Hugging Face's logo

nanotron
/
old_bench

Model card Files Community
3outeille HF staff commited on
Commit
8566631
•
1 Parent(s): 5e2919f

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

Browse files
Files changed (8) hide show
  1. .gitattributes +1 -0
  2. llama-1B/8_GPUS/dp-2_tp-1_pp-4_mbz-4/bench.slurm +111 -0
  3. llama-1B/8_GPUS/dp-2_tp-1_pp-4_mbz-4/config.yaml +90 -0
  4. llama-1B/8_GPUS/dp-2_tp-1_pp-4_mbz-4/log.out +250 -0
  5. llama-1B/8_GPUS/dp-2_tp-1_pp-4_mbz-4/log_metrics.csv +21 -0
  6. llama-1B/8_GPUS/dp-2_tp-1_pp-4_mbz-4/profiler.csv +2 -0
  7. llama-1B/8_GPUS/dp-2_tp-1_pp-4_mbz-4/profiler/ip-26-0-164-207_605035.1720046681997701534.pt.trace.json +3 -0
  8. llama-1B/8_GPUS/dp-2_tp-1_pp-4_mbz-4/status.txt +1 -0
.gitattributes CHANGED
@@ -108,3 +108,4 @@ llama-1B/8_GPUS/dp-1_tp-1_pp-8_mbz-4/profiler/ip-26-0-160-225_333932.17200423641
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
 
 
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
111
+ llama-1B/8_GPUS/dp-2_tp-1_pp-4_mbz-4/profiler/ip-26-0-164-207_605035.1720046681997701534.pt.trace.json filter=lfs diff=lfs merge=lfs -text
llama-1B/8_GPUS/dp-2_tp-1_pp-4_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-2_tp-1_pp-4_mbz-4/log.out
13
+ #SBATCH --error=/fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-2_tp-1_pp-4_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-2_tp-1_pp-4_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-2_tp-1_pp-4_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-2_tp-1_pp-4_mbz-4/status.txt
84
+ else
85
+ if grep -q "OutOfMemoryError" /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-2_tp-1_pp-4_mbz-4/log.out; then
86
+ printf "oom" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-2_tp-1_pp-4_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-2_tp-1_pp-4_mbz-4/log.out; then
88
+ printf "oom" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-2_tp-1_pp-4_mbz-4/status.txt
89
+ elif grep -q "Timeout at NCCL" /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-2_tp-1_pp-4_mbz-4/log.out; then
90
+ printf "timeout" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-2_tp-1_pp-4_mbz-4/status.txt
91
+ else
92
+ printf "fail" > /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-2_tp-1_pp-4_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-2_tp-1_pp-4_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-2_tp-1_pp-4_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-2_tp-1_pp-4_mbz-4 llama-1B/8_GPUS/dp-2_tp-1_pp-4_mbz-4 --commit-message "Upload llama-1B/8_GPUS/dp-2_tp-1_pp-4_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-2_tp-1_pp-4_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: 2
49
+ expert_parallel_size: 1
50
+ pp: 4
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-2_tp-1_pp-4_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: 128
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-2_tp-1_pp-4_mbz-4/log.out ADDED
@@ -0,0 +1,250 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ========================
2
+ START TIME: Wed Jul 3 22:39:17 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 22:39:20.273000 139671673374528 torch/distributed/run.py:757]
18
+ W0703 22:39:20.273000 139671673374528 torch/distributed/run.py:757] *****************************************
19
+ W0703 22:39:20.273000 139671673374528 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 22:39:20.273000 139671673374528 torch/distributed/run.py:757] *****************************************
21
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: Config:
22
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: Config(general=GeneralArgs(project='bench_cluster',
23
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: run='%date_%jobid',
24
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: seed=42,
25
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: step=None,
26
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: consumed_train_samples=None,
27
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: benchmark_csv_path=None,
28
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: ignore_sanity_checks=True),
29
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: parallelism=ParallelismArgs(dp=2,
30
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: pp=4,
31
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: tp=1,
32
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: pp_engine=<nanotron.parallel.pipeline_parallel.engine.OneForwardOneBackwardPipelineEngine object at 0x7f4edc87c5e0>,
33
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: tp_mode=<TensorParallelLinearMode.REDUCE_SCATTER: 2>,
34
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: tp_linear_async_communication=False,
35
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: expert_parallel_size=1),
36
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: model=ModelArgs(model_config=LlamaConfig(bos_token_id=1,
37
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: eos_token_id=2,
38
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: hidden_act='silu',
39
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: hidden_size=2048,
40
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: initializer_range=0.02,
41
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: intermediate_size=4096,
42
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: is_llama_config=True,
43
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: max_position_embeddings=4096,
44
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: num_attention_heads=32,
45
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: num_hidden_layers=24,
46
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: num_key_value_heads=32,
47
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: pad_token_id=None,
48
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: pretraining_tp=1,
49
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: rms_norm_eps=1e-05,
50
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: rope_scaling=None,
51
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: rope_theta=10000.0,
52
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: tie_word_embeddings=True,
53
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: use_cache=True,
54
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: vocab_size=50257),
55
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: init_method=RandomInit(std=0.025),
56
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: dtype=torch.bfloat16,
57
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: make_vocab_size_divisible_by=1,
58
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: ddp_bucket_cap_mb=25),
59
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: tokenizer=TokenizerArgs(tokenizer_name_or_path='openai-community/gpt2',
60
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: tokenizer_revision=None,
61
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: tokenizer_max_length=None),
62
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: checkpoints=CheckpointsArgs(checkpoints_path=Path('/dev/null'),
63
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: checkpoint_interval=100000,
64
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: save_initial_state=False,
65
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: resume_checkpoint_path=None,
66
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: checkpoints_path_is_shared_file_system=False),
67
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: logging=LoggingArgs(log_level='info',
68
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: log_level_replica='info',
69
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: iteration_step_info_interval=1),
70
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: tokens=TokensArgs(sequence_length=4096,
71
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: train_steps=20,
72
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: micro_batch_size=4,
73
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: batch_accumulation_per_replica=128,
74
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: val_check_interval=-1,
75
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: limit_val_batches=0,
76
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: limit_test_batches=0),
77
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: optimizer=OptimizerArgs(optimizer_factory=AdamWOptimizerArgs(adam_eps=1e-08,
78
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: adam_beta1=0.9,
79
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: adam_beta2=0.95,
80
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: torch_adam_is_fused=True,
81
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: name='adamW'),
82
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: zero_stage=1,
83
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: weight_decay=0.01,
84
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: clip_grad=1.0,
85
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: accumulate_grad_in_fp32=True,
86
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: learning_rate_scheduler=LRSchedulerArgs(learning_rate=0.0001,
87
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: lr_warmup_steps=1,
88
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: lr_warmup_style='linear',
89
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: lr_decay_style='linear',
90
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: lr_decay_steps=19,
91
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: lr_decay_starting_step=None,
92
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: min_decay_lr=1e-05)),
93
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: data_stages=[DatasetStageArgs(name='Training Stage',
94
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: start_training_step=1,
95
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: data=DataArgs(dataset=PretrainDatasetsArgs(hf_dataset_or_datasets='roneneldan/TinyStories',
96
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: hf_dataset_splits='train',
97
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: hf_dataset_config_name=None,
98
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: dataset_processing_num_proc_per_process=64,
99
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: dataset_overwrite_cache=False,
100
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: text_column_name='text'),
101
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: seed=42,
102
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: num_loading_workers=0))],
103
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: profiler=ProfilerArgs(profiler_export_path=Path('/fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-2_tp-1_pp-4_mbz-4')),
104
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: lighteval=None)
105
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: Model Config:
106
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: LlamaConfig(bos_token_id=1,
107
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: eos_token_id=2,
108
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: hidden_act='silu',
109
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: hidden_size=2048,
110
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: initializer_range=0.02,
111
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: intermediate_size=4096,
112
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: is_llama_config=True,
113
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: max_position_embeddings=4096,
114
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: num_attention_heads=32,
115
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: num_hidden_layers=24,
116
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: num_key_value_heads=32,
117
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: pad_token_id=None,
118
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: pretraining_tp=1,
119
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: rms_norm_eps=1e-05,
120
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: rope_scaling=None,
121
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: rope_theta=10000.0,
122
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: tie_word_embeddings=True,
123
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: use_cache=True,
124
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: vocab_size=50257)
125
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: Building model..
126
+ [default0]:07/03/2024 22:39:37 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: Setting PP block ranks...
127
+ [default2]:07/03/2024 22:39:50 [INFO|DP=0|PP=1|TP=0|ip-26-0-164-207]: Local number of parameters: 294M (560.05MiB)
128
+ [default2]:07/03/2024 22:39:50 [INFO|DP=0|PP=1|TP=0|ip-26-0-164-207]: [After model building] Memory usage: 567.07MiB. Peak allocated: 569.10MiB Peak reserved: 594.00MiB
129
+ [default2]:07/03/2024 22:39:50 [INFO|DP=0|PP=1|TP=0|ip-26-0-164-207]: No checkpoint path provided.
130
+ [default3]:07/03/2024 22:39:50 [INFO|DP=1|PP=1|TP=0|ip-26-0-164-207]: No checkpoint path provided.
131
+ [default7]:07/03/2024 22:39:50 [INFO|DP=1|PP=3|TP=0|ip-26-0-164-207]: No checkpoint path provided.
132
+ [default0]:07/03/2024 22:39:50 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: Total number of parameters: 1.21G (2312.82MiB)
133
+ [default0]:07/03/2024 22:39:50 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: Local number of parameters: 397M (756.37MiB)
134
+ [default0]:07/03/2024 22:39:50 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: [After model building] Memory usage: 763.38MiB. Peak allocated: 765.41MiB Peak reserved: 792.00MiB
135
+ [default0]:07/03/2024 22:39:50 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: No checkpoint path provided.
136
+ [default0]:07/03/2024 22:39:50 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: Parametrizing model parameters using StandardParametrizator
137
+ [default1]:07/03/2024 22:39:50 [INFO|DP=1|PP=0|TP=0|ip-26-0-164-207]: No checkpoint path provided.
138
+ [default4]:07/03/2024 22:39:50 [INFO|DP=0|PP=2|TP=0|ip-26-0-164-207]: Local number of parameters: 252M (480.05MiB)
139
+ [default6]:07/03/2024 22:39:50 [INFO|DP=0|PP=3|TP=0|ip-26-0-164-207]: Local number of parameters: 271M (516.35MiB)
140
+ [default6]:07/03/2024 22:39:50 [INFO|DP=0|PP=3|TP=0|ip-26-0-164-207]: [After model building] Memory usage: 520.36MiB. Peak allocated: 522.39MiB Peak reserved: 534.00MiB
141
+ [default6]:07/03/2024 22:39:50 [INFO|DP=0|PP=3|TP=0|ip-26-0-164-207]: No checkpoint path provided.
142
+ [default5]:07/03/2024 22:39:50 [INFO|DP=1|PP=2|TP=0|ip-26-0-164-207]: No checkpoint path provided.
143
+ [default4]:07/03/2024 22:39:50 [INFO|DP=0|PP=2|TP=0|ip-26-0-164-207]: [After model building] Memory usage: 486.06MiB. Peak allocated: 488.09MiB Peak reserved: 502.00MiB
144
+ [default4]:07/03/2024 22:39:50 [INFO|DP=0|PP=2|TP=0|ip-26-0-164-207]: No checkpoint path provided.
145
+ [default0]:07/03/2024 22:39:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: [Optimizer Building] Using LearningRateForSP as learning rate
146
+ [default0]:07/03/2024 22:39:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: [ZeRO sharding] Size of optimizer params per rank:
147
+ [default0]:07/03/2024 22:39:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: [ZeRO sharding] DP Rank 0 has 198M out of 397M (50.00%) params' optimizer states
148
+ [default0]:07/03/2024 22:39:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: [ZeRO sharding] DP Rank 1 has 198M out of 397M (50.00%) params' optimizer states
149
+ [default0]:07/03/2024 22:39:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: [Training Plan] Stage Training Stage has 19 remaining training steps and has consumed 0 samples
150
+ [default0]:07/03/2024 22:39:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: Using `datasets` library
151
+ [default0]:07/03/2024 22:39:54 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: Loading tokenizer from openai-community/gpt2 and transformers/hf_hub versions ('4.41.2', '0.23.4')
152
+ [default0]:07/03/2024 22:39:55 [WARNING|DP=0|PP=0|TP=0|ip-26-0-164-207]: 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 22:39:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: [Training Plan] There are 1 training stages
155
+ [default0]:07/03/2024 22:39:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: [Stage Training Stage] start from step 1
156
+ [default0]:07/03/2024 22:39:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]:
157
+ [default0]:07/03/2024 22:39:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: [Start training] datetime: 2024-07-03 22:39:55.752615 | mbs: 4 | grad_accum: 128 | global_batch_size: 1024 | sequence_length: 4096 | train_steps: 20 | start_iteration_step: 0 | consumed_train_samples: 0
158
+ [default0]:07/03/2024 22:39:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: Resuming training from stage Training Stage, it has trained for 0 samples and has 19 remaining train steps
159
+ [default0]:07/03/2024 22:39:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: Memory usage: 3032.50MiB. Peak allocated 3032.50MiB. Peak reserved: 3064.00MiB
160
+ [default3]:07/03/2024 22:39:55 [WARNING|DP=1|PP=1|TP=0|ip-26-0-164-207]: Repo card metadata block was not found. Setting CardData to empty.
161
+ [default7]:07/03/2024 22:39:55 [WARNING|DP=1|PP=3|TP=0|ip-26-0-164-207]: Repo card metadata block was not found. Setting CardData to empty.
162
+ [default1]:07/03/2024 22:39:55 [WARNING|DP=1|PP=0|TP=0|ip-26-0-164-207]: Repo card metadata block was not found. Setting CardData to empty.
163
+ [default5]:07/03/2024 22:39:55 [WARNING|DP=1|PP=2|TP=0|ip-26-0-164-207]: Repo card metadata block was not found. Setting CardData to empty.
164
+ [default6]:07/03/2024 22:39:55 [WARNING|DP=0|PP=3|TP=0|ip-26-0-164-207]: Repo card metadata block was not found. Setting CardData to empty.
165
+ [default4]:07/03/2024 22:39:55 [WARNING|DP=0|PP=2|TP=0|ip-26-0-164-207]: Repo card metadata block was not found. Setting CardData to empty.
166
+ [default6]:Repo card metadata block was not found. Setting CardData to empty.
167
+ [default7]:Repo card metadata block was not found. Setting CardData to empty.
168
+ [default3]: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
+ [default1]:Repo card metadata block was not found. Setting CardData to empty.
171
+ [default5]:Repo card metadata block was not found. Setting CardData to empty.
172
+ [default2]:07/03/2024 22:39:55 [WARNING|DP=0|PP=1|TP=0|ip-26-0-164-207]: Repo card metadata block was not found. Setting CardData to empty.
173
+ [default2]:Repo card metadata block was not found. Setting CardData to empty.
174
+ [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.)
175
+ [default6]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
176
+ [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.)
177
+ [default4]: 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
+ [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.)
181
+ [default2]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
182
+ [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.)
183
+ [default5]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
184
+ [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.)
185
+ [default0]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
186
+ [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.)
187
+ [default0]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
188
+ [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.)
189
+ [default3]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
190
+ [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.)
191
+ [default1]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
192
+ [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
193
+ [default6]: warnings.warn(
194
+ [default0]:07/03/2024 22:40:29 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: Memory usage: 3100.14MiB. Peak allocated 37544.07MiB. Peak reserved: 37812.00MiB
195
+ [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
196
+ [default0]: warnings.warn(
197
+ [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
198
+ [default7]: warnings.warn(
199
+ [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
200
+ [default1]: warnings.warn(
201
+ [default0]:07/03/2024 22:40:33 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: Memory usage: 4612.90MiB. Peak allocated 6503.83MiB. Peak reserved: 39710.00MiB
202
+ [default6]:07/03/2024 22:40:33 [INFO|DP=0|PP=3|TP=0|ip-26-0-164-207]: iteration: 1 / 20 | consumed_tokens: 4.19M | elapsed_time_per_iteration_ms: 36.3K | tokens_per_sec: 115K | tokens_per_sec_per_gpu: 14.4K | global_batch_size: 1.02K | lm_loss: 11.1 | lr: 0.0001 | model_tflops_per_gpu: 131 | hardware_tflops_per_gpu: 131 | grad_norm: 25.1 | cuda_memory_allocated: 3.32G | cuda_max_memory_reserved: 15.8G | hd_total_memory_tb: 312G | hd_used_memory_tb: 65.7G | hd_free_memory_tb: 246G
203
+ [default0]:07/03/2024 22:40:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: Memory usage: 4612.90MiB. Peak allocated 39056.83MiB. Peak reserved: 39710.00MiB
204
+ [default6]:07/03/2024 22:40:55 [INFO|DP=0|PP=3|TP=0|ip-26-0-164-207]: iteration: 2 / 20 | consumed_tokens: 8.39M | elapsed_time_per_iteration_ms: 22K | tokens_per_sec: 190K | tokens_per_sec_per_gpu: 23.8K | global_batch_size: 1.02K | lm_loss: 11.1 | lr: 9.53e-05 | model_tflops_per_gpu: 216 | hardware_tflops_per_gpu: 216 | grad_norm: 25.2 | cuda_memory_allocated: 3.32G | cuda_max_memory_reserved: 15.8G | hd_total_memory_tb: 312G | hd_used_memory_tb: 65.7G | hd_free_memory_tb: 246G
205
+ [default0]:07/03/2024 22:40:55 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: Memory usage: 4612.90MiB. Peak allocated 6503.83MiB. Peak reserved: 39710.00MiB
206
+ [default0]:07/03/2024 22:41:17 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: Memory usage: 4612.90MiB. Peak allocated 39056.83MiB. Peak reserved: 39710.00MiB
207
+ [default0]:07/03/2024 22:41:17 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: Memory usage: 4612.90MiB. Peak allocated 6503.83MiB. Peak reserved: 39710.00MiB
208
+ [default0]:STAGE:2024-07-03 22:41:17 605035:605035 ActivityProfilerController.cpp:314] Completed Stage: Warm Up
209
+ [default6]:07/03/2024 22:41:17 [INFO|DP=0|PP=3|TP=0|ip-26-0-164-207]: iteration: 3 / 20 | consumed_tokens: 12.6M | elapsed_time_per_iteration_ms: 22K | tokens_per_sec: 191K | tokens_per_sec_per_gpu: 23.8K | global_batch_size: 1.02K | lm_loss: 11.4 | lr: 9.05e-05 | model_tflops_per_gpu: 216 | hardware_tflops_per_gpu: 216 | grad_norm: 217 | cuda_memory_allocated: 3.32G | cuda_max_memory_reserved: 15.8G | hd_total_memory_tb: 312G | hd_used_memory_tb: 65.7G | hd_free_memory_tb: 246G
210
+ [default0]:07/03/2024 22:41:39 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: Memory usage: 4612.90MiB. Peak allocated 39056.83MiB. Peak reserved: 39710.00MiB
211
+ [default0]:07/03/2024 22:41:39 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: Memory usage: 4612.90MiB. Peak allocated 6503.83MiB. Peak reserved: 39710.00MiB
212
+ [default6]:07/03/2024 22:41:39 [INFO|DP=0|PP=3|TP=0|ip-26-0-164-207]: iteration: 4 / 20 | consumed_tokens: 16.8M | elapsed_time_per_iteration_ms: 22K | tokens_per_sec: 191K | tokens_per_sec_per_gpu: 23.8K | global_batch_size: 1.02K | lm_loss: 13.8 | lr: 8.58e-05 | model_tflops_per_gpu: 216 | hardware_tflops_per_gpu: 216 | grad_norm: 22.5 | cuda_memory_allocated: 3.32G | cuda_max_memory_reserved: 15.8G | hd_total_memory_tb: 312G | hd_used_memory_tb: 65.7G | hd_free_memory_tb: 246G
213
+ [default6]:07/03/2024 22:42:01 [INFO|DP=0|PP=3|TP=0|ip-26-0-164-207]: iteration: 5 / 20 | consumed_tokens: 21M | elapsed_time_per_iteration_ms: 22K | tokens_per_sec: 191K | tokens_per_sec_per_gpu: 23.8K | global_batch_size: 1.02K | lm_loss: 9.98 | lr: 8.11e-05 | model_tflops_per_gpu: 216 | hardware_tflops_per_gpu: 216 | grad_norm: 16.4
214
+ [default0]:07/03/2024 22:42:01 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: Memory usage: 4612.90MiB. Peak allocated 39056.83MiB. Peak reserved: 39710.00MiB
215
+ [default6]:07/03/2024 22:42:23 [INFO|DP=0|PP=3|TP=0|ip-26-0-164-207]: iteration: 6 / 20 | consumed_tokens: 25.2M | elapsed_time_per_iteration_ms: 22K | tokens_per_sec: 191K | tokens_per_sec_per_gpu: 23.9K | global_batch_size: 1.02K | lm_loss: 10.9 | lr: 7.63e-05 | model_tflops_per_gpu: 217 | hardware_tflops_per_gpu: 217 | grad_norm: 93.8
216
+ [default0]:STAGE:2024-07-03 22:42:43 605035:605035 ActivityProfilerController.cpp:320] Completed Stage: Collection
217
+ [default0]:STAGE:2024-07-03 22:42:45 605035:605035 ActivityProfilerController.cpp:324] Completed Stage: Post Processing
218
+ [default0]:07/03/2024 22:45:18 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: Memory usage: 4612.90MiB. Peak allocated 39056.83MiB. Peak reserved: 39710.00MiB
219
+ [default0]:07/03/2024 22:45:40 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: Memory usage: 4612.90MiB. Peak allocated 39056.83MiB. Peak reserved: 39710.00MiB
220
+ [default6]:07/03/2024 22:45:40 [INFO|DP=0|PP=3|TP=0|ip-26-0-164-207]: iteration: 7 / 20 | consumed_tokens: 29.4M | elapsed_time_per_iteration_ms: 198K | tokens_per_sec: 21.2K | tokens_per_sec_per_gpu: 2.65K | global_batch_size: 1.02K | lm_loss: 9.16 | lr: 7.16e-05 | model_tflops_per_gpu: 24 | hardware_tflops_per_gpu: 24 | grad_norm: 19.8
221
+ [default6]:07/03/2024 22:46:02 [INFO|DP=0|PP=3|TP=0|ip-26-0-164-207]: iteration: 8 / 20 | consumed_tokens: 33.6M | elapsed_time_per_iteration_ms: 22K | tokens_per_sec: 191K | tokens_per_sec_per_gpu: 23.9K | global_batch_size: 1.02K | lm_loss: 8.83 | lr: 6.68e-05 | model_tflops_per_gpu: 216 | hardware_tflops_per_gpu: 216 | grad_norm: 6.08
222
+ [default0]:07/03/2024 22:46:02 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: Memory usage: 4612.90MiB. Peak allocated 39056.83MiB. Peak reserved: 39710.00MiB
223
+ [default6]:07/03/2024 22:46:24 [INFO|DP=0|PP=3|TP=0|ip-26-0-164-207]: iteration: 9 / 20 | consumed_tokens: 37.7M | elapsed_time_per_iteration_ms: 22K | tokens_per_sec: 190K | tokens_per_sec_per_gpu: 23.8K | global_batch_size: 1.02K | lm_loss: 8.47 | lr: 6.21e-05 | model_tflops_per_gpu: 216 | hardware_tflops_per_gpu: 216 | grad_norm: 5.23
224
+ [default0]:07/03/2024 22:46:24 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: Memory usage: 4612.90MiB. Peak allocated 39056.83MiB. Peak reserved: 39710.00MiB
225
+ [default0]:07/03/2024 22:46:46 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: Memory usage: 4612.90MiB. Peak allocated 39056.83MiB. Peak reserved: 39710.00MiB
226
+ [default6]:07/03/2024 22:46:46 [INFO|DP=0|PP=3|TP=0|ip-26-0-164-207]: iteration: 10 / 20 | consumed_tokens: 41.9M | elapsed_time_per_iteration_ms: 22K | tokens_per_sec: 191K | tokens_per_sec_per_gpu: 23.9K | global_batch_size: 1.02K | lm_loss: 8.17 | lr: 5.74e-05 | model_tflops_per_gpu: 217 | hardware_tflops_per_gpu: 217 | grad_norm: 7.71
227
+ [default0]:07/03/2024 22:47:08 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: Memory usage: 4612.90MiB. Peak allocated 39056.83MiB. Peak reserved: 39710.00MiB
228
+ [default6]:07/03/2024 22:47:08 [INFO|DP=0|PP=3|TP=0|ip-26-0-164-207]: iteration: 11 / 20 | consumed_tokens: 46.1M | elapsed_time_per_iteration_ms: 22K | tokens_per_sec: 191K | tokens_per_sec_per_gpu: 23.9K | global_batch_size: 1.02K | lm_loss: 7.93 | lr: 5.26e-05 | model_tflops_per_gpu: 217 | hardware_tflops_per_gpu: 217 | grad_norm: 5.53
229
+ [default6]:07/03/2024 22:47:30 [INFO|DP=0|PP=3|TP=0|ip-26-0-164-207]: iteration: 12 / 20 | consumed_tokens: 50.3M | elapsed_time_per_iteration_ms: 22K | tokens_per_sec: 191K | tokens_per_sec_per_gpu: 23.9K | global_batch_size: 1.02K | lm_loss: 7.75 | lr: 4.79e-05 | model_tflops_per_gpu: 216 | hardware_tflops_per_gpu: 216 | grad_norm: 4.64
230
+ [default0]:07/03/2024 22:47:30 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: Memory usage: 4612.90MiB. Peak allocated 39056.83MiB. Peak reserved: 39710.00MiB
231
+ [default0]:07/03/2024 22:47:52 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: Memory usage: 4612.90MiB. Peak allocated 39056.83MiB. Peak reserved: 39710.00MiB
232
+ [default6]:07/03/2024 22:47:52 [INFO|DP=0|PP=3|TP=0|ip-26-0-164-207]: iteration: 13 / 20 | consumed_tokens: 54.5M | elapsed_time_per_iteration_ms: 22K | tokens_per_sec: 191K | tokens_per_sec_per_gpu: 23.8K | global_batch_size: 1.02K | lm_loss: 7.58 | lr: 4.32e-05 | model_tflops_per_gpu: 216 | hardware_tflops_per_gpu: 216 | grad_norm: 2.9
233
+ [default6]:07/03/2024 22:48:14 [INFO|DP=0|PP=3|TP=0|ip-26-0-164-207]: iteration: 14 / 20 | consumed_tokens: 58.7M | elapsed_time_per_iteration_ms: 22K | tokens_per_sec: 191K | tokens_per_sec_per_gpu: 23.9K | global_batch_size: 1.02K | lm_loss: 7.5 | lr: 3.84e-05 | model_tflops_per_gpu: 216 | hardware_tflops_per_gpu: 216 | grad_norm: 4.18
234
+ [default0]:07/03/2024 22:48:14 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: Memory usage: 4612.90MiB. Peak allocated 39056.83MiB. Peak reserved: 39710.00MiB
235
+ [default0]:07/03/2024 22:48:36 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: Memory usage: 4612.90MiB. Peak allocated 39056.83MiB. Peak reserved: 39710.00MiB
236
+ [default6]:07/03/2024 22:48:36 [INFO|DP=0|PP=3|TP=0|ip-26-0-164-207]: iteration: 15 / 20 | consumed_tokens: 62.9M | elapsed_time_per_iteration_ms: 22K | tokens_per_sec: 191K | tokens_per_sec_per_gpu: 23.9K | global_batch_size: 1.02K | lm_loss: 7.4 | lr: 3.37e-05 | model_tflops_per_gpu: 216 | hardware_tflops_per_gpu: 216 | grad_norm: 3.86
237
+ [default0]:07/03/2024 22:48:58 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: Memory usage: 4612.90MiB. Peak allocated 39056.83MiB. Peak reserved: 39710.00MiB
238
+ [default6]:07/03/2024 22:48:58 [INFO|DP=0|PP=3|TP=0|ip-26-0-164-207]: iteration: 16 / 20 | consumed_tokens: 67.1M | elapsed_time_per_iteration_ms: 22K | tokens_per_sec: 190K | tokens_per_sec_per_gpu: 23.8K | global_batch_size: 1.02K | lm_loss: 7.28 | lr: 2.89e-05 | model_tflops_per_gpu: 216 | hardware_tflops_per_gpu: 216 | grad_norm: 3.06
239
+ [default0]:07/03/2024 22:49:20 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: Memory usage: 4612.90MiB. Peak allocated 39056.83MiB. Peak reserved: 39710.00MiB
240
+ [default6]:07/03/2024 22:49:20 [INFO|DP=0|PP=3|TP=0|ip-26-0-164-207]: iteration: 17 / 20 | consumed_tokens: 71.3M | elapsed_time_per_iteration_ms: 22K | tokens_per_sec: 190K | tokens_per_sec_per_gpu: 23.8K | global_batch_size: 1.02K | lm_loss: 7.19 | lr: 2.42e-05 | model_tflops_per_gpu: 216 | hardware_tflops_per_gpu: 216 | grad_norm: 2.39
241
+ [default0]:07/03/2024 22:49:42 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: Memory usage: 4612.90MiB. Peak allocated 39056.83MiB. Peak reserved: 39710.00MiB
242
+ [default6]:07/03/2024 22:49:42 [INFO|DP=0|PP=3|TP=0|ip-26-0-164-207]: iteration: 18 / 20 | consumed_tokens: 75.5M | elapsed_time_per_iteration_ms: 22K | tokens_per_sec: 190K | tokens_per_sec_per_gpu: 23.8K | global_batch_size: 1.02K | lm_loss: 7.13 | lr: 1.95e-05 | model_tflops_per_gpu: 216 | hardware_tflops_per_gpu: 216 | grad_norm: 2.21
243
+ [default0]:07/03/2024 22:50:04 [INFO|DP=0|PP=0|TP=0|ip-26-0-164-207]: Memory usage: 4612.90MiB. Peak allocated 39056.83MiB. Peak reserved: 39710.00MiB
244
+ [default6]:07/03/2024 22:50:04 [INFO|DP=0|PP=3|TP=0|ip-26-0-164-207]: iteration: 19 / 20 | consumed_tokens: 79.7M | elapsed_time_per_iteration_ms: 22K | tokens_per_sec: 190K | tokens_per_sec_per_gpu: 23.8K | global_batch_size: 1.02K | lm_loss: 7.08 | lr: 1.47e-05 | model_tflops_per_gpu: 216 | hardware_tflops_per_gpu: 216 | grad_norm: 2.64
245
+ [default6]:07/03/2024 22:50:26 [INFO|DP=0|PP=3|TP=0|ip-26-0-164-207]: iteration: 20 / 20 | consumed_tokens: 83.9M | elapsed_time_per_iteration_ms: 22K | tokens_per_sec: 191K | tokens_per_sec_per_gpu: 23.8K | global_batch_size: 1.02K | lm_loss: 7.03 | lr: 1e-05 | model_tflops_per_gpu: 216 | hardware_tflops_per_gpu: 216 | grad_norm: 2.3
246
+ Saved 1 csv files over 1 completed logs
247
+ Processing file: /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-2_tp-1_pp-4_mbz-4/profiler/ip-26-0-164-207_605035.1720046681997701534.pt.trace.json
248
+ Results written to /fsx/ferdinandmom/ferdinand-hf/bench_cluster/results/llama-1B/8_GPUS/dp-2_tp-1_pp-4_mbz-4/profiler.csv
249
+ 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.
250
+
llama-1B/8_GPUS/dp-2_tp-1_pp-4_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
+ 1,4190000.0000000005,36300.0,115000.0,14400.0,1020.0,11.1,0.0001,131.0,131.0,25.1,4612.9,39056.83,39710.0
3
+ 2,8390000.0,22000.0,190000.0,23800.0,1020.0,11.1,9.53e-05,216.0,216.0,25.2,4612.9,6503.83,39710.0
4
+ 3,12600000.0,22000.0,191000.0,23800.0,1020.0,11.4,9.05e-05,216.0,216.0,217.0,4612.9,6503.83,39710.0
5
+ 4,16800000.0,22000.0,191000.0,23800.0,1020.0,13.8,8.58e-05,216.0,216.0,22.5,,,
6
+ 5,21000000.0,22000.0,191000.0,23800.0,1020.0,9.98,8.11e-05,216.0,216.0,16.4,4612.9,39056.83,39710.0
7
+ 6,25200000.0,22000.0,191000.0,23900.0,1020.0,10.9,7.63e-05,217.0,217.0,93.8,4612.9,39056.83,39710.0
8
+ 7,29400000.0,198000.0,21200.0,2650.0,1020.0,9.16,7.16e-05,24.0,24.0,19.8,,,
9
+ 8,33600000.0,22000.0,191000.0,23900.0,1020.0,8.83,6.68e-05,216.0,216.0,6.08,4612.9,39056.83,39710.0
10
+ 9,37700000.0,22000.0,190000.0,23800.0,1020.0,8.47,6.21e-05,216.0,216.0,5.23,4612.9,39056.83,39710.0
11
+ 10,41900000.0,22000.0,191000.0,23900.0,1020.0,8.17,5.74e-05,217.0,217.0,7.71,4612.9,39056.83,39710.0
12
+ 11,46100000.0,22000.0,191000.0,23900.0,1020.0,7.93,5.26e-05,217.0,217.0,5.53,,,
13
+ 12,50300000.0,22000.0,191000.0,23900.0,1020.0,7.75,4.79e-05,216.0,216.0,4.64,4612.9,39056.83,39710.0
14
+ 13,54500000.0,22000.0,191000.0,23800.0,1020.0,7.58,4.32e-05,216.0,216.0,2.9,,,
15
+ 14,58700000.0,22000.0,191000.0,23900.0,1020.0,7.5,3.84e-05,216.0,216.0,4.18,4612.9,39056.83,39710.0
16
+ 15,62900000.0,22000.0,191000.0,23900.0,1020.0,7.4,3.37e-05,216.0,216.0,3.86,4612.9,39056.83,39710.0
17
+ 16,67099999.99999999,22000.0,190000.0,23800.0,1020.0,7.28,2.89e-05,216.0,216.0,3.06,4612.9,39056.83,39710.0
18
+ 17,71300000.0,22000.0,190000.0,23800.0,1020.0,7.19,2.42e-05,216.0,216.0,2.39,4612.9,39056.83,39710.0
19
+ 18,75500000.0,22000.0,190000.0,23800.0,1020.0,7.13,1.95e-05,216.0,216.0,2.21,4612.9,39056.83,39710.0
20
+ 19,79700000.0,22000.0,190000.0,23800.0,1020.0,7.08,1.47e-05,216.0,216.0,2.64,,,
21
+ 20,83900000.0,22000.0,191000.0,23800.0,1020.0,7.03,1e-05,216.0,216.0,2.3,,,
llama-1B/8_GPUS/dp-2_tp-1_pp-4_mbz-4/profiler.csv ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ forward,backward
2
+ 0ms 798μs,1ms 258μs
llama-1B/8_GPUS/dp-2_tp-1_pp-4_mbz-4/profiler/ip-26-0-164-207_605035.1720046681997701534.pt.trace.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7869b63681b355e0a935df130807ee592f0ae02414b17ffa8329c1b1497dbc6c
3
+ size 5129472127
llama-1B/8_GPUS/dp-2_tp-1_pp-4_mbz-4/status.txt ADDED
@@ -0,0 +1 @@
 
 
1
+ completed