diff --git "a/attnserver.run_attnserver.slurm.sh.343248.err.log" "b/attnserver.run_attnserver.slurm.sh.343248.err.log" new file mode 100644--- /dev/null +++ "b/attnserver.run_attnserver.slurm.sh.343248.err.log" @@ -0,0 +1,1981 @@ ++ source /mnt/weka/home/hao.zhang/conda/miniconda/bin/activate +++ _CONDA_ROOT=/mnt/weka/home/hao.zhang/conda/miniconda +++ . /mnt/weka/home/hao.zhang/conda/miniconda/etc/profile.d/conda.sh ++++ export CONDA_EXE=/mnt/weka/home/hao.zhang/conda/miniconda/bin/conda ++++ CONDA_EXE=/mnt/weka/home/hao.zhang/conda/miniconda/bin/conda ++++ export _CE_M= ++++ _CE_M= ++++ export _CE_CONDA= ++++ _CE_CONDA= ++++ export CONDA_PYTHON_EXE=/mnt/weka/home/hao.zhang/conda/miniconda/bin/python ++++ CONDA_PYTHON_EXE=/mnt/weka/home/hao.zhang/conda/miniconda/bin/python ++++ '[' -z x ']' +++ conda activate +++ local cmd=activate +++ case "$cmd" in +++ __conda_activate activate +++ '[' -n '' ']' +++ local ask_conda ++++ PS1= ++++ __conda_exe shell.posix activate ++++ '[' -n '' ']' ++++ /mnt/weka/home/hao.zhang/conda/miniconda/bin/conda shell.posix activate +++ ask_conda='unset _CE_M +unset _CE_CONDA +PS1='\''(base) '\'' +export PATH='\''/mnt/weka/home/hao.zhang/.local/bin:/mnt/weka/home/hao.zhang/.local/bin:/mnt/weka/home/hao.zhang/.local/bin:/mnt/weka/home/hao.zhang/conda/miniconda/bin:/mnt/weka/home/hao.zhang/conda/miniconda/condabin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin:/usr/games:/usr/local/games:/snap/bin'\'' +export CONDA_SHLVL='\''1'\'' +export CONDA_PROMPT_MODIFIER='\''(base) '\'' +export CONDA_EXE='\''/mnt/weka/home/hao.zhang/conda/miniconda/bin/conda'\'' +export CONDA_PYTHON_EXE='\''/mnt/weka/home/hao.zhang/conda/miniconda/bin/python'\''' +++ eval 'unset _CE_M +unset _CE_CONDA +PS1='\''(base) '\'' +export PATH='\''/mnt/weka/home/hao.zhang/.local/bin:/mnt/weka/home/hao.zhang/.local/bin:/mnt/weka/home/hao.zhang/.local/bin:/mnt/weka/home/hao.zhang/conda/miniconda/bin:/mnt/weka/home/hao.zhang/conda/miniconda/condabin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin:/usr/games:/usr/local/games:/snap/bin'\'' +export CONDA_SHLVL='\''1'\'' +export CONDA_PROMPT_MODIFIER='\''(base) '\'' +export CONDA_EXE='\''/mnt/weka/home/hao.zhang/conda/miniconda/bin/conda'\'' +export CONDA_PYTHON_EXE='\''/mnt/weka/home/hao.zhang/conda/miniconda/bin/python'\''' ++++ unset _CE_M ++++ unset _CE_CONDA ++++ PS1='(base) ' ++++ export PATH=/mnt/weka/home/hao.zhang/.local/bin:/mnt/weka/home/hao.zhang/.local/bin:/mnt/weka/home/hao.zhang/.local/bin:/mnt/weka/home/hao.zhang/conda/miniconda/bin:/mnt/weka/home/hao.zhang/conda/miniconda/condabin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin:/usr/games:/usr/local/games:/snap/bin ++++ PATH=/mnt/weka/home/hao.zhang/.local/bin:/mnt/weka/home/hao.zhang/.local/bin:/mnt/weka/home/hao.zhang/.local/bin:/mnt/weka/home/hao.zhang/conda/miniconda/bin:/mnt/weka/home/hao.zhang/conda/miniconda/condabin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin:/usr/games:/usr/local/games:/snap/bin ++++ export CONDA_SHLVL=1 ++++ CONDA_SHLVL=1 ++++ export 'CONDA_PROMPT_MODIFIER=(base) ' ++++ CONDA_PROMPT_MODIFIER='(base) ' ++++ export CONDA_EXE=/mnt/weka/home/hao.zhang/conda/miniconda/bin/conda ++++ CONDA_EXE=/mnt/weka/home/hao.zhang/conda/miniconda/bin/conda ++++ export CONDA_PYTHON_EXE=/mnt/weka/home/hao.zhang/conda/miniconda/bin/python ++++ CONDA_PYTHON_EXE=/mnt/weka/home/hao.zhang/conda/miniconda/bin/python +++ __conda_hashr +++ '[' -n '' ']' +++ '[' -n '' ']' +++ hash -r ++ conda activate junda-attnserver ++ local cmd=activate ++ case "$cmd" in ++ __conda_activate activate junda-attnserver ++ '[' -n '' ']' ++ local ask_conda +++ PS1='(base) ' +++ __conda_exe shell.posix activate junda-attnserver +++ '[' -n '' ']' +++ /mnt/weka/home/hao.zhang/conda/miniconda/bin/conda shell.posix activate junda-attnserver ++ ask_conda='unset _CE_M +unset _CE_CONDA +PS1='\''(junda-attnserver) '\'' +export PATH='\''/mnt/weka/home/hao.zhang/.local/bin:/mnt/weka/home/hao.zhang/.local/bin:/mnt/weka/home/hao.zhang/.local/bin:/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/bin:/mnt/weka/home/hao.zhang/conda/miniconda/condabin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin:/usr/games:/usr/local/games:/snap/bin'\'' +export CONDA_PREFIX='\''/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver'\'' +export CONDA_SHLVL='\''2'\'' +export CONDA_DEFAULT_ENV='\''junda-attnserver'\'' +export CONDA_PROMPT_MODIFIER='\''(junda-attnserver) '\'' +export CONDA_PREFIX_1='\''/mnt/weka/home/hao.zhang/conda/miniconda'\'' +export CONDA_EXE='\''/mnt/weka/home/hao.zhang/conda/miniconda/bin/conda'\'' +export CONDA_PYTHON_EXE='\''/mnt/weka/home/hao.zhang/conda/miniconda/bin/python'\''' ++ eval 'unset _CE_M +unset _CE_CONDA +PS1='\''(junda-attnserver) '\'' +export PATH='\''/mnt/weka/home/hao.zhang/.local/bin:/mnt/weka/home/hao.zhang/.local/bin:/mnt/weka/home/hao.zhang/.local/bin:/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/bin:/mnt/weka/home/hao.zhang/conda/miniconda/condabin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin:/usr/games:/usr/local/games:/snap/bin'\'' +export CONDA_PREFIX='\''/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver'\'' +export CONDA_SHLVL='\''2'\'' +export CONDA_DEFAULT_ENV='\''junda-attnserver'\'' +export CONDA_PROMPT_MODIFIER='\''(junda-attnserver) '\'' +export CONDA_PREFIX_1='\''/mnt/weka/home/hao.zhang/conda/miniconda'\'' +export CONDA_EXE='\''/mnt/weka/home/hao.zhang/conda/miniconda/bin/conda'\'' +export CONDA_PYTHON_EXE='\''/mnt/weka/home/hao.zhang/conda/miniconda/bin/python'\''' +++ unset _CE_M +++ unset _CE_CONDA +++ PS1='(junda-attnserver) ' +++ export PATH=/mnt/weka/home/hao.zhang/.local/bin:/mnt/weka/home/hao.zhang/.local/bin:/mnt/weka/home/hao.zhang/.local/bin:/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/bin:/mnt/weka/home/hao.zhang/conda/miniconda/condabin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin:/usr/games:/usr/local/games:/snap/bin +++ PATH=/mnt/weka/home/hao.zhang/.local/bin:/mnt/weka/home/hao.zhang/.local/bin:/mnt/weka/home/hao.zhang/.local/bin:/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/bin:/mnt/weka/home/hao.zhang/conda/miniconda/condabin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin:/usr/games:/usr/local/games:/snap/bin +++ export CONDA_PREFIX=/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver +++ CONDA_PREFIX=/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver +++ export CONDA_SHLVL=2 +++ CONDA_SHLVL=2 +++ export CONDA_DEFAULT_ENV=junda-attnserver +++ CONDA_DEFAULT_ENV=junda-attnserver +++ export 'CONDA_PROMPT_MODIFIER=(junda-attnserver) ' +++ CONDA_PROMPT_MODIFIER='(junda-attnserver) ' +++ export CONDA_PREFIX_1=/mnt/weka/home/hao.zhang/conda/miniconda +++ CONDA_PREFIX_1=/mnt/weka/home/hao.zhang/conda/miniconda +++ export CONDA_EXE=/mnt/weka/home/hao.zhang/conda/miniconda/bin/conda +++ CONDA_EXE=/mnt/weka/home/hao.zhang/conda/miniconda/bin/conda +++ export CONDA_PYTHON_EXE=/mnt/weka/home/hao.zhang/conda/miniconda/bin/python +++ CONDA_PYTHON_EXE=/mnt/weka/home/hao.zhang/conda/miniconda/bin/python ++ __conda_hashr ++ '[' -n '' ']' ++ '[' -n '' ']' ++ hash -r ++ export CHROME_TRACE_PREFIX=/mnt/sharefs/users/hao.zhang/junda/megatron-prof-data--unstable-v5 ++ CHROME_TRACE_PREFIX=/mnt/sharefs/users/hao.zhang/junda/megatron-prof-data--unstable-v5 ++ mkdir -p /mnt/sharefs/users/hao.zhang/junda/megatron-prof-data--unstable-v5 ++ export PROF_TP_SIZE=2 ++ PROF_TP_SIZE=2 ++ export PROF_CP_SIZE=4 ++ PROF_CP_SIZE=4 ++ export PROF_BS=32 ++ PROF_BS=32 ++ for ctx_length in 1024 2048 4096 8192 12288 16384 24576 32768 40960 49152 65536 81920 98304 131072 ++ export PROF_CTX_LENGTH=1024 ++ PROF_CTX_LENGTH=1024 ++ name='/mnt/sharefs/users/hao.zhang/junda/megatron-prof-data--unstable-v5/mytrace.L1024*tp2.cp4.bs32.json' ++ '[' -f '/mnt/sharefs/users/hao.zhang/junda/megatron-prof-data--unstable-v5/mytrace.L1024*tp2.cp4.bs32.json' ']' ++ echo 'Running ctx_length=1024, TP_SIZE=2, CP_SIZE=4, BATCH_SIZE=32' ++ srun bash ./attnserver.sh ++ which python3 ++ python3 -m torch.distributed.launch --nproc_per_node 8 --nnodes 1 --node_rank 0 --rdzv_id 343248 --rdzv_backend c10d --rdzv_endpoint fs-mbz-gpu-791:29500 ./pretrain_gpt_profile.py --tensor-model-parallel-size 2 --context-parallel-size 4 --num-layers 2 --hidden-size 4096 --num-attention-heads 64 --group-query-attention --num-query-groups 16 --seq-length 1024 --max-position-embeddings 1024 --micro-batch-size 1 --global-batch-size 1 --lr 0.0005 --train-iters 10 --lr-decay-iters 150000 --lr-decay-style cosine --lr-warmup-iters 2 --weight-decay .1 --adam-beta2 .999 --fp16 --log-interval 1 --save-interval 16 --eval-interval 16 --eval-iters 1 --vocab-file vocab.json --merge-file merges.txt --save gpt-checkpoint --load gpt-checkpoint --logging-level 0 --mock-data --tensorboard-dir tensorboard-logs/ +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/launch.py:207: FutureWarning: The module torch.distributed.launch is deprecated +and will be removed in future. Use torchrun. +Note that --use-env is set by default in torchrun. +If your script expects `--local-rank` argument to be set, please +change it to read from `os.environ['LOCAL_RANK']` instead. See +https://pytorch.org/docs/stable/distributed.html#launch-utility for +further instructions + + main() +W0621 22:03:02.732000 2388828 site-packages/torch/distributed/run.py:766] +W0621 22:03:02.732000 2388828 site-packages/torch/distributed/run.py:766] ***************************************** +W0621 22:03:02.732000 2388828 site-packages/torch/distributed/run.py:766] 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. +W0621 22:03:02.732000 2388828 site-packages/torch/distributed/run.py:766] ***************************************** +[rank6]:[W621 22:03:25.379133613 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 6] using GPU 6 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank4]:[W621 22:03:25.379140210 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 4] using GPU 4 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank2]:[W621 22:03:25.379140116 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 2] using GPU 2 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank5]:[W621 22:03:26.387887715 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 5] using GPU 5 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank1]:[W621 22:03:26.387899823 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 1] using GPU 1 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank7]:[W621 22:03:26.387917210 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 7] using GPU 7 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank3]:[W621 22:03:26.388422226 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 3] using GPU 3 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank0]:[W621 22:03:26.510140421 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 0] using GPU 0 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +[rank7]: Traceback (most recent call last): +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 554, in +[rank7]: pretrain( +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 863, in pretrain +[rank7]: iteration, num_floating_point_operations_so_far = train( +[rank7]: ^^^^^^ +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 2229, in train +[rank7]: ) = train_step( +[rank7]: ^^^^^^^^^^^ +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 1382, in train_step +[rank7]: losses_reduced = forward_backward_func( +[rank7]: ^^^^^^^^^^^^^^^^^^^^^^ +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 535, in forward_backward_no_pipelining +[rank7]: backward_step(input_tensor, output_tensor, output_tensor_grad, model_type, config) +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 409, in backward_step +[rank7]: custom_backward(output_tensor[0], output_tensor_grad[0]) +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 160, in custom_backward +[rank7]: Variable._execution_engine.run_backward( +[rank7]: File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/autograd/function.py", line 307, in apply +[rank7]: return user_fn(self, *args) +[rank7]: ^^^^^^^^^^^^^^^^^^^^ +[rank7]: File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/amp/autocast_mode.py", line 556, in decorate_bwd +[rank7]: return bwd(*args, **kwargs) +[rank7]: ^^^^^^^^^^^^^^^^^^^^ +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/tensor_parallel/layers.py", line 495, in backward +[rank7]: grad_input = grad_output.matmul(weight) +[rank7]: ^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank7]: RuntimeError: CUDA error: CUBLAS_STATUS_ALLOC_FAILED when calling `cublasCreate(handle)` +[rank6]: Traceback (most recent call last): +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 554, in +[rank6]: pretrain( +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 863, in pretrain +[rank6]: iteration, num_floating_point_operations_so_far = train( +[rank6]: ^^^^^^ +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 2229, in train +[rank6]: ) = train_step( +[rank6]: ^^^^^^^^^^^ +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 1382, in train_step +[rank6]: losses_reduced = forward_backward_func( +[rank6]: ^^^^^^^^^^^^^^^^^^^^^^ +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 535, in forward_backward_no_pipelining +[rank6]: backward_step(input_tensor, output_tensor, output_tensor_grad, model_type, config) +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 409, in backward_step +[rank6]: custom_backward(output_tensor[0], output_tensor_grad[0]) +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 160, in custom_backward +[rank6]: Variable._execution_engine.run_backward( +[rank6]: File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/autograd/function.py", line 307, in apply +[rank6]: return user_fn(self, *args) +[rank6]: ^^^^^^^^^^^^^^^^^^^^ +[rank6]: File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/amp/autocast_mode.py", line 556, in decorate_bwd +[rank6]: return bwd(*args, **kwargs) +[rank6]: ^^^^^^^^^^^^^^^^^^^^ +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/tensor_parallel/layers.py", line 495, in backward +[rank6]: grad_input = grad_output.matmul(weight) +[rank6]: ^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank6]: RuntimeError: CUDA error: CUBLAS_STATUS_ALLOC_FAILED when calling `cublasCreate(handle)` +W0621 22:03:53.321000 2388828 site-packages/torch/distributed/elastic/multiprocessing/api.py:900] Sending process 2388899 closing signal SIGTERM +W0621 22:03:53.324000 2388828 site-packages/torch/distributed/elastic/multiprocessing/api.py:900] Sending process 2388900 closing signal SIGTERM +W0621 22:03:53.329000 2388828 site-packages/torch/distributed/elastic/multiprocessing/api.py:900] Sending process 2388901 closing signal SIGTERM +W0621 22:03:53.333000 2388828 site-packages/torch/distributed/elastic/multiprocessing/api.py:900] Sending process 2388902 closing signal SIGTERM +W0621 22:03:53.344000 2388828 site-packages/torch/distributed/elastic/multiprocessing/api.py:900] Sending process 2388903 closing signal SIGTERM +W0621 22:03:53.347000 2388828 site-packages/torch/distributed/elastic/multiprocessing/api.py:900] Sending process 2388904 closing signal SIGTERM +W0621 22:03:53.351000 2388828 site-packages/torch/distributed/elastic/multiprocessing/api.py:900] Sending process 2388905 closing signal SIGTERM +E0621 22:03:56.055000 2388828 site-packages/torch/distributed/elastic/multiprocessing/api.py:874] failed (exitcode: 1) local_rank: 7 (pid: 2388906) of binary: /mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/bin/python3 +Traceback (most recent call last): + File "", line 198, in _run_module_as_main + File "", line 88, in _run_code + File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/launch.py", line 207, in + main() + File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/typing_extensions.py", line 3253, in wrapper + return arg(*args, **kwargs) + ^^^^^^^^^^^^^^^^^^^^ + File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/launch.py", line 203, in main + launch(args) + File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/launch.py", line 188, in launch + run(args) + File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/run.py", line 883, in run + elastic_launch( + File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/launcher/api.py", line 139, in __call__ + return launch_agent(self._config, self._entrypoint, list(args)) + ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ + File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/launcher/api.py", line 270, in launch_agent + raise ChildFailedError( +torch.distributed.elastic.multiprocessing.errors.ChildFailedError: +============================================================ +./pretrain_gpt_profile.py FAILED +------------------------------------------------------------ +Failures: + +------------------------------------------------------------ +Root Cause (first observed failure): +[0]: + time : 2025-06-21_22:03:53 + host : fs-mbz-gpu-791 + rank : 7 (local_rank: 7) + exitcode : 1 (pid: 2388906) + error_file: + traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html +============================================================ ++ set +x ++ for ctx_length in 1024 2048 4096 8192 12288 16384 24576 32768 40960 49152 65536 81920 98304 131072 ++ export PROF_CTX_LENGTH=2048 ++ PROF_CTX_LENGTH=2048 ++ name='/mnt/sharefs/users/hao.zhang/junda/megatron-prof-data--unstable-v5/mytrace.L2048*tp2.cp4.bs32.json' ++ '[' -f '/mnt/sharefs/users/hao.zhang/junda/megatron-prof-data--unstable-v5/mytrace.L2048*tp2.cp4.bs32.json' ']' ++ echo 'Running ctx_length=2048, TP_SIZE=2, CP_SIZE=4, BATCH_SIZE=32' ++ srun bash ./attnserver.sh ++ which python3 ++ python3 -m torch.distributed.launch --nproc_per_node 8 --nnodes 1 --node_rank 0 --rdzv_id 343248 --rdzv_backend c10d --rdzv_endpoint fs-mbz-gpu-791:29500 ./pretrain_gpt_profile.py --tensor-model-parallel-size 2 --context-parallel-size 4 --num-layers 2 --hidden-size 4096 --num-attention-heads 64 --group-query-attention --num-query-groups 16 --seq-length 2048 --max-position-embeddings 2048 --micro-batch-size 1 --global-batch-size 1 --lr 0.0005 --train-iters 10 --lr-decay-iters 150000 --lr-decay-style cosine --lr-warmup-iters 2 --weight-decay .1 --adam-beta2 .999 --fp16 --log-interval 1 --save-interval 16 --eval-interval 16 --eval-iters 1 --vocab-file vocab.json --merge-file merges.txt --save gpt-checkpoint --load gpt-checkpoint --logging-level 0 --mock-data --tensorboard-dir tensorboard-logs/ +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/launch.py:207: FutureWarning: The module torch.distributed.launch is deprecated +and will be removed in future. Use torchrun. +Note that --use-env is set by default in torchrun. +If your script expects `--local-rank` argument to be set, please +change it to read from `os.environ['LOCAL_RANK']` instead. See +https://pytorch.org/docs/stable/distributed.html#launch-utility for +further instructions + + main() +W0621 22:03:59.105000 2391961 site-packages/torch/distributed/run.py:766] +W0621 22:03:59.105000 2391961 site-packages/torch/distributed/run.py:766] ***************************************** +W0621 22:03:59.105000 2391961 site-packages/torch/distributed/run.py:766] 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. +W0621 22:03:59.105000 2391961 site-packages/torch/distributed/run.py:766] ***************************************** +[rank4]:[W621 22:04:20.468105838 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 4] using GPU 4 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank6]:[W621 22:04:20.468108295 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 6] using GPU 6 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank2]:[W621 22:04:20.468134882 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 2] using GPU 2 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank5]:[W621 22:04:20.474106212 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 5] using GPU 5 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank1]:[W621 22:04:20.474189208 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 1] using GPU 1 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank7]:[W621 22:04:20.474870802 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 7] using GPU 7 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank3]:[W621 22:04:20.474983592 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 3] using GPU 3 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank0]:[W621 22:04:20.609302527 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 0] using GPU 0 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +[rank3]: Traceback (most recent call last): +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 554, in +[rank3]: pretrain( +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 863, in pretrain +[rank3]: iteration, num_floating_point_operations_so_far = train( +[rank3]: ^^^^^^ +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 2229, in train +[rank3]: ) = train_step( +[rank3]: ^^^^^^^^^^^ +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 1382, in train_step +[rank3]: losses_reduced = forward_backward_func( +[rank3]: ^^^^^^^^^^^^^^^^^^^^^^ +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 518, in forward_backward_no_pipelining +[rank3]: output_tensor, num_tokens = forward_step( +[rank3]: ^^^^^^^^^^^^^ +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 289, in forward_step +[rank3]: output_tensor, loss_func = forward_step_func(data_iterator, model) +[rank3]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 446, in forward_step +[rank3]: (tokens, labels, loss_mask, attention_mask, position_ids), token_lens = get_batch(data_iterator) +[rank3]: ^^^^^^^^^^^^^^^^^^^^^^^^ +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 293, in get_batch +[rank3]: batch = get_batch_on_this_cp_rank(batch) +[rank3]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/utils.py", line 1765, in get_batch_on_this_cp_rank +[rank3]: val = val.index_select(seq_dim, index) +[rank3]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank3]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 32.00 GiB. GPU 3 has a total capacity of 139.81 GiB of which 8.10 GiB is free. Including non-PyTorch memory, this process has 131.70 GiB memory in use. Of the allocated memory 130.21 GiB is allocated by PyTorch, and 37.54 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank7]: Traceback (most recent call last): +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 554, in +[rank7]: pretrain( +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 863, in pretrain +[rank7]: iteration, num_floating_point_operations_so_far = train( +[rank7]: ^^^^^^ +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 2229, in train +[rank7]: ) = train_step( +[rank7]: ^^^^^^^^^^^ +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 1382, in train_step +[rank7]: losses_reduced = forward_backward_func( +[rank7]: ^^^^^^^^^^^^^^^^^^^^^^ +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 518, in forward_backward_no_pipelining +[rank7]: output_tensor, num_tokens = forward_step( +[rank7]: ^^^^^^^^^^^^^ +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 289, in forward_step +[rank7]: output_tensor, loss_func = forward_step_func(data_iterator, model) +[rank7]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 446, in forward_step +[rank7]: (tokens, labels, loss_mask, attention_mask, position_ids), token_lens = get_batch(data_iterator) +[rank7]: ^^^^^^^^^^^^^^^^^^^^^^^^ +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 293, in get_batch +[rank7]: batch = get_batch_on_this_cp_rank(batch) +[rank7]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/utils.py", line 1765, in get_batch_on_this_cp_rank +[rank7]: val = val.index_select(seq_dim, index) +[rank7]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank7]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 32.00 GiB. GPU 7 has a total capacity of 139.81 GiB of which 8.10 GiB is free. Including non-PyTorch memory, this process has 131.70 GiB memory in use. Of the allocated memory 130.21 GiB is allocated by PyTorch, and 37.54 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank1]: Traceback (most recent call last): +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 554, in +[rank1]: pretrain( +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 863, in pretrain +[rank1]: iteration, num_floating_point_operations_so_far = train( +[rank1]: ^^^^^^ +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 2229, in train +[rank1]: ) = train_step( +[rank1]: ^^^^^^^^^^^ +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 1382, in train_step +[rank1]: losses_reduced = forward_backward_func( +[rank1]: ^^^^^^^^^^^^^^^^^^^^^^ +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 518, in forward_backward_no_pipelining +[rank1]: output_tensor, num_tokens = forward_step( +[rank1]: ^^^^^^^^^^^^^ +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 289, in forward_step +[rank1]: output_tensor, loss_func = forward_step_func(data_iterator, model) +[rank1]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 446, in forward_step +[rank1]: (tokens, labels, loss_mask, attention_mask, position_ids), token_lens = get_batch(data_iterator) +[rank1]: ^^^^^^^^^^^^^^^^^^^^^^^^ +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 293, in get_batch +[rank1]: batch = get_batch_on_this_cp_rank(batch) +[rank1]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/utils.py", line 1765, in get_batch_on_this_cp_rank +[rank1]: val = val.index_select(seq_dim, index) +[rank1]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank1]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 32.00 GiB. GPU 1 has a total capacity of 139.81 GiB of which 8.10 GiB is free. Including non-PyTorch memory, this process has 131.70 GiB memory in use. Of the allocated memory 130.21 GiB is allocated by PyTorch, and 37.54 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank5]: Traceback (most recent call last): +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 554, in +[rank5]: pretrain( +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 863, in pretrain +[rank5]: iteration, num_floating_point_operations_so_far = train( +[rank5]: ^^^^^^ +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 2229, in train +[rank5]: ) = train_step( +[rank5]: ^^^^^^^^^^^ +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 1382, in train_step +[rank5]: losses_reduced = forward_backward_func( +[rank5]: ^^^^^^^^^^^^^^^^^^^^^^ +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 518, in forward_backward_no_pipelining +[rank5]: output_tensor, num_tokens = forward_step( +[rank5]: ^^^^^^^^^^^^^ +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 289, in forward_step +[rank5]: output_tensor, loss_func = forward_step_func(data_iterator, model) +[rank5]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 446, in forward_step +[rank5]: (tokens, labels, loss_mask, attention_mask, position_ids), token_lens = get_batch(data_iterator) +[rank5]: ^^^^^^^^^^^^^^^^^^^^^^^^ +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 293, in get_batch +[rank5]: batch = get_batch_on_this_cp_rank(batch) +[rank5]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/utils.py", line 1765, in get_batch_on_this_cp_rank +[rank5]: val = val.index_select(seq_dim, index) +[rank5]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank5]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 32.00 GiB. GPU 5 has a total capacity of 139.81 GiB of which 8.10 GiB is free. Including non-PyTorch memory, this process has 131.70 GiB memory in use. Of the allocated memory 130.21 GiB is allocated by PyTorch, and 37.54 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank6]: Traceback (most recent call last): +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 554, in +[rank6]: pretrain( +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 863, in pretrain +[rank6]: iteration, num_floating_point_operations_so_far = train( +[rank6]: ^^^^^^ +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 2229, in train +[rank6]: ) = train_step( +[rank6]: ^^^^^^^^^^^ +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 1382, in train_step +[rank6]: losses_reduced = forward_backward_func( +[rank6]: ^^^^^^^^^^^^^^^^^^^^^^ +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 518, in forward_backward_no_pipelining +[rank6]: output_tensor, num_tokens = forward_step( +[rank6]: ^^^^^^^^^^^^^ +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 289, in forward_step +[rank6]: output_tensor, loss_func = forward_step_func(data_iterator, model) +[rank6]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 446, in forward_step +[rank6]: (tokens, labels, loss_mask, attention_mask, position_ids), token_lens = get_batch(data_iterator) +[rank6]: ^^^^^^^^^^^^^^^^^^^^^^^^ +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 293, in get_batch +[rank6]: batch = get_batch_on_this_cp_rank(batch) +[rank6]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/utils.py", line 1765, in get_batch_on_this_cp_rank +[rank6]: val = val.index_select(seq_dim, index) +[rank6]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank6]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 32.00 GiB. GPU 6 has a total capacity of 139.81 GiB of which 8.10 GiB is free. Including non-PyTorch memory, this process has 131.70 GiB memory in use. Of the allocated memory 130.21 GiB is allocated by PyTorch, and 37.54 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank4]: Traceback (most recent call last): +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 554, in +[rank4]: pretrain( +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 863, in pretrain +[rank4]: iteration, num_floating_point_operations_so_far = train( +[rank4]: ^^^^^^ +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 2229, in train +[rank4]: ) = train_step( +[rank4]: ^^^^^^^^^^^ +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 1382, in train_step +[rank4]: losses_reduced = forward_backward_func( +[rank4]: ^^^^^^^^^^^^^^^^^^^^^^ +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 518, in forward_backward_no_pipelining +[rank4]: output_tensor, num_tokens = forward_step( +[rank4]: ^^^^^^^^^^^^^ +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 289, in forward_step +[rank4]: output_tensor, loss_func = forward_step_func(data_iterator, model) +[rank4]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 446, in forward_step +[rank4]: (tokens, labels, loss_mask, attention_mask, position_ids), token_lens = get_batch(data_iterator) +[rank4]: ^^^^^^^^^^^^^^^^^^^^^^^^ +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 293, in get_batch +[rank4]: batch = get_batch_on_this_cp_rank(batch) +[rank4]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/utils.py", line 1765, in get_batch_on_this_cp_rank +[rank4]: val = val.index_select(seq_dim, index) +[rank4]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank4]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 32.00 GiB. GPU 4 has a total capacity of 139.81 GiB of which 8.10 GiB is free. Including non-PyTorch memory, this process has 131.70 GiB memory in use. Of the allocated memory 130.21 GiB is allocated by PyTorch, and 37.54 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank0]: Traceback (most recent call last): +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 554, in +[rank0]: pretrain( +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 863, in pretrain +[rank0]: iteration, num_floating_point_operations_so_far = train( +[rank0]: ^^^^^^ +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 2229, in train +[rank0]: ) = train_step( +[rank0]: ^^^^^^^^^^^ +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 1382, in train_step +[rank0]: losses_reduced = forward_backward_func( +[rank0]: ^^^^^^^^^^^^^^^^^^^^^^ +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 518, in forward_backward_no_pipelining +[rank0]: output_tensor, num_tokens = forward_step( +[rank0]: ^^^^^^^^^^^^^ +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 289, in forward_step +[rank0]: output_tensor, loss_func = forward_step_func(data_iterator, model) +[rank0]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 446, in forward_step +[rank0]: (tokens, labels, loss_mask, attention_mask, position_ids), token_lens = get_batch(data_iterator) +[rank0]: ^^^^^^^^^^^^^^^^^^^^^^^^ +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 293, in get_batch +[rank0]: batch = get_batch_on_this_cp_rank(batch) +[rank0]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/utils.py", line 1765, in get_batch_on_this_cp_rank +[rank0]: val = val.index_select(seq_dim, index) +[rank0]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank0]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 32.00 GiB. GPU 0 has a total capacity of 139.81 GiB of which 8.10 GiB is free. Including non-PyTorch memory, this process has 131.70 GiB memory in use. Of the allocated memory 130.21 GiB is allocated by PyTorch, and 37.54 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank2]: Traceback (most recent call last): +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 554, in +[rank2]: pretrain( +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 863, in pretrain +[rank2]: iteration, num_floating_point_operations_so_far = train( +[rank2]: ^^^^^^ +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 2229, in train +[rank2]: ) = train_step( +[rank2]: ^^^^^^^^^^^ +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 1382, in train_step +[rank2]: losses_reduced = forward_backward_func( +[rank2]: ^^^^^^^^^^^^^^^^^^^^^^ +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 518, in forward_backward_no_pipelining +[rank2]: output_tensor, num_tokens = forward_step( +[rank2]: ^^^^^^^^^^^^^ +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 289, in forward_step +[rank2]: output_tensor, loss_func = forward_step_func(data_iterator, model) +[rank2]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 446, in forward_step +[rank2]: (tokens, labels, loss_mask, attention_mask, position_ids), token_lens = get_batch(data_iterator) +[rank2]: ^^^^^^^^^^^^^^^^^^^^^^^^ +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 293, in get_batch +[rank2]: batch = get_batch_on_this_cp_rank(batch) +[rank2]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/utils.py", line 1765, in get_batch_on_this_cp_rank +[rank2]: val = val.index_select(seq_dim, index) +[rank2]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank2]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 32.00 GiB. GPU 2 has a total capacity of 139.81 GiB of which 8.10 GiB is free. Including non-PyTorch memory, this process has 131.70 GiB memory in use. Of the allocated memory 130.21 GiB is allocated by PyTorch, and 37.54 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank1]:[W621 22:04:28.347232593 ProcessGroupNCCL.cpp:1476] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank7]:[W621 22:04:28.355827157 ProcessGroupNCCL.cpp:1476] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank5]:[W621 22:04:28.363466299 ProcessGroupNCCL.cpp:1476] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank3]:[W621 22:04:28.363509140 ProcessGroupNCCL.cpp:1476] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +W0621 22:04:31.059000 2391961 site-packages/torch/distributed/elastic/multiprocessing/api.py:900] Sending process 2392033 closing signal SIGTERM +W0621 22:04:31.063000 2391961 site-packages/torch/distributed/elastic/multiprocessing/api.py:900] Sending process 2392034 closing signal SIGTERM +W0621 22:04:31.064000 2391961 site-packages/torch/distributed/elastic/multiprocessing/api.py:900] Sending process 2392035 closing signal SIGTERM +W0621 22:04:31.066000 2391961 site-packages/torch/distributed/elastic/multiprocessing/api.py:900] Sending process 2392036 closing signal SIGTERM +W0621 22:04:31.067000 2391961 site-packages/torch/distributed/elastic/multiprocessing/api.py:900] Sending process 2392037 closing signal SIGTERM +W0621 22:04:31.070000 2391961 site-packages/torch/distributed/elastic/multiprocessing/api.py:900] Sending process 2392038 closing signal SIGTERM +W0621 22:04:31.071000 2391961 site-packages/torch/distributed/elastic/multiprocessing/api.py:900] Sending process 2392039 closing signal SIGTERM +E0621 22:04:32.412000 2391961 site-packages/torch/distributed/elastic/multiprocessing/api.py:874] failed (exitcode: 1) local_rank: 7 (pid: 2392040) of binary: /mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/bin/python3 +Traceback (most recent call last): + File "", line 198, in _run_module_as_main + File "", line 88, in _run_code + File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/launch.py", line 207, in + main() + File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/typing_extensions.py", line 3253, in wrapper + return arg(*args, **kwargs) + ^^^^^^^^^^^^^^^^^^^^ + File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/launch.py", line 203, in main + launch(args) + File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/launch.py", line 188, in launch + run(args) + File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/run.py", line 883, in run + elastic_launch( + File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/launcher/api.py", line 139, in __call__ + return launch_agent(self._config, self._entrypoint, list(args)) + ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ + File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/launcher/api.py", line 270, in launch_agent + raise ChildFailedError( +torch.distributed.elastic.multiprocessing.errors.ChildFailedError: +============================================================ +./pretrain_gpt_profile.py FAILED +------------------------------------------------------------ +Failures: + +------------------------------------------------------------ +Root Cause (first observed failure): +[0]: + time : 2025-06-21_22:04:31 + host : fs-mbz-gpu-791 + rank : 7 (local_rank: 7) + exitcode : 1 (pid: 2392040) + error_file: + traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html +============================================================ ++ set +x ++ for ctx_length in 1024 2048 4096 8192 12288 16384 24576 32768 40960 49152 65536 81920 98304 131072 ++ export PROF_CTX_LENGTH=4096 ++ PROF_CTX_LENGTH=4096 ++ name='/mnt/sharefs/users/hao.zhang/junda/megatron-prof-data--unstable-v5/mytrace.L4096*tp2.cp4.bs32.json' ++ '[' -f '/mnt/sharefs/users/hao.zhang/junda/megatron-prof-data--unstable-v5/mytrace.L4096*tp2.cp4.bs32.json' ']' ++ echo 'Running ctx_length=4096, TP_SIZE=2, CP_SIZE=4, BATCH_SIZE=32' ++ srun bash ./attnserver.sh ++ which python3 ++ python3 -m torch.distributed.launch --nproc_per_node 8 --nnodes 1 --node_rank 0 --rdzv_id 343248 --rdzv_backend c10d --rdzv_endpoint fs-mbz-gpu-791:29500 ./pretrain_gpt_profile.py --tensor-model-parallel-size 2 --context-parallel-size 4 --num-layers 2 --hidden-size 4096 --num-attention-heads 64 --group-query-attention --num-query-groups 16 --seq-length 4096 --max-position-embeddings 4096 --micro-batch-size 1 --global-batch-size 1 --lr 0.0005 --train-iters 10 --lr-decay-iters 150000 --lr-decay-style cosine --lr-warmup-iters 2 --weight-decay .1 --adam-beta2 .999 --fp16 --log-interval 1 --save-interval 16 --eval-interval 16 --eval-iters 1 --vocab-file vocab.json --merge-file merges.txt --save gpt-checkpoint --load gpt-checkpoint --logging-level 0 --mock-data --tensorboard-dir tensorboard-logs/ +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/launch.py:207: FutureWarning: The module torch.distributed.launch is deprecated +and will be removed in future. Use torchrun. +Note that --use-env is set by default in torchrun. +If your script expects `--local-rank` argument to be set, please +change it to read from `os.environ['LOCAL_RANK']` instead. See +https://pytorch.org/docs/stable/distributed.html#launch-utility for +further instructions + + main() +W0621 22:04:35.375000 2393858 site-packages/torch/distributed/run.py:766] +W0621 22:04:35.375000 2393858 site-packages/torch/distributed/run.py:766] ***************************************** +W0621 22:04:35.375000 2393858 site-packages/torch/distributed/run.py:766] 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. +W0621 22:04:35.375000 2393858 site-packages/torch/distributed/run.py:766] ***************************************** +[rank1]:[W621 22:04:57.784459168 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 1] using GPU 1 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank7]:[W621 22:04:57.784473236 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 7] using GPU 7 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank4]:[W621 22:04:57.784512309 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 4] using GPU 4 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank6]:[W621 22:04:57.784512316 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 6] using GPU 6 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank5]:[W621 22:04:57.784547747 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 5] using GPU 5 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank2]:[W621 22:04:57.784578657 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 2] using GPU 2 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank3]:[W621 22:04:57.784618341 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 3] using GPU 3 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank0]:[W621 22:04:57.916239336 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 0] using GPU 0 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +[rank1]: Traceback (most recent call last): +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 554, in +[rank1]: pretrain( +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 863, in pretrain +[rank1]: iteration, num_floating_point_operations_so_far = train( +[rank1]: ^^^^^^ +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 2229, in train +[rank1]: ) = train_step( +[rank1]: ^^^^^^^^^^^ +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 1382, in train_step +[rank1]: losses_reduced = forward_backward_func( +[rank1]: ^^^^^^^^^^^^^^^^^^^^^^ +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 518, in forward_backward_no_pipelining +[rank1]: output_tensor, num_tokens = forward_step( +[rank1]: ^^^^^^^^^^^^^ +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 289, in forward_step +[rank1]: output_tensor, loss_func = forward_step_func(data_iterator, model) +[rank1]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 446, in forward_step +[rank1]: (tokens, labels, loss_mask, attention_mask, position_ids), token_lens = get_batch(data_iterator) +[rank1]: ^^^^^^^^^^^^^^^^^^^^^^^^ +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 284, in get_batch +[rank1]: batch = next(global_batches) +[rank1]: ^^^^^^^^^^^^^^^^^^^^ +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 226, in setup_batches +[rank1]: attention_mask = torch.ones( +[rank1]: ^^^^^^^^^^^ +[rank1]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 512.00 GiB. GPU 1 has a total capacity of 139.81 GiB of which 135.94 GiB is free. Including non-PyTorch memory, this process has 3.86 GiB memory in use. Of the allocated memory 2.29 GiB is allocated by PyTorch, and 119.54 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank5]: Traceback (most recent call last): +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 554, in +[rank5]: pretrain( +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 863, in pretrain +[rank5]: iteration, num_floating_point_operations_so_far = train( +[rank5]: ^^^^^^ +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 2229, in train +[rank5]: ) = train_step( +[rank5]: ^^^^^^^^^^^ +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 1382, in train_step +[rank5]: losses_reduced = forward_backward_func( +[rank5]: ^^^^^^^^^^^^^^^^^^^^^^ +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 518, in forward_backward_no_pipelining +[rank5]: output_tensor, num_tokens = forward_step( +[rank5]: ^^^^^^^^^^^^^ +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 289, in forward_step +[rank5]: output_tensor, loss_func = forward_step_func(data_iterator, model) +[rank5]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 446, in forward_step +[rank5]: (tokens, labels, loss_mask, attention_mask, position_ids), token_lens = get_batch(data_iterator) +[rank5]: ^^^^^^^^^^^^^^^^^^^^^^^^ +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 284, in get_batch +[rank5]: batch = next(global_batches) +[rank5]: ^^^^^^^^^^^^^^^^^^^^ +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 226, in setup_batches +[rank5]: attention_mask = torch.ones( +[rank5]: ^^^^^^^^^^^ +[rank5]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 512.00 GiB. GPU 5 has a total capacity of 139.81 GiB of which 135.94 GiB is free. Including non-PyTorch memory, this process has 3.86 GiB memory in use. Of the allocated memory 2.29 GiB is allocated by PyTorch, and 119.54 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank7]: Traceback (most recent call last): +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 554, in +[rank7]: pretrain( +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 863, in pretrain +[rank7]: iteration, num_floating_point_operations_so_far = train( +[rank7]: ^^^^^^ +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 2229, in train +[rank7]: ) = train_step( +[rank7]: ^^^^^^^^^^^ +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 1382, in train_step +[rank7]: losses_reduced = forward_backward_func( +[rank7]: ^^^^^^^^^^^^^^^^^^^^^^ +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 518, in forward_backward_no_pipelining +[rank7]: output_tensor, num_tokens = forward_step( +[rank7]: ^^^^^^^^^^^^^ +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 289, in forward_step +[rank7]: output_tensor, loss_func = forward_step_func(data_iterator, model) +[rank7]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 446, in forward_step +[rank7]: (tokens, labels, loss_mask, attention_mask, position_ids), token_lens = get_batch(data_iterator) +[rank7]: ^^^^^^^^^^^^^^^^^^^^^^^^ +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 284, in get_batch +[rank7]: batch = next(global_batches) +[rank7]: ^^^^^^^^^^^^^^^^^^^^ +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 226, in setup_batches +[rank7]: attention_mask = torch.ones( +[rank7]: ^^^^^^^^^^^ +[rank7]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 512.00 GiB. GPU 7 has a total capacity of 139.81 GiB of which 135.94 GiB is free. Including non-PyTorch memory, this process has 3.86 GiB memory in use. Of the allocated memory 2.29 GiB is allocated by PyTorch, and 119.54 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank3]: Traceback (most recent call last): +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 554, in +[rank3]: pretrain( +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 863, in pretrain +[rank3]: iteration, num_floating_point_operations_so_far = train( +[rank3]: ^^^^^^ +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 2229, in train +[rank3]: ) = train_step( +[rank3]: ^^^^^^^^^^^ +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 1382, in train_step +[rank3]: losses_reduced = forward_backward_func( +[rank3]: ^^^^^^^^^^^^^^^^^^^^^^ +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 518, in forward_backward_no_pipelining +[rank3]: output_tensor, num_tokens = forward_step( +[rank3]: ^^^^^^^^^^^^^ +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 289, in forward_step +[rank3]: output_tensor, loss_func = forward_step_func(data_iterator, model) +[rank3]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 446, in forward_step +[rank3]: (tokens, labels, loss_mask, attention_mask, position_ids), token_lens = get_batch(data_iterator) +[rank3]: ^^^^^^^^^^^^^^^^^^^^^^^^ +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 284, in get_batch +[rank3]: batch = next(global_batches) +[rank3]: ^^^^^^^^^^^^^^^^^^^^ +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 226, in setup_batches +[rank3]: attention_mask = torch.ones( +[rank3]: ^^^^^^^^^^^ +[rank3]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 512.00 GiB. GPU 3 has a total capacity of 139.81 GiB of which 135.94 GiB is free. Including non-PyTorch memory, this process has 3.86 GiB memory in use. Of the allocated memory 2.29 GiB is allocated by PyTorch, and 119.54 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank4]: Traceback (most recent call last): +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 554, in +[rank4]: pretrain( +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 863, in pretrain +[rank4]: iteration, num_floating_point_operations_so_far = train( +[rank4]: ^^^^^^ +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 2229, in train +[rank4]: ) = train_step( +[rank4]: ^^^^^^^^^^^ +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 1382, in train_step +[rank4]: losses_reduced = forward_backward_func( +[rank4]: ^^^^^^^^^^^^^^^^^^^^^^ +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 518, in forward_backward_no_pipelining +[rank4]: output_tensor, num_tokens = forward_step( +[rank4]: ^^^^^^^^^^^^^ +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 289, in forward_step +[rank4]: output_tensor, loss_func = forward_step_func(data_iterator, model) +[rank4]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 446, in forward_step +[rank4]: (tokens, labels, loss_mask, attention_mask, position_ids), token_lens = get_batch(data_iterator) +[rank4]: ^^^^^^^^^^^^^^^^^^^^^^^^ +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 284, in get_batch +[rank4]: batch = next(global_batches) +[rank4]: ^^^^^^^^^^^^^^^^^^^^ +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 226, in setup_batches +[rank4]: attention_mask = torch.ones( +[rank4]: ^^^^^^^^^^^ +[rank4]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 512.00 GiB. GPU 4 has a total capacity of 139.81 GiB of which 135.94 GiB is free. Including non-PyTorch memory, this process has 3.86 GiB memory in use. Of the allocated memory 2.29 GiB is allocated by PyTorch, and 119.54 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank6]: Traceback (most recent call last): +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 554, in +[rank6]: pretrain( +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 863, in pretrain +[rank6]: iteration, num_floating_point_operations_so_far = train( +[rank6]: ^^^^^^ +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 2229, in train +[rank6]: ) = train_step( +[rank6]: ^^^^^^^^^^^ +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 1382, in train_step +[rank6]: losses_reduced = forward_backward_func( +[rank6]: ^^^^^^^^^^^^^^^^^^^^^^ +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 518, in forward_backward_no_pipelining +[rank6]: output_tensor, num_tokens = forward_step( +[rank6]: ^^^^^^^^^^^^^ +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 289, in forward_step +[rank6]: output_tensor, loss_func = forward_step_func(data_iterator, model) +[rank6]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 446, in forward_step +[rank6]: (tokens, labels, loss_mask, attention_mask, position_ids), token_lens = get_batch(data_iterator) +[rank6]: ^^^^^^^^^^^^^^^^^^^^^^^^ +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 284, in get_batch +[rank6]: batch = next(global_batches) +[rank6]: ^^^^^^^^^^^^^^^^^^^^ +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 226, in setup_batches +[rank6]: attention_mask = torch.ones( +[rank6]: ^^^^^^^^^^^ +[rank6]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 512.00 GiB. GPU 6 has a total capacity of 139.81 GiB of which 135.94 GiB is free. Including non-PyTorch memory, this process has 3.86 GiB memory in use. Of the allocated memory 2.29 GiB is allocated by PyTorch, and 119.54 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank0]: Traceback (most recent call last): +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 554, in +[rank0]: pretrain( +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 863, in pretrain +[rank0]: iteration, num_floating_point_operations_so_far = train( +[rank0]: ^^^^^^ +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 2229, in train +[rank0]: ) = train_step( +[rank0]: ^^^^^^^^^^^ +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 1382, in train_step +[rank0]: losses_reduced = forward_backward_func( +[rank0]: ^^^^^^^^^^^^^^^^^^^^^^ +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 518, in forward_backward_no_pipelining +[rank0]: output_tensor, num_tokens = forward_step( +[rank0]: ^^^^^^^^^^^^^ +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 289, in forward_step +[rank0]: output_tensor, loss_func = forward_step_func(data_iterator, model) +[rank0]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 446, in forward_step +[rank0]: (tokens, labels, loss_mask, attention_mask, position_ids), token_lens = get_batch(data_iterator) +[rank0]: ^^^^^^^^^^^^^^^^^^^^^^^^ +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 284, in get_batch +[rank0]: batch = next(global_batches) +[rank0]: ^^^^^^^^^^^^^^^^^^^^ +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 226, in setup_batches +[rank0]: attention_mask = torch.ones( +[rank0]: ^^^^^^^^^^^ +[rank0]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 512.00 GiB. GPU 0 has a total capacity of 139.81 GiB of which 135.94 GiB is free. Including non-PyTorch memory, this process has 3.86 GiB memory in use. Of the allocated memory 2.29 GiB is allocated by PyTorch, and 119.54 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank2]: Traceback (most recent call last): +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 554, in +[rank2]: pretrain( +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 863, in pretrain +[rank2]: iteration, num_floating_point_operations_so_far = train( +[rank2]: ^^^^^^ +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 2229, in train +[rank2]: ) = train_step( +[rank2]: ^^^^^^^^^^^ +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 1382, in train_step +[rank2]: losses_reduced = forward_backward_func( +[rank2]: ^^^^^^^^^^^^^^^^^^^^^^ +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 518, in forward_backward_no_pipelining +[rank2]: output_tensor, num_tokens = forward_step( +[rank2]: ^^^^^^^^^^^^^ +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 289, in forward_step +[rank2]: output_tensor, loss_func = forward_step_func(data_iterator, model) +[rank2]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 446, in forward_step +[rank2]: (tokens, labels, loss_mask, attention_mask, position_ids), token_lens = get_batch(data_iterator) +[rank2]: ^^^^^^^^^^^^^^^^^^^^^^^^ +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 284, in get_batch +[rank2]: batch = next(global_batches) +[rank2]: ^^^^^^^^^^^^^^^^^^^^ +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 226, in setup_batches +[rank2]: attention_mask = torch.ones( +[rank2]: ^^^^^^^^^^^ +[rank2]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 512.00 GiB. GPU 2 has a total capacity of 139.81 GiB of which 135.94 GiB is free. Including non-PyTorch memory, this process has 3.86 GiB memory in use. Of the allocated memory 2.29 GiB is allocated by PyTorch, and 119.54 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank1]:[W621 22:05:05.162431269 ProcessGroupNCCL.cpp:1476] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank3]:[W621 22:05:05.272000047 ProcessGroupNCCL.cpp:1476] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank5]:[W621 22:05:05.281982273 ProcessGroupNCCL.cpp:1476] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank7]:[W621 22:05:05.282025681 ProcessGroupNCCL.cpp:1476] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +W0621 22:05:07.417000 2393858 site-packages/torch/distributed/elastic/multiprocessing/api.py:900] Sending process 2393928 closing signal SIGTERM +W0621 22:05:07.419000 2393858 site-packages/torch/distributed/elastic/multiprocessing/api.py:900] Sending process 2393929 closing signal SIGTERM +W0621 22:05:07.420000 2393858 site-packages/torch/distributed/elastic/multiprocessing/api.py:900] Sending process 2393930 closing signal SIGTERM +W0621 22:05:07.422000 2393858 site-packages/torch/distributed/elastic/multiprocessing/api.py:900] Sending process 2393931 closing signal SIGTERM +W0621 22:05:07.422000 2393858 site-packages/torch/distributed/elastic/multiprocessing/api.py:900] Sending process 2393932 closing signal SIGTERM +W0621 22:05:07.424000 2393858 site-packages/torch/distributed/elastic/multiprocessing/api.py:900] Sending process 2393933 closing signal SIGTERM +W0621 22:05:07.425000 2393858 site-packages/torch/distributed/elastic/multiprocessing/api.py:900] Sending process 2393934 closing signal SIGTERM +E0621 22:05:08.407000 2393858 site-packages/torch/distributed/elastic/multiprocessing/api.py:874] failed (exitcode: 1) local_rank: 7 (pid: 2393935) of binary: /mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/bin/python3 +Traceback (most recent call last): + File "", line 198, in _run_module_as_main + File "", line 88, in _run_code + File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/launch.py", line 207, in + main() + File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/typing_extensions.py", line 3253, in wrapper + return arg(*args, **kwargs) + ^^^^^^^^^^^^^^^^^^^^ + File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/launch.py", line 203, in main + launch(args) + File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/launch.py", line 188, in launch + run(args) + File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/run.py", line 883, in run + elastic_launch( + File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/launcher/api.py", line 139, in __call__ + return launch_agent(self._config, self._entrypoint, list(args)) + ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ + File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/launcher/api.py", line 270, in launch_agent + raise ChildFailedError( +torch.distributed.elastic.multiprocessing.errors.ChildFailedError: +============================================================ +./pretrain_gpt_profile.py FAILED +------------------------------------------------------------ +Failures: + +------------------------------------------------------------ +Root Cause (first observed failure): +[0]: + time : 2025-06-21_22:05:07 + host : fs-mbz-gpu-791 + rank : 7 (local_rank: 7) + exitcode : 1 (pid: 2393935) + error_file: + traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html +============================================================ ++ set +x ++ for ctx_length in 1024 2048 4096 8192 12288 16384 24576 32768 40960 49152 65536 81920 98304 131072 ++ export PROF_CTX_LENGTH=8192 ++ PROF_CTX_LENGTH=8192 ++ name='/mnt/sharefs/users/hao.zhang/junda/megatron-prof-data--unstable-v5/mytrace.L8192*tp2.cp4.bs32.json' ++ '[' -f '/mnt/sharefs/users/hao.zhang/junda/megatron-prof-data--unstable-v5/mytrace.L8192*tp2.cp4.bs32.json' ']' ++ echo 'Running ctx_length=8192, TP_SIZE=2, CP_SIZE=4, BATCH_SIZE=32' ++ srun bash ./attnserver.sh ++ which python3 ++ python3 -m torch.distributed.launch --nproc_per_node 8 --nnodes 1 --node_rank 0 --rdzv_id 343248 --rdzv_backend c10d --rdzv_endpoint fs-mbz-gpu-791:29500 ./pretrain_gpt_profile.py --tensor-model-parallel-size 2 --context-parallel-size 4 --num-layers 2 --hidden-size 4096 --num-attention-heads 64 --group-query-attention --num-query-groups 16 --seq-length 8192 --max-position-embeddings 8192 --micro-batch-size 1 --global-batch-size 1 --lr 0.0005 --train-iters 10 --lr-decay-iters 150000 --lr-decay-style cosine --lr-warmup-iters 2 --weight-decay .1 --adam-beta2 .999 --fp16 --log-interval 1 --save-interval 16 --eval-interval 16 --eval-iters 1 --vocab-file vocab.json --merge-file merges.txt --save gpt-checkpoint --load gpt-checkpoint --logging-level 0 --mock-data --tensorboard-dir tensorboard-logs/ +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/launch.py:207: FutureWarning: The module torch.distributed.launch is deprecated +and will be removed in future. Use torchrun. +Note that --use-env is set by default in torchrun. +If your script expects `--local-rank` argument to be set, please +change it to read from `os.environ['LOCAL_RANK']` instead. See +https://pytorch.org/docs/stable/distributed.html#launch-utility for +further instructions + + main() +W0621 22:05:11.376000 2395912 site-packages/torch/distributed/run.py:766] +W0621 22:05:11.376000 2395912 site-packages/torch/distributed/run.py:766] ***************************************** +W0621 22:05:11.376000 2395912 site-packages/torch/distributed/run.py:766] 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. +W0621 22:05:11.376000 2395912 site-packages/torch/distributed/run.py:766] ***************************************** +[rank4]:[W621 22:05:33.843598789 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 4] using GPU 4 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank6]:[W621 22:05:33.843855608 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 6] using GPU 6 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank2]:[W621 22:05:33.844029778 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 2] using GPU 2 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank7]:[W621 22:05:33.849795526 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 7] using GPU 7 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank1]:[W621 22:05:33.849868136 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 1] using GPU 1 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank5]:[W621 22:05:33.850265181 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 5] using GPU 5 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank3]:[W621 22:05:33.852705629 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 3] using GPU 3 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank0]:[W621 22:05:33.977552574 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 0] using GPU 0 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +[rank1]: Traceback (most recent call last): +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 554, in +[rank1]: pretrain( +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 863, in pretrain +[rank1]: iteration, num_floating_point_operations_so_far = train( +[rank1]: ^^^^^^ +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 2229, in train +[rank1]: ) = train_step( +[rank1]: ^^^^^^^^^^^ +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 1382, in train_step +[rank1]: losses_reduced = forward_backward_func( +[rank1]: ^^^^^^^^^^^^^^^^^^^^^^ +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 518, in forward_backward_no_pipelining +[rank1]: output_tensor, num_tokens = forward_step( +[rank1]: ^^^^^^^^^^^^^ +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 289, in forward_step +[rank1]: output_tensor, loss_func = forward_step_func(data_iterator, model) +[rank1]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 446, in forward_step +[rank1]: (tokens, labels, loss_mask, attention_mask, position_ids), token_lens = get_batch(data_iterator) +[rank1]: ^^^^^^^^^^^^^^^^^^^^^^^^ +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 284, in get_batch +[rank1]: batch = next(global_batches) +[rank1]: ^^^^^^^^^^^^^^^^^^^^ +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 226, in setup_batches +[rank1]: attention_mask = torch.ones( +[rank1]: ^^^^^^^^^^^ +[rank1]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 2048.00 GiB. GPU 1 has a total capacity of 139.81 GiB of which 135.73 GiB is free. Including non-PyTorch memory, this process has 4.08 GiB memory in use. Of the allocated memory 2.49 GiB is allocated by PyTorch, and 135.54 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank5]: Traceback (most recent call last): +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 554, in +[rank5]: pretrain( +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 863, in pretrain +[rank5]: iteration, num_floating_point_operations_so_far = train( +[rank5]: ^^^^^^ +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 2229, in train +[rank5]: ) = train_step( +[rank5]: ^^^^^^^^^^^ +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 1382, in train_step +[rank5]: losses_reduced = forward_backward_func( +[rank5]: ^^^^^^^^^^^^^^^^^^^^^^ +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 518, in forward_backward_no_pipelining +[rank5]: output_tensor, num_tokens = forward_step( +[rank5]: ^^^^^^^^^^^^^ +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 289, in forward_step +[rank5]: output_tensor, loss_func = forward_step_func(data_iterator, model) +[rank5]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 446, in forward_step +[rank5]: (tokens, labels, loss_mask, attention_mask, position_ids), token_lens = get_batch(data_iterator) +[rank5]: ^^^^^^^^^^^^^^^^^^^^^^^^ +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 284, in get_batch +[rank5]: batch = next(global_batches) +[rank5]: ^^^^^^^^^^^^^^^^^^^^ +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 226, in setup_batches +[rank5]: attention_mask = torch.ones( +[rank5]: ^^^^^^^^^^^ +[rank5]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 2048.00 GiB. GPU 5 has a total capacity of 139.81 GiB of which 135.73 GiB is free. Including non-PyTorch memory, this process has 4.08 GiB memory in use. Of the allocated memory 2.49 GiB is allocated by PyTorch, and 135.54 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank3]: Traceback (most recent call last): +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 554, in +[rank3]: pretrain( +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 863, in pretrain +[rank3]: iteration, num_floating_point_operations_so_far = train( +[rank3]: ^^^^^^ +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 2229, in train +[rank3]: ) = train_step( +[rank3]: ^^^^^^^^^^^ +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 1382, in train_step +[rank3]: losses_reduced = forward_backward_func( +[rank3]: ^^^^^^^^^^^^^^^^^^^^^^ +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 518, in forward_backward_no_pipelining +[rank3]: output_tensor, num_tokens = forward_step( +[rank3]: ^^^^^^^^^^^^^ +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 289, in forward_step +[rank3]: output_tensor, loss_func = forward_step_func(data_iterator, model) +[rank3]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 446, in forward_step +[rank3]: (tokens, labels, loss_mask, attention_mask, position_ids), token_lens = get_batch(data_iterator) +[rank3]: ^^^^^^^^^^^^^^^^^^^^^^^^ +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 284, in get_batch +[rank3]: batch = next(global_batches) +[rank3]: ^^^^^^^^^^^^^^^^^^^^ +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 226, in setup_batches +[rank3]: attention_mask = torch.ones( +[rank3]: ^^^^^^^^^^^ +[rank3]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 2048.00 GiB. GPU 3 has a total capacity of 139.81 GiB of which 135.73 GiB is free. Including non-PyTorch memory, this process has 4.08 GiB memory in use. Of the allocated memory 2.49 GiB is allocated by PyTorch, and 135.54 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank6]: Traceback (most recent call last): +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 554, in +[rank6]: pretrain( +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 863, in pretrain +[rank6]: iteration, num_floating_point_operations_so_far = train( +[rank6]: ^^^^^^ +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 2229, in train +[rank6]: ) = train_step( +[rank6]: ^^^^^^^^^^^ +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 1382, in train_step +[rank6]: losses_reduced = forward_backward_func( +[rank6]: ^^^^^^^^^^^^^^^^^^^^^^ +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 518, in forward_backward_no_pipelining +[rank6]: output_tensor, num_tokens = forward_step( +[rank6]: ^^^^^^^^^^^^^ +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 289, in forward_step +[rank6]: output_tensor, loss_func = forward_step_func(data_iterator, model) +[rank6]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 446, in forward_step +[rank6]: (tokens, labels, loss_mask, attention_mask, position_ids), token_lens = get_batch(data_iterator) +[rank6]: ^^^^^^^^^^^^^^^^^^^^^^^^ +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 284, in get_batch +[rank6]: batch = next(global_batches) +[rank6]: ^^^^^^^^^^^^^^^^^^^^ +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 226, in setup_batches +[rank6]: attention_mask = torch.ones( +[rank6]: ^^^^^^^^^^^ +[rank6]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 2048.00 GiB. GPU 6 has a total capacity of 139.81 GiB of which 135.73 GiB is free. Including non-PyTorch memory, this process has 4.08 GiB memory in use. Of the allocated memory 2.49 GiB is allocated by PyTorch, and 135.54 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank0]: Traceback (most recent call last): +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 554, in +[rank0]: pretrain( +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 863, in pretrain +[rank0]: iteration, num_floating_point_operations_so_far = train( +[rank0]: ^^^^^^ +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 2229, in train +[rank0]: ) = train_step( +[rank0]: ^^^^^^^^^^^ +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 1382, in train_step +[rank0]: losses_reduced = forward_backward_func( +[rank0]: ^^^^^^^^^^^^^^^^^^^^^^ +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 518, in forward_backward_no_pipelining +[rank0]: output_tensor, num_tokens = forward_step( +[rank0]: ^^^^^^^^^^^^^ +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 289, in forward_step +[rank0]: output_tensor, loss_func = forward_step_func(data_iterator, model) +[rank0]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 446, in forward_step +[rank0]: (tokens, labels, loss_mask, attention_mask, position_ids), token_lens = get_batch(data_iterator) +[rank0]: ^^^^^^^^^^^^^^^^^^^^^^^^ +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 284, in get_batch +[rank0]: batch = next(global_batches) +[rank0]: ^^^^^^^^^^^^^^^^^^^^ +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 226, in setup_batches +[rank0]: attention_mask = torch.ones( +[rank0]: ^^^^^^^^^^^ +[rank0]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 2048.00 GiB. GPU 0 has a total capacity of 139.81 GiB of which 135.73 GiB is free. Including non-PyTorch memory, this process has 4.08 GiB memory in use. Of the allocated memory 2.49 GiB is allocated by PyTorch, and 135.54 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank2]: Traceback (most recent call last): +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 554, in +[rank2]: pretrain( +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 863, in pretrain +[rank2]: iteration, num_floating_point_operations_so_far = train( +[rank2]: ^^^^^^ +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 2229, in train +[rank2]: ) = train_step( +[rank2]: ^^^^^^^^^^^ +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 1382, in train_step +[rank2]: losses_reduced = forward_backward_func( +[rank2]: ^^^^^^^^^^^^^^^^^^^^^^ +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 518, in forward_backward_no_pipelining +[rank2]: output_tensor, num_tokens = forward_step( +[rank2]: ^^^^^^^^^^^^^ +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 289, in forward_step +[rank2]: output_tensor, loss_func = forward_step_func(data_iterator, model) +[rank2]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 446, in forward_step +[rank2]: (tokens, labels, loss_mask, attention_mask, position_ids), token_lens = get_batch(data_iterator) +[rank2]: ^^^^^^^^^^^^^^^^^^^^^^^^ +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 284, in get_batch +[rank2]: batch = next(global_batches) +[rank2]: ^^^^^^^^^^^^^^^^^^^^ +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 226, in setup_batches +[rank2]: attention_mask = torch.ones( +[rank2]: ^^^^^^^^^^^ +[rank2]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 2048.00 GiB. GPU 2 has a total capacity of 139.81 GiB of which 135.73 GiB is free. Including non-PyTorch memory, this process has 4.08 GiB memory in use. Of the allocated memory 2.49 GiB is allocated by PyTorch, and 135.54 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank4]: Traceback (most recent call last): +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 554, in +[rank4]: pretrain( +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 863, in pretrain +[rank4]: iteration, num_floating_point_operations_so_far = train( +[rank4]: ^^^^^^ +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 2229, in train +[rank4]: ) = train_step( +[rank4]: ^^^^^^^^^^^ +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 1382, in train_step +[rank4]: losses_reduced = forward_backward_func( +[rank4]: ^^^^^^^^^^^^^^^^^^^^^^ +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 518, in forward_backward_no_pipelining +[rank4]: output_tensor, num_tokens = forward_step( +[rank4]: ^^^^^^^^^^^^^ +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 289, in forward_step +[rank4]: output_tensor, loss_func = forward_step_func(data_iterator, model) +[rank4]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 446, in forward_step +[rank4]: (tokens, labels, loss_mask, attention_mask, position_ids), token_lens = get_batch(data_iterator) +[rank4]: ^^^^^^^^^^^^^^^^^^^^^^^^ +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 284, in get_batch +[rank4]: batch = next(global_batches) +[rank4]: ^^^^^^^^^^^^^^^^^^^^ +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 226, in setup_batches +[rank4]: attention_mask = torch.ones( +[rank4]: ^^^^^^^^^^^ +[rank4]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 2048.00 GiB. GPU 4 has a total capacity of 139.81 GiB of which 135.73 GiB is free. Including non-PyTorch memory, this process has 4.08 GiB memory in use. Of the allocated memory 2.49 GiB is allocated by PyTorch, and 135.54 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank7]: Traceback (most recent call last): +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 554, in +[rank7]: pretrain( +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 863, in pretrain +[rank7]: iteration, num_floating_point_operations_so_far = train( +[rank7]: ^^^^^^ +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 2229, in train +[rank7]: ) = train_step( +[rank7]: ^^^^^^^^^^^ +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 1382, in train_step +[rank7]: losses_reduced = forward_backward_func( +[rank7]: ^^^^^^^^^^^^^^^^^^^^^^ +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 518, in forward_backward_no_pipelining +[rank7]: output_tensor, num_tokens = forward_step( +[rank7]: ^^^^^^^^^^^^^ +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 289, in forward_step +[rank7]: output_tensor, loss_func = forward_step_func(data_iterator, model) +[rank7]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 446, in forward_step +[rank7]: (tokens, labels, loss_mask, attention_mask, position_ids), token_lens = get_batch(data_iterator) +[rank7]: ^^^^^^^^^^^^^^^^^^^^^^^^ +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 284, in get_batch +[rank7]: batch = next(global_batches) +[rank7]: ^^^^^^^^^^^^^^^^^^^^ +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 226, in setup_batches +[rank7]: attention_mask = torch.ones( +[rank7]: ^^^^^^^^^^^ +[rank7]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 2048.00 GiB. GPU 7 has a total capacity of 139.81 GiB of which 135.73 GiB is free. Including non-PyTorch memory, this process has 4.08 GiB memory in use. Of the allocated memory 2.49 GiB is allocated by PyTorch, and 135.54 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank3]:[W621 22:05:42.462456864 ProcessGroupNCCL.cpp:1476] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank7]:[W621 22:05:42.481447535 ProcessGroupNCCL.cpp:1476] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank1]:[W621 22:05:42.552115679 ProcessGroupNCCL.cpp:1476] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank5]:[W621 22:05:42.662766947 ProcessGroupNCCL.cpp:1476] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +W0621 22:05:43.365000 2395912 site-packages/torch/distributed/elastic/multiprocessing/api.py:900] Sending process 2395982 closing signal SIGTERM +W0621 22:05:43.368000 2395912 site-packages/torch/distributed/elastic/multiprocessing/api.py:900] Sending process 2395983 closing signal SIGTERM +W0621 22:05:43.369000 2395912 site-packages/torch/distributed/elastic/multiprocessing/api.py:900] Sending process 2395984 closing signal SIGTERM +W0621 22:05:43.371000 2395912 site-packages/torch/distributed/elastic/multiprocessing/api.py:900] Sending process 2395986 closing signal SIGTERM +W0621 22:05:43.374000 2395912 site-packages/torch/distributed/elastic/multiprocessing/api.py:900] Sending process 2395987 closing signal SIGTERM +W0621 22:05:43.375000 2395912 site-packages/torch/distributed/elastic/multiprocessing/api.py:900] Sending process 2395988 closing signal SIGTERM +W0621 22:05:43.378000 2395912 site-packages/torch/distributed/elastic/multiprocessing/api.py:900] Sending process 2395989 closing signal SIGTERM +E0621 22:05:44.172000 2395912 site-packages/torch/distributed/elastic/multiprocessing/api.py:874] failed (exitcode: 1) local_rank: 3 (pid: 2395985) of binary: /mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/bin/python3 +Traceback (most recent call last): + File "", line 198, in _run_module_as_main + File "", line 88, in _run_code + File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/launch.py", line 207, in + main() + File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/typing_extensions.py", line 3253, in wrapper + return arg(*args, **kwargs) + ^^^^^^^^^^^^^^^^^^^^ + File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/launch.py", line 203, in main + launch(args) + File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/launch.py", line 188, in launch + run(args) + File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/run.py", line 883, in run + elastic_launch( + File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/launcher/api.py", line 139, in __call__ + return launch_agent(self._config, self._entrypoint, list(args)) + ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ + File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/launcher/api.py", line 270, in launch_agent + raise ChildFailedError( +torch.distributed.elastic.multiprocessing.errors.ChildFailedError: +============================================================ +./pretrain_gpt_profile.py FAILED +------------------------------------------------------------ +Failures: + +------------------------------------------------------------ +Root Cause (first observed failure): +[0]: + time : 2025-06-21_22:05:43 + host : fs-mbz-gpu-791 + rank : 3 (local_rank: 3) + exitcode : 1 (pid: 2395985) + error_file: + traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html +============================================================ ++ set +x ++ for ctx_length in 1024 2048 4096 8192 12288 16384 24576 32768 40960 49152 65536 81920 98304 131072 ++ export PROF_CTX_LENGTH=12288 ++ PROF_CTX_LENGTH=12288 ++ name='/mnt/sharefs/users/hao.zhang/junda/megatron-prof-data--unstable-v5/mytrace.L12288*tp2.cp4.bs32.json' ++ '[' -f '/mnt/sharefs/users/hao.zhang/junda/megatron-prof-data--unstable-v5/mytrace.L12288*tp2.cp4.bs32.json' ']' ++ echo 'Running ctx_length=12288, TP_SIZE=2, CP_SIZE=4, BATCH_SIZE=32' ++ srun bash ./attnserver.sh ++ which python3 ++ python3 -m torch.distributed.launch --nproc_per_node 8 --nnodes 1 --node_rank 0 --rdzv_id 343248 --rdzv_backend c10d --rdzv_endpoint fs-mbz-gpu-791:29500 ./pretrain_gpt_profile.py --tensor-model-parallel-size 2 --context-parallel-size 4 --num-layers 2 --hidden-size 4096 --num-attention-heads 64 --group-query-attention --num-query-groups 16 --seq-length 12288 --max-position-embeddings 12288 --micro-batch-size 1 --global-batch-size 1 --lr 0.0005 --train-iters 10 --lr-decay-iters 150000 --lr-decay-style cosine --lr-warmup-iters 2 --weight-decay .1 --adam-beta2 .999 --fp16 --log-interval 1 --save-interval 16 --eval-interval 16 --eval-iters 1 --vocab-file vocab.json --merge-file merges.txt --save gpt-checkpoint --load gpt-checkpoint --logging-level 0 --mock-data --tensorboard-dir tensorboard-logs/ +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/launch.py:207: FutureWarning: The module torch.distributed.launch is deprecated +and will be removed in future. Use torchrun. +Note that --use-env is set by default in torchrun. +If your script expects `--local-rank` argument to be set, please +change it to read from `os.environ['LOCAL_RANK']` instead. See +https://pytorch.org/docs/stable/distributed.html#launch-utility for +further instructions + + main() +W0621 22:05:47.065000 2397787 site-packages/torch/distributed/run.py:766] +W0621 22:05:47.065000 2397787 site-packages/torch/distributed/run.py:766] ***************************************** +W0621 22:05:47.065000 2397787 site-packages/torch/distributed/run.py:766] 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. +W0621 22:05:47.065000 2397787 site-packages/torch/distributed/run.py:766] ***************************************** +[rank5]:[W621 22:06:08.185640603 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 5] using GPU 5 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank7]:[W621 22:06:08.185655799 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 7] using GPU 7 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank3]:[W621 22:06:08.185674309 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 3] using GPU 3 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank6]:[W621 22:06:08.185700047 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 6] using GPU 6 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank2]:[W621 22:06:08.185740232 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 2] using GPU 2 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank1]:[W621 22:06:08.185765895 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 1] using GPU 1 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank4]:[W621 22:06:08.185851446 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 4] using GPU 4 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank0]:[W621 22:06:08.360874750 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 0] using GPU 0 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +[rank1]: Traceback (most recent call last): +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 554, in +[rank1]: pretrain( +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 863, in pretrain +[rank1]: iteration, num_floating_point_operations_so_far = train( +[rank1]: ^^^^^^ +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 2229, in train +[rank1]: ) = train_step( +[rank1]: ^^^^^^^^^^^ +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 1382, in train_step +[rank1]: losses_reduced = forward_backward_func( +[rank1]: ^^^^^^^^^^^^^^^^^^^^^^ +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 518, in forward_backward_no_pipelining +[rank1]: output_tensor, num_tokens = forward_step( +[rank1]: ^^^^^^^^^^^^^ +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 289, in forward_step +[rank1]: output_tensor, loss_func = forward_step_func(data_iterator, model) +[rank1]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 446, in forward_step +[rank1]: (tokens, labels, loss_mask, attention_mask, position_ids), token_lens = get_batch(data_iterator) +[rank1]: ^^^^^^^^^^^^^^^^^^^^^^^^ +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 284, in get_batch +[rank1]: batch = next(global_batches) +[rank1]: ^^^^^^^^^^^^^^^^^^^^ +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 226, in setup_batches +[rank1]: attention_mask = torch.ones( +[rank1]: ^^^^^^^^^^^ +[rank1]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 4608.00 GiB. GPU 1 has a total capacity of 139.81 GiB of which 135.55 GiB is free. Including non-PyTorch memory, this process has 4.25 GiB memory in use. Of the allocated memory 2.69 GiB is allocated by PyTorch, and 103.54 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank6]: Traceback (most recent call last): +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 554, in +[rank6]: pretrain( +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 863, in pretrain +[rank6]: iteration, num_floating_point_operations_so_far = train( +[rank6]: ^^^^^^ +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 2229, in train +[rank6]: ) = train_step( +[rank6]: ^^^^^^^^^^^ +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 1382, in train_step +[rank6]: losses_reduced = forward_backward_func( +[rank6]: ^^^^^^^^^^^^^^^^^^^^^^ +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 518, in forward_backward_no_pipelining +[rank6]: output_tensor, num_tokens = forward_step( +[rank6]: ^^^^^^^^^^^^^ +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 289, in forward_step +[rank6]: output_tensor, loss_func = forward_step_func(data_iterator, model) +[rank6]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 446, in forward_step +[rank6]: (tokens, labels, loss_mask, attention_mask, position_ids), token_lens = get_batch(data_iterator) +[rank6]: ^^^^^^^^^^^^^^^^^^^^^^^^ +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 284, in get_batch +[rank6]: batch = next(global_batches) +[rank6]: ^^^^^^^^^^^^^^^^^^^^ +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 226, in setup_batches +[rank6]: attention_mask = torch.ones( +[rank6]: ^^^^^^^^^^^ +[rank6]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 4608.00 GiB. GPU 6 has a total capacity of 139.81 GiB of which 135.55 GiB is free. Including non-PyTorch memory, this process has 4.25 GiB memory in use. Of the allocated memory 2.69 GiB is allocated by PyTorch, and 103.54 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank5]: Traceback (most recent call last): +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 554, in +[rank5]: pretrain( +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 863, in pretrain +[rank5]: iteration, num_floating_point_operations_so_far = train( +[rank5]: ^^^^^^ +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 2229, in train +[rank5]: ) = train_step( +[rank5]: ^^^^^^^^^^^ +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 1382, in train_step +[rank5]: losses_reduced = forward_backward_func( +[rank5]: ^^^^^^^^^^^^^^^^^^^^^^ +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 518, in forward_backward_no_pipelining +[rank5]: output_tensor, num_tokens = forward_step( +[rank5]: ^^^^^^^^^^^^^ +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 289, in forward_step +[rank5]: output_tensor, loss_func = forward_step_func(data_iterator, model) +[rank5]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 446, in forward_step +[rank5]: (tokens, labels, loss_mask, attention_mask, position_ids), token_lens = get_batch(data_iterator) +[rank5]: ^^^^^^^^^^^^^^^^^^^^^^^^ +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 284, in get_batch +[rank5]: batch = next(global_batches) +[rank5]: ^^^^^^^^^^^^^^^^^^^^ +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 226, in setup_batches +[rank5]: attention_mask = torch.ones( +[rank5]: ^^^^^^^^^^^ +[rank5]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 4608.00 GiB. GPU 5 has a total capacity of 139.81 GiB of which 135.55 GiB is free. Including non-PyTorch memory, this process has 4.25 GiB memory in use. Of the allocated memory 2.69 GiB is allocated by PyTorch, and 103.54 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank2]: Traceback (most recent call last): +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 554, in +[rank2]: pretrain( +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 863, in pretrain +[rank2]: iteration, num_floating_point_operations_so_far = train( +[rank2]: ^^^^^^ +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 2229, in train +[rank2]: ) = train_step( +[rank2]: ^^^^^^^^^^^ +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 1382, in train_step +[rank2]: losses_reduced = forward_backward_func( +[rank2]: ^^^^^^^^^^^^^^^^^^^^^^ +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 518, in forward_backward_no_pipelining +[rank2]: output_tensor, num_tokens = forward_step( +[rank2]: ^^^^^^^^^^^^^ +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 289, in forward_step +[rank2]: output_tensor, loss_func = forward_step_func(data_iterator, model) +[rank2]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 446, in forward_step +[rank2]: (tokens, labels, loss_mask, attention_mask, position_ids), token_lens = get_batch(data_iterator) +[rank2]: ^^^^^^^^^^^^^^^^^^^^^^^^ +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 284, in get_batch +[rank2]: batch = next(global_batches) +[rank2]: ^^^^^^^^^^^^^^^^^^^^ +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 226, in setup_batches +[rank2]: attention_mask = torch.ones( +[rank2]: ^^^^^^^^^^^ +[rank2]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 4608.00 GiB. GPU 2 has a total capacity of 139.81 GiB of which 135.55 GiB is free. Including non-PyTorch memory, this process has 4.25 GiB memory in use. Of the allocated memory 2.69 GiB is allocated by PyTorch, and 103.54 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank7]: Traceback (most recent call last): +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 554, in +[rank7]: pretrain( +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 863, in pretrain +[rank7]: iteration, num_floating_point_operations_so_far = train( +[rank7]: ^^^^^^ +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 2229, in train +[rank7]: ) = train_step( +[rank7]: ^^^^^^^^^^^ +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 1382, in train_step +[rank7]: losses_reduced = forward_backward_func( +[rank7]: ^^^^^^^^^^^^^^^^^^^^^^ +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 518, in forward_backward_no_pipelining +[rank7]: output_tensor, num_tokens = forward_step( +[rank7]: ^^^^^^^^^^^^^ +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 289, in forward_step +[rank7]: output_tensor, loss_func = forward_step_func(data_iterator, model) +[rank7]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 446, in forward_step +[rank7]: (tokens, labels, loss_mask, attention_mask, position_ids), token_lens = get_batch(data_iterator) +[rank7]: ^^^^^^^^^^^^^^^^^^^^^^^^ +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 284, in get_batch +[rank7]: batch = next(global_batches) +[rank7]: ^^^^^^^^^^^^^^^^^^^^ +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 226, in setup_batches +[rank7]: attention_mask = torch.ones( +[rank7]: ^^^^^^^^^^^ +[rank7]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 4608.00 GiB. GPU 7 has a total capacity of 139.81 GiB of which 135.55 GiB is free. Including non-PyTorch memory, this process has 4.25 GiB memory in use. Of the allocated memory 2.69 GiB is allocated by PyTorch, and 103.54 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank3]: Traceback (most recent call last): +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 554, in +[rank3]: pretrain( +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 863, in pretrain +[rank3]: iteration, num_floating_point_operations_so_far = train( +[rank3]: ^^^^^^ +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 2229, in train +[rank3]: ) = train_step( +[rank3]: ^^^^^^^^^^^ +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 1382, in train_step +[rank3]: losses_reduced = forward_backward_func( +[rank3]: ^^^^^^^^^^^^^^^^^^^^^^ +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 518, in forward_backward_no_pipelining +[rank3]: output_tensor, num_tokens = forward_step( +[rank3]: ^^^^^^^^^^^^^ +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 289, in forward_step +[rank3]: output_tensor, loss_func = forward_step_func(data_iterator, model) +[rank3]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 446, in forward_step +[rank3]: (tokens, labels, loss_mask, attention_mask, position_ids), token_lens = get_batch(data_iterator) +[rank3]: ^^^^^^^^^^^^^^^^^^^^^^^^ +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 284, in get_batch +[rank3]: batch = next(global_batches) +[rank3]: ^^^^^^^^^^^^^^^^^^^^ +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 226, in setup_batches +[rank3]: attention_mask = torch.ones( +[rank3]: ^^^^^^^^^^^ +[rank3]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 4608.00 GiB. GPU 3 has a total capacity of 139.81 GiB of which 135.55 GiB is free. Including non-PyTorch memory, this process has 4.25 GiB memory in use. Of the allocated memory 2.69 GiB is allocated by PyTorch, and 103.54 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank4]: Traceback (most recent call last): +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 554, in +[rank4]: pretrain( +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 863, in pretrain +[rank4]: iteration, num_floating_point_operations_so_far = train( +[rank4]: ^^^^^^ +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 2229, in train +[rank4]: ) = train_step( +[rank4]: ^^^^^^^^^^^ +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 1382, in train_step +[rank4]: losses_reduced = forward_backward_func( +[rank4]: ^^^^^^^^^^^^^^^^^^^^^^ +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 518, in forward_backward_no_pipelining +[rank4]: output_tensor, num_tokens = forward_step( +[rank4]: ^^^^^^^^^^^^^ +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 289, in forward_step +[rank4]: output_tensor, loss_func = forward_step_func(data_iterator, model) +[rank4]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 446, in forward_step +[rank4]: (tokens, labels, loss_mask, attention_mask, position_ids), token_lens = get_batch(data_iterator) +[rank4]: ^^^^^^^^^^^^^^^^^^^^^^^^ +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 284, in get_batch +[rank4]: batch = next(global_batches) +[rank4]: ^^^^^^^^^^^^^^^^^^^^ +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 226, in setup_batches +[rank4]: attention_mask = torch.ones( +[rank4]: ^^^^^^^^^^^ +[rank4]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 4608.00 GiB. GPU 4 has a total capacity of 139.81 GiB of which 135.55 GiB is free. Including non-PyTorch memory, this process has 4.25 GiB memory in use. Of the allocated memory 2.69 GiB is allocated by PyTorch, and 103.54 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank0]: Traceback (most recent call last): +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 554, in +[rank0]: pretrain( +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 863, in pretrain +[rank0]: iteration, num_floating_point_operations_so_far = train( +[rank0]: ^^^^^^ +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 2229, in train +[rank0]: ) = train_step( +[rank0]: ^^^^^^^^^^^ +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 1382, in train_step +[rank0]: losses_reduced = forward_backward_func( +[rank0]: ^^^^^^^^^^^^^^^^^^^^^^ +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 518, in forward_backward_no_pipelining +[rank0]: output_tensor, num_tokens = forward_step( +[rank0]: ^^^^^^^^^^^^^ +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 289, in forward_step +[rank0]: output_tensor, loss_func = forward_step_func(data_iterator, model) +[rank0]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 446, in forward_step +[rank0]: (tokens, labels, loss_mask, attention_mask, position_ids), token_lens = get_batch(data_iterator) +[rank0]: ^^^^^^^^^^^^^^^^^^^^^^^^ +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 284, in get_batch +[rank0]: batch = next(global_batches) +[rank0]: ^^^^^^^^^^^^^^^^^^^^ +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 226, in setup_batches +[rank0]: attention_mask = torch.ones( +[rank0]: ^^^^^^^^^^^ +[rank0]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 4608.00 GiB. GPU 0 has a total capacity of 139.81 GiB of which 135.55 GiB is free. Including non-PyTorch memory, this process has 4.25 GiB memory in use. Of the allocated memory 2.69 GiB is allocated by PyTorch, and 103.54 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank3]:[W621 22:06:17.367431724 ProcessGroupNCCL.cpp:1476] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank1]:[W621 22:06:17.369708610 ProcessGroupNCCL.cpp:1476] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank5]:[W621 22:06:18.478455353 ProcessGroupNCCL.cpp:1476] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank7]:[W621 22:06:18.549108533 ProcessGroupNCCL.cpp:1476] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +W0621 22:06:19.913000 2397787 site-packages/torch/distributed/elastic/multiprocessing/api.py:900] Sending process 2397858 closing signal SIGTERM +W0621 22:06:19.916000 2397787 site-packages/torch/distributed/elastic/multiprocessing/api.py:900] Sending process 2397859 closing signal SIGTERM +W0621 22:06:19.917000 2397787 site-packages/torch/distributed/elastic/multiprocessing/api.py:900] Sending process 2397860 closing signal SIGTERM +W0621 22:06:19.922000 2397787 site-packages/torch/distributed/elastic/multiprocessing/api.py:900] Sending process 2397862 closing signal SIGTERM +W0621 22:06:19.926000 2397787 site-packages/torch/distributed/elastic/multiprocessing/api.py:900] Sending process 2397863 closing signal SIGTERM +W0621 22:06:19.926000 2397787 site-packages/torch/distributed/elastic/multiprocessing/api.py:900] Sending process 2397864 closing signal SIGTERM +W0621 22:06:19.939000 2397787 site-packages/torch/distributed/elastic/multiprocessing/api.py:900] Sending process 2397865 closing signal SIGTERM +E0621 22:06:21.722000 2397787 site-packages/torch/distributed/elastic/multiprocessing/api.py:874] failed (exitcode: 1) local_rank: 3 (pid: 2397861) of binary: /mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/bin/python3 +Traceback (most recent call last): + File "", line 198, in _run_module_as_main + File "", line 88, in _run_code + File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/launch.py", line 207, in + main() + File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/typing_extensions.py", line 3253, in wrapper + return arg(*args, **kwargs) + ^^^^^^^^^^^^^^^^^^^^ + File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/launch.py", line 203, in main + launch(args) + File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/launch.py", line 188, in launch + run(args) + File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/run.py", line 883, in run + elastic_launch( + File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/launcher/api.py", line 139, in __call__ + return launch_agent(self._config, self._entrypoint, list(args)) + ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ + File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/launcher/api.py", line 270, in launch_agent + raise ChildFailedError( +torch.distributed.elastic.multiprocessing.errors.ChildFailedError: +============================================================ +./pretrain_gpt_profile.py FAILED +------------------------------------------------------------ +Failures: + +------------------------------------------------------------ +Root Cause (first observed failure): +[0]: + time : 2025-06-21_22:06:19 + host : fs-mbz-gpu-791 + rank : 3 (local_rank: 3) + exitcode : 1 (pid: 2397861) + error_file: + traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html +============================================================ ++ set +x ++ for ctx_length in 1024 2048 4096 8192 12288 16384 24576 32768 40960 49152 65536 81920 98304 131072 ++ export PROF_CTX_LENGTH=16384 ++ PROF_CTX_LENGTH=16384 ++ name='/mnt/sharefs/users/hao.zhang/junda/megatron-prof-data--unstable-v5/mytrace.L16384*tp2.cp4.bs32.json' ++ '[' -f '/mnt/sharefs/users/hao.zhang/junda/megatron-prof-data--unstable-v5/mytrace.L16384*tp2.cp4.bs32.json' ']' ++ echo 'Running ctx_length=16384, TP_SIZE=2, CP_SIZE=4, BATCH_SIZE=32' ++ srun bash ./attnserver.sh ++ which python3 ++ python3 -m torch.distributed.launch --nproc_per_node 8 --nnodes 1 --node_rank 0 --rdzv_id 343248 --rdzv_backend c10d --rdzv_endpoint fs-mbz-gpu-791:29500 ./pretrain_gpt_profile.py --tensor-model-parallel-size 2 --context-parallel-size 4 --num-layers 2 --hidden-size 4096 --num-attention-heads 64 --group-query-attention --num-query-groups 16 --seq-length 16384 --max-position-embeddings 16384 --micro-batch-size 1 --global-batch-size 1 --lr 0.0005 --train-iters 10 --lr-decay-iters 150000 --lr-decay-style cosine --lr-warmup-iters 2 --weight-decay .1 --adam-beta2 .999 --fp16 --log-interval 1 --save-interval 16 --eval-interval 16 --eval-iters 1 --vocab-file vocab.json --merge-file merges.txt --save gpt-checkpoint --load gpt-checkpoint --logging-level 0 --mock-data --tensorboard-dir tensorboard-logs/ +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/launch.py:207: FutureWarning: The module torch.distributed.launch is deprecated +and will be removed in future. Use torchrun. +Note that --use-env is set by default in torchrun. +If your script expects `--local-rank` argument to be set, please +change it to read from `os.environ['LOCAL_RANK']` instead. See +https://pytorch.org/docs/stable/distributed.html#launch-utility for +further instructions + + main() +W0621 22:06:24.539000 2399666 site-packages/torch/distributed/run.py:766] +W0621 22:06:24.539000 2399666 site-packages/torch/distributed/run.py:766] ***************************************** +W0621 22:06:24.539000 2399666 site-packages/torch/distributed/run.py:766] 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. +W0621 22:06:24.539000 2399666 site-packages/torch/distributed/run.py:766] ***************************************** +[rank2]:[W621 22:06:46.548048371 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 2] using GPU 2 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank4]:[W621 22:06:46.548096911 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 4] using GPU 4 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank6]:[W621 22:06:46.548771728 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 6] using GPU 6 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank7]:[W621 22:06:46.558668751 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 7] using GPU 7 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank5]:[W621 22:06:46.559044277 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 5] using GPU 5 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank3]:[W621 22:06:46.559182545 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 3] using GPU 3 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank1]:[W621 22:06:46.559685460 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 1] using GPU 1 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank0]:[W621 22:06:46.685116528 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 0] using GPU 0 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/models/gpt/gpt_layer_specs.py:94: UserWarning: The fp8 argument in "get_gpt_layer_with_transformer_engine_spec" has been deprecated and will be removed soon. Please update your code accordingly. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/transformer_engine/pytorch/cpu_offload.py:595: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +[rank5]: Traceback (most recent call last): +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 554, in +[rank5]: pretrain( +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 863, in pretrain +[rank5]: iteration, num_floating_point_operations_so_far = train( +[rank5]: ^^^^^^ +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 2229, in train +[rank5]: ) = train_step( +[rank5]: ^^^^^^^^^^^ +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 1382, in train_step +[rank5]: losses_reduced = forward_backward_func( +[rank5]: ^^^^^^^^^^^^^^^^^^^^^^ +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 518, in forward_backward_no_pipelining +[rank5]: output_tensor, num_tokens = forward_step( +[rank5]: ^^^^^^^^^^^^^ +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 289, in forward_step +[rank5]: output_tensor, loss_func = forward_step_func(data_iterator, model) +[rank5]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 446, in forward_step +[rank5]: (tokens, labels, loss_mask, attention_mask, position_ids), token_lens = get_batch(data_iterator) +[rank5]: ^^^^^^^^^^^^^^^^^^^^^^^^ +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 284, in get_batch +[rank5]: batch = next(global_batches) +[rank5]: ^^^^^^^^^^^^^^^^^^^^ +[rank5]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 226, in setup_batches +[rank5]: attention_mask = torch.ones( +[rank5]: ^^^^^^^^^^^ +[rank5]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 8192.00 GiB. GPU 5 has a total capacity of 139.81 GiB of which 135.36 GiB is free. Including non-PyTorch memory, this process has 4.45 GiB memory in use. Of the allocated memory 2.90 GiB is allocated by PyTorch, and 97.54 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank3]: Traceback (most recent call last): +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 554, in +[rank3]: pretrain( +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 863, in pretrain +[rank3]: iteration, num_floating_point_operations_so_far = train( +[rank3]: ^^^^^^ +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 2229, in train +[rank3]: ) = train_step( +[rank3]: ^^^^^^^^^^^ +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 1382, in train_step +[rank3]: losses_reduced = forward_backward_func( +[rank3]: ^^^^^^^^^^^^^^^^^^^^^^ +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 518, in forward_backward_no_pipelining +[rank3]: output_tensor, num_tokens = forward_step( +[rank3]: ^^^^^^^^^^^^^ +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 289, in forward_step +[rank3]: output_tensor, loss_func = forward_step_func(data_iterator, model) +[rank3]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 446, in forward_step +[rank3]: (tokens, labels, loss_mask, attention_mask, position_ids), token_lens = get_batch(data_iterator) +[rank3]: ^^^^^^^^^^^^^^^^^^^^^^^^ +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 284, in get_batch +[rank3]: batch = next(global_batches) +[rank3]: ^^^^^^^^^^^^^^^^^^^^ +[rank3]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 226, in setup_batches +[rank3]: attention_mask = torch.ones( +[rank3]: ^^^^^^^^^^^ +[rank3]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 8192.00 GiB. GPU 3 has a total capacity of 139.81 GiB of which 135.36 GiB is free. Including non-PyTorch memory, this process has 4.45 GiB memory in use. Of the allocated memory 2.90 GiB is allocated by PyTorch, and 97.54 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank0]: Traceback (most recent call last): +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 554, in +[rank0]: pretrain( +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 863, in pretrain +[rank0]: iteration, num_floating_point_operations_so_far = train( +[rank0]: ^^^^^^ +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 2229, in train +[rank0]: ) = train_step( +[rank0]: ^^^^^^^^^^^ +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 1382, in train_step +[rank0]: losses_reduced = forward_backward_func( +[rank0]: ^^^^^^^^^^^^^^^^^^^^^^ +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 518, in forward_backward_no_pipelining +[rank0]: output_tensor, num_tokens = forward_step( +[rank0]: ^^^^^^^^^^^^^ +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 289, in forward_step +[rank0]: output_tensor, loss_func = forward_step_func(data_iterator, model) +[rank0]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 446, in forward_step +[rank0]: (tokens, labels, loss_mask, attention_mask, position_ids), token_lens = get_batch(data_iterator) +[rank0]: ^^^^^^^^^^^^^^^^^^^^^^^^ +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 284, in get_batch +[rank0]: batch = next(global_batches) +[rank0]: ^^^^^^^^^^^^^^^^^^^^ +[rank0]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 226, in setup_batches +[rank0]: attention_mask = torch.ones( +[rank0]: ^^^^^^^^^^^ +[rank0]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 8192.00 GiB. GPU 0 has a total capacity of 139.81 GiB of which 135.36 GiB is free. Including non-PyTorch memory, this process has 4.45 GiB memory in use. Of the allocated memory 2.90 GiB is allocated by PyTorch, and 97.54 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank6]: Traceback (most recent call last): +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 554, in +[rank6]: pretrain( +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 863, in pretrain +[rank6]: iteration, num_floating_point_operations_so_far = train( +[rank6]: ^^^^^^ +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 2229, in train +[rank6]: ) = train_step( +[rank6]: ^^^^^^^^^^^ +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 1382, in train_step +[rank6]: losses_reduced = forward_backward_func( +[rank6]: ^^^^^^^^^^^^^^^^^^^^^^ +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 518, in forward_backward_no_pipelining +[rank6]: output_tensor, num_tokens = forward_step( +[rank6]: ^^^^^^^^^^^^^ +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 289, in forward_step +[rank6]: output_tensor, loss_func = forward_step_func(data_iterator, model) +[rank6]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 446, in forward_step +[rank6]: (tokens, labels, loss_mask, attention_mask, position_ids), token_lens = get_batch(data_iterator) +[rank6]: ^^^^^^^^^^^^^^^^^^^^^^^^ +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 284, in get_batch +[rank6]: batch = next(global_batches) +[rank6]: ^^^^^^^^^^^^^^^^^^^^ +[rank6]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 226, in setup_batches +[rank6]: attention_mask = torch.ones( +[rank6]: ^^^^^^^^^^^ +[rank6]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 8192.00 GiB. GPU 6 has a total capacity of 139.81 GiB of which 135.36 GiB is free. Including non-PyTorch memory, this process has 4.45 GiB memory in use. Of the allocated memory 2.90 GiB is allocated by PyTorch, and 97.54 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank1]: Traceback (most recent call last): +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 554, in +[rank1]: pretrain( +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 863, in pretrain +[rank1]: iteration, num_floating_point_operations_so_far = train( +[rank1]: ^^^^^^ +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 2229, in train +[rank1]: ) = train_step( +[rank1]: ^^^^^^^^^^^ +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 1382, in train_step +[rank1]: losses_reduced = forward_backward_func( +[rank1]: ^^^^^^^^^^^^^^^^^^^^^^ +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 518, in forward_backward_no_pipelining +[rank1]: output_tensor, num_tokens = forward_step( +[rank1]: ^^^^^^^^^^^^^ +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 289, in forward_step +[rank1]: output_tensor, loss_func = forward_step_func(data_iterator, model) +[rank1]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 446, in forward_step +[rank1]: (tokens, labels, loss_mask, attention_mask, position_ids), token_lens = get_batch(data_iterator) +[rank1]: ^^^^^^^^^^^^^^^^^^^^^^^^ +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 284, in get_batch +[rank1]: batch = next(global_batches) +[rank1]: ^^^^^^^^^^^^^^^^^^^^ +[rank1]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 226, in setup_batches +[rank1]: attention_mask = torch.ones( +[rank1]: ^^^^^^^^^^^ +[rank1]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 8192.00 GiB. GPU 1 has a total capacity of 139.81 GiB of which 135.36 GiB is free. Including non-PyTorch memory, this process has 4.45 GiB memory in use. Of the allocated memory 2.90 GiB is allocated by PyTorch, and 97.54 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank4]: Traceback (most recent call last): +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 554, in +[rank4]: pretrain( +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 863, in pretrain +[rank4]: iteration, num_floating_point_operations_so_far = train( +[rank4]: ^^^^^^ +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 2229, in train +[rank4]: ) = train_step( +[rank4]: ^^^^^^^^^^^ +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 1382, in train_step +[rank4]: losses_reduced = forward_backward_func( +[rank4]: ^^^^^^^^^^^^^^^^^^^^^^ +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 518, in forward_backward_no_pipelining +[rank4]: output_tensor, num_tokens = forward_step( +[rank4]: ^^^^^^^^^^^^^ +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 289, in forward_step +[rank4]: output_tensor, loss_func = forward_step_func(data_iterator, model) +[rank4]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 446, in forward_step +[rank4]: (tokens, labels, loss_mask, attention_mask, position_ids), token_lens = get_batch(data_iterator) +[rank4]: ^^^^^^^^^^^^^^^^^^^^^^^^ +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 284, in get_batch +[rank4]: batch = next(global_batches) +[rank4]: ^^^^^^^^^^^^^^^^^^^^ +[rank4]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 226, in setup_batches +[rank4]: attention_mask = torch.ones( +[rank4]: ^^^^^^^^^^^ +[rank4]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 8192.00 GiB. GPU 4 has a total capacity of 139.81 GiB of which 135.36 GiB is free. Including non-PyTorch memory, this process has 4.45 GiB memory in use. Of the allocated memory 2.90 GiB is allocated by PyTorch, and 97.54 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank7]: Traceback (most recent call last): +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 554, in +[rank7]: pretrain( +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 863, in pretrain +[rank7]: iteration, num_floating_point_operations_so_far = train( +[rank7]: ^^^^^^ +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 2229, in train +[rank7]: ) = train_step( +[rank7]: ^^^^^^^^^^^ +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 1382, in train_step +[rank7]: losses_reduced = forward_backward_func( +[rank7]: ^^^^^^^^^^^^^^^^^^^^^^ +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 518, in forward_backward_no_pipelining +[rank7]: output_tensor, num_tokens = forward_step( +[rank7]: ^^^^^^^^^^^^^ +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 289, in forward_step +[rank7]: output_tensor, loss_func = forward_step_func(data_iterator, model) +[rank7]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 446, in forward_step +[rank7]: (tokens, labels, loss_mask, attention_mask, position_ids), token_lens = get_batch(data_iterator) +[rank7]: ^^^^^^^^^^^^^^^^^^^^^^^^ +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 284, in get_batch +[rank7]: batch = next(global_batches) +[rank7]: ^^^^^^^^^^^^^^^^^^^^ +[rank7]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 226, in setup_batches +[rank7]: attention_mask = torch.ones( +[rank7]: ^^^^^^^^^^^ +[rank7]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 8192.00 GiB. GPU 7 has a total capacity of 139.81 GiB of which 135.36 GiB is free. Including non-PyTorch memory, this process has 4.45 GiB memory in use. Of the allocated memory 2.90 GiB is allocated by PyTorch, and 97.54 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank2]: Traceback (most recent call last): +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 554, in +[rank2]: pretrain( +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 863, in pretrain +[rank2]: iteration, num_floating_point_operations_so_far = train( +[rank2]: ^^^^^^ +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 2229, in train +[rank2]: ) = train_step( +[rank2]: ^^^^^^^^^^^ +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/training/training.py", line 1382, in train_step +[rank2]: losses_reduced = forward_backward_func( +[rank2]: ^^^^^^^^^^^^^^^^^^^^^^ +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 518, in forward_backward_no_pipelining +[rank2]: output_tensor, num_tokens = forward_step( +[rank2]: ^^^^^^^^^^^^^ +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/megatron/core/pipeline_parallel/schedules.py", line 289, in forward_step +[rank2]: output_tensor, loss_func = forward_step_func(data_iterator, model) +[rank2]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 446, in forward_step +[rank2]: (tokens, labels, loss_mask, attention_mask, position_ids), token_lens = get_batch(data_iterator) +[rank2]: ^^^^^^^^^^^^^^^^^^^^^^^^ +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 284, in get_batch +[rank2]: batch = next(global_batches) +[rank2]: ^^^^^^^^^^^^^^^^^^^^ +[rank2]: File "/mnt/weka/home/hao.zhang/junda/attnserver-megatron/./pretrain_gpt_profile.py", line 226, in setup_batches +[rank2]: attention_mask = torch.ones( +[rank2]: ^^^^^^^^^^^ +[rank2]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 8192.00 GiB. GPU 2 has a total capacity of 139.81 GiB of which 135.36 GiB is free. Including non-PyTorch memory, this process has 4.45 GiB memory in use. Of the allocated memory 2.90 GiB is allocated by PyTorch, and 97.54 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) +[rank3]:[W621 22:06:55.121006921 ProcessGroupNCCL.cpp:1476] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank7]:[W621 22:06:55.151434692 ProcessGroupNCCL.cpp:1476] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank1]:[W621 22:06:55.152193095 ProcessGroupNCCL.cpp:1476] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank5]:[W621 22:06:55.212093457 ProcessGroupNCCL.cpp:1476] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +W0621 22:06:57.981000 2399666 site-packages/torch/distributed/elastic/multiprocessing/api.py:900] Sending process 2399735 closing signal SIGTERM +W0621 22:06:57.983000 2399666 site-packages/torch/distributed/elastic/multiprocessing/api.py:900] Sending process 2399736 closing signal SIGTERM +W0621 22:06:57.983000 2399666 site-packages/torch/distributed/elastic/multiprocessing/api.py:900] Sending process 2399737 closing signal SIGTERM +W0621 22:06:57.986000 2399666 site-packages/torch/distributed/elastic/multiprocessing/api.py:900] Sending process 2399738 closing signal SIGTERM +W0621 22:06:57.986000 2399666 site-packages/torch/distributed/elastic/multiprocessing/api.py:900] Sending process 2399739 closing signal SIGTERM +W0621 22:06:57.992000 2399666 site-packages/torch/distributed/elastic/multiprocessing/api.py:900] Sending process 2399741 closing signal SIGTERM +W0621 22:06:58.000000 2399666 site-packages/torch/distributed/elastic/multiprocessing/api.py:900] Sending process 2399742 closing signal SIGTERM +E0621 22:06:58.265000 2399666 site-packages/torch/distributed/elastic/multiprocessing/api.py:874] failed (exitcode: 1) local_rank: 5 (pid: 2399740) of binary: /mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/bin/python3 +Traceback (most recent call last): + File "", line 198, in _run_module_as_main + File "", line 88, in _run_code + File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/launch.py", line 207, in + main() + File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/typing_extensions.py", line 3253, in wrapper + return arg(*args, **kwargs) + ^^^^^^^^^^^^^^^^^^^^ + File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/launch.py", line 203, in main + launch(args) + File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/launch.py", line 188, in launch + run(args) + File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/run.py", line 883, in run + elastic_launch( + File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/launcher/api.py", line 139, in __call__ + return launch_agent(self._config, self._entrypoint, list(args)) + ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ + File "/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/launcher/api.py", line 270, in launch_agent + raise ChildFailedError( +torch.distributed.elastic.multiprocessing.errors.ChildFailedError: +============================================================ +./pretrain_gpt_profile.py FAILED +------------------------------------------------------------ +Failures: + +------------------------------------------------------------ +Root Cause (first observed failure): +[0]: + time : 2025-06-21_22:06:57 + host : fs-mbz-gpu-791 + rank : 5 (local_rank: 5) + exitcode : 1 (pid: 2399740) + error_file: + traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html +============================================================ ++ set +x ++ for ctx_length in 1024 2048 4096 8192 12288 16384 24576 32768 40960 49152 65536 81920 98304 131072 ++ export PROF_CTX_LENGTH=24576 ++ PROF_CTX_LENGTH=24576 ++ name='/mnt/sharefs/users/hao.zhang/junda/megatron-prof-data--unstable-v5/mytrace.L24576*tp2.cp4.bs32.json' ++ '[' -f '/mnt/sharefs/users/hao.zhang/junda/megatron-prof-data--unstable-v5/mytrace.L24576*tp2.cp4.bs32.json' ']' ++ echo 'Running ctx_length=24576, TP_SIZE=2, CP_SIZE=4, BATCH_SIZE=32' ++ srun bash ./attnserver.sh ++ which python3 ++ python3 -m torch.distributed.launch --nproc_per_node 8 --nnodes 1 --node_rank 0 --rdzv_id 343248 --rdzv_backend c10d --rdzv_endpoint fs-mbz-gpu-791:29500 ./pretrain_gpt_profile.py --tensor-model-parallel-size 2 --context-parallel-size 4 --num-layers 2 --hidden-size 4096 --num-attention-heads 64 --group-query-attention --num-query-groups 16 --seq-length 24576 --max-position-embeddings 24576 --micro-batch-size 1 --global-batch-size 1 --lr 0.0005 --train-iters 10 --lr-decay-iters 150000 --lr-decay-style cosine --lr-warmup-iters 2 --weight-decay .1 --adam-beta2 .999 --fp16 --log-interval 1 --save-interval 16 --eval-interval 16 --eval-iters 1 --vocab-file vocab.json --merge-file merges.txt --save gpt-checkpoint --load gpt-checkpoint --logging-level 0 --mock-data --tensorboard-dir tensorboard-logs/ +/mnt/weka/home/hao.zhang/conda/miniconda/envs/junda-attnserver/lib/python3.12/site-packages/torch/distributed/launch.py:207: FutureWarning: The module torch.distributed.launch is deprecated +and will be removed in future. Use torchrun. +Note that --use-env is set by default in torchrun. +If your script expects `--local-rank` argument to be set, please +change it to read from `os.environ['LOCAL_RANK']` instead. See +https://pytorch.org/docs/stable/distributed.html#launch-utility for +further instructions + + main() +W0621 22:07:02.709000 2401561 site-packages/torch/distributed/run.py:766] +W0621 22:07:02.709000 2401561 site-packages/torch/distributed/run.py:766] ***************************************** +W0621 22:07:02.709000 2401561 site-packages/torch/distributed/run.py:766] 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. +W0621 22:07:02.709000 2401561 site-packages/torch/distributed/run.py:766] ***************************************** +[rank7]:[W621 22:07:25.570566222 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 7] using GPU 7 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank1]:[W621 22:07:25.570579576 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 1] using GPU 1 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank3]:[W621 22:07:25.570639362 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 3] using GPU 3 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank5]:[W621 22:07:25.570752528 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 5] using GPU 5 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank6]:[W621 22:07:25.571044062 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 6] using GPU 6 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank2]:[W621 22:07:25.571074360 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 2] using GPU 2 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank4]:[W621 22:07:25.571438088 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 4] using GPU 4 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device. +[rank0]:[W621 22:07:25.706691733 ProcessGroupNCCL.cpp:4715] [PG ID 0 PG GUID 0 Rank 0] using GPU 0 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can pecify device_id in init_process_group() to force use of a particular device.