: iteration: 13 / 20 | consumed_tokens: 54.5M | elapsed_time_per_iteration_ms: 14.4K | tokens_per_sec: 292K | tokens_per_sec_per_gpu: 36.5K | global_batch_size: 1.02K | lm_loss: 7.7 | lr: 4.32e-05 | model_tflops_per_gpu: 331 | hardware_tflops_per_gpu: 331 | grad_norm: 4.72
222
+ [default0]:07/03/2024 22:08:22 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Memory usage: 8040.99MiB. Peak allocated 43688.81MiB. Peak reserved: 44258.00MiB
223
+ [default0]:07/03/2024 22:08:36 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: iteration: 14 / 20 | consumed_tokens: 58.7M | elapsed_time_per_iteration_ms: 14.4K | tokens_per_sec: 291K | tokens_per_sec_per_gpu: 36.3K | global_batch_size: 1.02K | lm_loss: 7.56 | lr: 3.84e-05 | model_tflops_per_gpu: 330 | hardware_tflops_per_gpu: 330 | grad_norm: 5.13
224
+ [default0]:07/03/2024 22:08:36 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Memory usage: 8040.99MiB. Peak allocated 43688.81MiB. Peak reserved: 44258.00MiB
225
+ [default0]:07/03/2024 22:08:51 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: iteration: 15 / 20 | consumed_tokens: 62.9M | elapsed_time_per_iteration_ms: 14.4K | tokens_per_sec: 291K | tokens_per_sec_per_gpu: 36.4K | global_batch_size: 1.02K | lm_loss: 7.4 | lr: 3.37e-05 | model_tflops_per_gpu: 330 | hardware_tflops_per_gpu: 330 | grad_norm: 5.17
226
+ [default0]:07/03/2024 22:08:51 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Memory usage: 8040.99MiB. Peak allocated 43688.81MiB. Peak reserved: 44258.00MiB
227
+ [default0]:07/03/2024 22:09:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: iteration: 16 / 20 | consumed_tokens: 67.1M | elapsed_time_per_iteration_ms: 14.4K | tokens_per_sec: 292K | tokens_per_sec_per_gpu: 36.5K | global_batch_size: 1.02K | lm_loss: 7.3 | lr: 2.89e-05 | model_tflops_per_gpu: 331 | hardware_tflops_per_gpu: 331 | grad_norm: 5.28
228
+ [default0]:07/03/2024 22:09:05 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Memory usage: 8040.99MiB. Peak allocated 43688.81MiB. Peak reserved: 44258.00MiB
229
+ [default0]:07/03/2024 22:09:20 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: iteration: 17 / 20 | consumed_tokens: 71.3M | elapsed_time_per_iteration_ms: 14.3K | tokens_per_sec: 293K | tokens_per_sec_per_gpu: 36.6K | global_batch_size: 1.02K | lm_loss: 7.22 | lr: 2.42e-05 | model_tflops_per_gpu: 332 | hardware_tflops_per_gpu: 332 | grad_norm: 5.23
230
+ [default0]:07/03/2024 22:09:20 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Memory usage: 8040.99MiB. Peak allocated 43688.81MiB. Peak reserved: 44258.00MiB
231
+ [default0]:07/03/2024 22:09:34 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: iteration: 18 / 20 | consumed_tokens: 75.5M | elapsed_time_per_iteration_ms: 14.4K | tokens_per_sec: 292K | tokens_per_sec_per_gpu: 36.5K | global_batch_size: 1.02K | lm_loss: 7.15 | lr: 1.95e-05 | model_tflops_per_gpu: 331 | hardware_tflops_per_gpu: 331 | grad_norm: 5.1
232
+ [default0]:07/03/2024 22:09:34 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Memory usage: 8040.99MiB. Peak allocated 43688.81MiB. Peak reserved: 44258.00MiB
233
+ [default0]:07/03/2024 22:09:48 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: iteration: 19 / 20 | consumed_tokens: 79.7M | elapsed_time_per_iteration_ms: 14.4K | tokens_per_sec: 292K | tokens_per_sec_per_gpu: 36.4K | global_batch_size: 1.02K | lm_loss: 7.08 | lr: 1.47e-05 | model_tflops_per_gpu: 331 | hardware_tflops_per_gpu: 331 | grad_norm: 3.96
234
+ [default0]:07/03/2024 22:09:48 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: Memory usage: 8040.99MiB. Peak allocated 43688.81MiB. Peak reserved: 44258.00MiB
235
+ [default0]:07/03/2024 22:10:03 [INFO|DP=0|PP=0|TP=0|ip-26-0-160-225]: iteration: 20 / 20 | consumed_tokens: 83.9M | elapsed_time_per_iteration_ms: 14.4K | tokens_per_sec: 291K | tokens_per_sec_per_gpu: 36.4K | global_batch_size: 1.02K | lm_loss: 7.03 | lr: 1e-05 | model_tflops_per_gpu: 330 | hardware_tflops_per_gpu: 330 | grad_norm: 2.93
236
+ Saved 1 csv files over 1 completed logs
237
+ Processing file: /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-8_tp-1_pp-1_mbz-4/profiler/ip-26-0-160-225_339725.1720044373966110627.pt.trace.json
238
+ Results written to /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-8_tp-1_pp-1_mbz-4/profiler.csv
239
+ 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.
240
+
llama-1B/8_GPUS/dp-8_tp-1_pp-1_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
+ 1,4190000.0000000005,23700.0,177000.0,22100.0,1020.0,11.3,0.0001,201.0,201.0,33.1,8041.01,43688.81,44258.0
3
+ 2,8390000.0,14500.0,290000.0,36300.0,1020.0,11.3,9.53e-05,329.0,329.0,33.3,8041.01,43688.81,44258.0
4
+ 3,12600000.0,14400.0,292000.0,36500.0,1020.0,16.0,9.05e-05,331.0,331.0,249.0,8041.01,43688.81,44258.0
5
+ 4,16800000.0,14400.0,290000.0,36300.0,1020.0,15.1,8.58e-05,329.0,329.0,41.6,8040.99,12538.59,44258.0
6
+ 5,21000000.0,14400.0,291000.0,36300.0,1020.0,10.8,8.11e-05,330.0,330.0,26.0,8040.99,43688.81,44258.0
7
+ 6,25200000.0,14400.0,292000.0,36500.0,1020.0,10.8,7.63e-05,331.0,331.0,18.9,8040.99,43688.81,44258.0
8
+ 7,29400000.0,14100.0,297000.0,37100.0,1020.0,10.2,7.16e-05,336.0,336.0,7.97,8040.99,43688.81,44258.0
9
+ 8,33600000.0,14300.0,293000.0,36600.0,1020.0,9.15,6.68e-05,332.0,332.0,6.46,8040.99,43688.81,44258.0
10
+ 9,37700000.0,14400.0,292000.0,36500.0,1020.0,11.2,6.21e-05,331.0,331.0,59.7,8040.99,43688.81,44258.0
11
+ 10,41900000.0,14400.0,291000.0,36300.0,1020.0,9.59,5.74e-05,330.0,330.0,44.1,8040.99,43688.81,44258.0
12
+ 11,46100000.0,14400.0,290000.0,36300.0,1020.0,8.08,5.26e-05,329.0,329.0,8.4,8040.99,43688.81,44258.0
13
+ 12,50300000.0,14400.0,292000.0,36400.0,1020.0,7.86,4.79e-05,331.0,331.0,5.08,8040.99,43688.81,44258.0
14
+ 13,54500000.0,14400.0,292000.0,36500.0,1020.0,7.7,4.32e-05,331.0,331.0,4.72,8040.99,43688.81,44258.0
15
+ 14,58700000.0,14400.0,291000.0,36300.0,1020.0,7.56,3.84e-05,330.0,330.0,5.13,8040.99,43688.81,44258.0
16
+ 15,62900000.0,14400.0,291000.0,36400.0,1020.0,7.4,3.37e-05,330.0,330.0,5.17,8040.99,43688.81,44258.0
17
+ 16,67099999.99999999,14400.0,292000.0,36500.0,1020.0,7.3,2.89e-05,331.0,331.0,5.28,8040.99,43688.81,44258.0
18
+ 17,71300000.0,14300.0,293000.0,36600.0,1020.0,7.22,2.42e-05,332.0,332.0,5.23,8040.99,43688.81,44258.0
19
+ 18,75500000.0,14400.0,292000.0,36500.0,1020.0,7.15,1.95e-05,331.0,331.0,5.1,8040.99,43688.81,44258.0
20
+ 19,79700000.0,14400.0,292000.0,36400.0,1020.0,7.08,1.47e-05,331.0,331.0,3.96,8040.99,43688.81,44258.0
21
+ 20,83900000.0,14400.0,291000.0,36400.0,1020.0,7.03,1e-05,330.0,330.0,2.93,,,
llama-1B/8_GPUS/dp-8_tp-1_pp-1_mbz-4/profiler.csv ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ forward,backward
2
+ 0ms 894μs,0ms 346μs
llama-1B/8_GPUS/dp-8_tp-1_pp-1_mbz-4/profiler/ip-26-0-160-225_339725.1720044373966110627.pt.trace.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1e5a7a3fbd59ca563f0f098cc071129a1883009560627e769b1211e3a76e46fd
3
+ size 4320505771
llama-1B/8_GPUS/dp-8_tp-1_pp-1_mbz-4/status.txt ADDED
@@ -0,0 +1 @@
 
 
1
+ completed