Megatron-LM icon indicating copy to clipboard operation
Megatron-LM copied to clipboard
Published 1 month ago • verified publisher icon NVIDIA

[BUG] Error pre-training BERT

Open fabiancpl opened this issue 1 year ago • 2 comments

Hi guys,

I am following the Megatron-LM example to pre-train a BERT model but I'm getting this error:

[rank0]: Traceback (most recent call last):
[rank0]:   File "/root/Megatron-LM/pretrain_bert.py", line 193, in <module>
[rank0]:     pretrain(train_valid_test_datasets_provider, model_provider,
[rank0]:   File "/root/Megatron-LM/megatron/training/training.py", line 274, in pretrain
[rank0]:     iteration, num_floating_point_operations_so_far = train(
[rank0]:                                                       ^^^^^^
[rank0]:   File "/root/Megatron-LM/megatron/training/training.py", line 1027, in train
[rank0]:     train_step(forward_step_func,
[rank0]:   File "/root/Megatron-LM/megatron/training/training.py", line 550, in train_step
[rank0]:     losses_reduced = forward_backward_func(
[rank0]:                      ^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/root/Megatron-LM/megatron/core/pipeline_parallel/schedules.py", line 381, in forward_backward_no_pipelining
[rank0]:     output_tensor, num_tokens = forward_step(
[rank0]:                                 ^^^^^^^^^^^^^
[rank0]:   File "/root/Megatron-LM/megatron/core/pipeline_parallel/schedules.py", line 206, in forward_step
[rank0]:     output_tensor, loss_func = forward_step_func(data_iterator, model)
[rank0]:                                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/root/Megatron-LM/pretrain_bert.py", line 139, in forward_step
[rank0]:     output_tensor = model(tokens, padding_mask,
[rank0]:                     ^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/opt/conda/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1553, in _wrapped_call_impl
[rank0]:     return self._call_impl(*args, **kwargs)
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/opt/conda/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1562, in _call_impl
[rank0]:     return forward_call(*args, **kwargs)
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/root/Megatron-LM/megatron/core/distributed/distributed_data_parallel.py", line 180, in forward
[rank0]:     return self.module(*inputs, **kwargs)
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/opt/conda/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1553, in _wrapped_call_impl
[rank0]:     return self._call_impl(*args, **kwargs)
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/opt/conda/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1562, in _call_impl
[rank0]:     return forward_call(*args, **kwargs)
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/root/Megatron-LM/megatron/legacy/model/module.py", line 190, in forward
[rank0]:     outputs = self.module(*inputs, **kwargs)
[rank0]:               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/opt/conda/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1553, in _wrapped_call_impl
[rank0]:     return self._call_impl(*args, **kwargs)
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/opt/conda/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1562, in _call_impl
[rank0]:     return forward_call(*args, **kwargs)
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/root/Megatron-LM/megatron/core/models/bert/bert_model.py", line 237, in forward
[rank0]:     hidden_states = self.encoder(
[rank0]:                     ^^^^^^^^^^^^^
[rank0]:   File "/opt/conda/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1553, in _wrapped_call_impl
[rank0]:     return self._call_impl(*args, **kwargs)
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/opt/conda/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1562, in _call_impl
[rank0]:     return forward_call(*args, **kwargs)
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/root/Megatron-LM/megatron/core/transformer/transformer_block.py", line 383, in forward
[rank0]:     hidden_states, context = layer(
[rank0]:                              ^^^^^^
[rank0]:   File "/opt/conda/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1553, in _wrapped_call_impl
[rank0]:     return self._call_impl(*args, **kwargs)
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/opt/conda/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1562, in _call_impl
[rank0]:     return forward_call(*args, **kwargs)
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/root/Megatron-LM/megatron/core/transformer/transformer_layer.py", line 178, in forward
[rank0]:     attention_output_with_bias = self.self_attention(
[rank0]:                                  ^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/opt/conda/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1553, in _wrapped_call_impl
[rank0]:     return self._call_impl(*args, **kwargs)
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/opt/conda/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1562, in _call_impl
[rank0]:     return forward_call(*args, **kwargs)
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/root/Megatron-LM/megatron/core/transformer/attention.py", line 315, in forward
[rank0]:     core_attn_out = self.core_attention(
[rank0]:                     ^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/opt/conda/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1553, in _wrapped_call_impl
[rank0]:     return self._call_impl(*args, **kwargs)
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/opt/conda/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1562, in _call_impl
[rank0]:     return forward_call(*args, **kwargs)
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/root/Megatron-LM/megatron/core/transformer/custom_layers/transformer_engine.py", line 514, in forward
[rank0]:     core_attn_out = super().forward(
[rank0]:                     ^^^^^^^^^^^^^^^^
[rank0]:   File "/opt/conda/lib/python3.12/site-packages/transformer_engine/pytorch/attention.py", line 5267, in forward
[rank0]:     return self.fused_attention(
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/opt/conda/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1553, in _wrapped_call_impl
[rank0]:     return self._call_impl(*args, **kwargs)
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/opt/conda/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1562, in _call_impl
[rank0]:     return forward_call(*args, **kwargs)
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/opt/conda/lib/python3.12/site-packages/torch/_dynamo/eval_frame.py", line 600, in _fn
[rank0]:     return fn(*args, **kwargs)
[rank0]:            ^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/opt/conda/lib/python3.12/site-packages/transformer_engine/pytorch/attention.py", line 4157, in forward
[rank0]:     cu_seqlens_q = get_cu_seqlens(attention_mask)
[rank0]:                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/opt/conda/lib/python3.12/site-packages/transformer_engine/pytorch/attention.py", line 247, in get_cu_seqlens
[rank0]:     cu_seqlens = torch.cat((zero, cu_seqlens))
[rank0]:                  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]: RuntimeError: Tensors must have same number of dimensions: got 1 and 2

I'm using transformer_engine==1.8.0+3ec998e with megatron core_v0.7.0. My pre-training script is like this:

#!/bin/bash

export CUDA_DEVICE_MAX_CONNECTIONS=1
export NVTE_FLASH_ATTN=0

CHECKPOINT_PATH="./checkpoints"
VOCAB_FILE="./vocabs/bert-base-cased-vocab.txt"
DATA_PATH="my-bert_text_sentence"

BERT_ARGS="
    --num-layers 24 \
    --hidden-size 1024 \
    --num-attention-heads 16 \
    --seq-length 512 \
    --max-position-embeddings 512 \
    --micro-batch-size 4 \
    --global-batch-size 8 \
    --lr 0.0001 \
    --train-iters 2000000 \
    --lr-decay-iters 990000 \
    --lr-decay-style linear \
    --min-lr 0.00001 \
    --weight-decay 1e-2 \
    --lr-warmup-fraction .01 \
    --clip-grad 1.0 \
    --fp16
    --use-mcore-models
"

DATA_ARGS="
    --data-path $DATA_PATH \
    --vocab-file $VOCAB_FILE \
    --split 949,50,1
"

OUTPUT_ARGS="
    --log-interval 100 \
    --save-interval 10000 \
    --eval-interval 1000 \
    --eval-iters 10
"

torchrun pretrain_bert.py \
    $BERT_ARGS \
    $DATA_ARGS \
    $OUTPUT_ARGS \
    --save $CHECKPOINT_PATH \
    --load $CHECKPOINT_PATH

I'm also interested in using the BERT cased checkpoint instead of pre-training from scratch.

Thanks in advance.

fabiancpl avatar Jul 25 '24 18:07 fabiancpl

same issue here

RobinAlgayres avatar Jul 30 '24 13:07 RobinAlgayres

Marking as stale. No activity in 60 days.

github-actions[bot] avatar Sep 28 '24 18:09 github-actions[bot]

Hi guys,

Just wondering if this issue has been resolved? I still face the same issue on my side with version core_v0.12.1

wiwiamlam avatar Jun 18 '25 15:06 wiwiamlam

This issue is stale because it has been open for 30 days with no activity. Remove stale label or comment or this will be closed in 7 days.

github-actions[bot] avatar Aug 01 '25 02:08 github-actions[bot]