job hangs or IndexError when train reward model with PP> 1

[zirui]

Describe the bug

I Attempt to train reward models of different size(3B/6B/30B), and found out that when PP > 1, two type of issues arise 3B/6B:

• TP=4, PP=1: ok
• TP=4, PP=2: the job hangs

30B:

• TP=8, PP=1: ok
• TP=8, PP=2/4/8: IndexError: index xx is out of range

parameters configuration:

python $CODE_DIR/NeMo-Aligner/examples/nlp/gpt/train_reward_model.py \
      trainer.num_nodes=$NNODES \
      trainer.devices=8 \
      trainer.rm.max_epochs=3 \
      trainer.rm.max_steps=-1 \
      ++model.encoder_seq_length=4096 \
      ++model.micro_batch_size=1 \
      ++model.global_batch_size=64 \
      ++model.data.data_impl=jsonl \
      +model.use_flash_attention=True \
      model.data.seq_length=2048 \
      pretrained_checkpoint.restore_from_path=$MODEL_PATH \
      "model.data.data_prefix={train: [${TRAIN_DATA_PATH}], validation: [${VALID_DATA_PATH}], test: [${VALID_DATA_PATH}]}" \
      trainer.rm.save_interval=10 \
      trainer.rm.val_check_interval=10 \
      exp_manager.create_wandb_logger=True \
      exp_manager.wandb_logger_kwargs.project=$WANDB_PROJECT \
      exp_manager.wandb_logger_kwargs.name=$EXP_NAME \
      exp_manager.explicit_log_dir=$EXP_OUTPUT_DIR \
      ++model.tensor_model_parallel_size=8 \
      ++model.pipeline_model_parallel_size=4 \
      ++model.optim.name=distributed_fused_adam \
      ++model.optim.lr=${LR}  \
      ++model.optim.sched.constant_steps=0 \
      model.optim.weight_decay=0.1

Environment details

• docker: nvcr.io/nvidia/nemo:24.05.01

  File "/code/NeMo-Aligner/examples/nlp/gpt/train_reward_model.py", line 153, in main
    rm_trainer.fit()
  File "/code/NeMo-Aligner/nemo_aligner/algorithms/supervised.py", line 210, in fit
    loss, metrics = self.train_single_step(batch)
  File "/code/NeMo-Aligner/nemo_aligner/algorithms/supervised.py", line 145, in train_single_step
    loss_mean, metrics = self.model.get_loss_and_metrics(batch=batch, forward_only=False)
  File "/code/NeMo-Aligner/nemo_aligner/models/nlp/gpt/megatron_gpt_reward_model.py", line 252, in get_loss_and_metrics
    losses_reduced_per_micro_batch = fwd_bwd_function(
  File "/opt/megatron-lm/megatron/core/pipeline_parallel/schedules.py", line 1271, in forward_backward_pipelining_without_interleaving
    output_tensor, num_tokens = forward_step(
  File "/opt/megatron-lm/megatron/core/pipeline_parallel/schedules.py", line 206, in forward_step
    output_tensor, loss_func = forward_step_func(data_iterator, model)
  File "/code/NeMo-Aligner/nemo_aligner/models/nlp/gpt/megatron_gpt_reward_model.py", line 158, in fwd_output_and_loss_func
    output_tensor = model(**forward_args)
  File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py", line 1511, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
  File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py", line 1520, in _call_impl
    return forward_call(*args, **kwargs)
  File "/opt/megatron-lm/megatron/core/transformer/module.py", line 173, in forward
    outputs = self.module(*inputs, **kwargs)
  File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py", line 1511, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
  File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py", line 1520, in _call_impl
    return forward_call(*args, **kwargs)
  File "/code/NeMo-Aligner/nemo_aligner/models/nlp/gpt/gpt_reward_model.py", line 214, in forward
    hidden_states = super().forward(
  File "/opt/megatron-lm/megatron/core/models/gpt/gpt_model.py", line 190, in forward
    hidden_states = self.decoder(
  File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py", line 1511, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
  File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py", line 1520, in _call_impl
    return forward_call(*args, **kwargs)
  File "/opt/megatron-lm/megatron/core/transformer/transformer_block.py", line 371, in forward
    hidden_states = self._checkpointed_forward(
  File "/opt/megatron-lm/megatron/core/transformer/transformer_block.py", line 257, in _checkpointed_forward
    hidden_states, context = checkpoint_handler(
  File "/opt/megatron-lm/megatron/core/transformer/transformer_block.py", line 240, in checkpoint_handler
    return tensor_parallel.checkpoint(
  File "/opt/megatron-lm/megatron/core/tensor_parallel/random.py", line 301, in checkpoint
    return CheckpointFunction.apply(function, distribute_saved_activations, *args)
  File "/usr/local/lib/python3.10/dist-packages/torch/autograd/function.py", line 553, in apply
    return super().apply(*args, **kwargs)  # type: ignore[misc]
  File "/opt/megatron-lm/megatron/core/tensor_parallel/random.py", line 240, in forward
    outputs = run_function(*args)
  File "/opt/megatron-lm/megatron/core/transformer/transformer_block.py", line 211, in custom_forward
    layer = self._get_layer(index)
  File "/opt/megatron-lm/megatron/core/transformer/transformer_block.py", line 188, in _get_layer
    return self.layers[layer_number]
  File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/container.py", line 295, in __getitem__
    return self._modules[self._get_abs_string_index(idx)]
  File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/container.py", line 285, in _get_abs_string_index
    raise IndexError(f'index {idx} is out of range')
IndexError: index 15 is out of range
Training steps:   0%|          | 0/423 [00:03<?, ?it/s]

[odelalleau]

Would you be able to share your model_config.yaml?

[zirui]

Would you be able to share your model_config.yaml?

I use the default config: NeMo-Aligner/examples/nlp/gpt/conf/training_rm.yaml

[odelalleau]

I was wondering about the model_config.yaml found in your .nemo file (you can extract it with tar xvf on your .nemo). I'm curious about your model's architecture.

[zirui]

I was wondering about the model_config.yaml found in your .nemo file (you can extract it with tar xvf on your .nemo). I'm curious about your model's architecture.

Below is the content of model_config.yaml (Actually, it is a llama-like model that has been converted into .nemo from huggingface model.)

mcore_gpt: true
micro_batch_size: 4
global_batch_size: 8
tensor_model_parallel_size: 1
pipeline_model_parallel_size: 1
virtual_pipeline_model_parallel_size: null
encoder_seq_length: 4096
max_position_embeddings: 4096
num_layers: 32
hidden_size: 4096
ffn_hidden_size: 11008
num_attention_heads: 32
init_method_std: 0.02
use_scaled_init_method: true
hidden_dropout: 0.0
attention_dropout: 0.0
ffn_dropout: 0.0
kv_channels: null
apply_query_key_layer_scaling: true
normalization: rmsnorm
layernorm_epsilon: 1.0e-06
do_layer_norm_weight_decay: false
make_vocab_size_divisible_by: 128
pre_process: true
post_process: true
persist_layer_norm: true
bias: false
activation: fast-swiglu
headscale: false
transformer_block_type: pre_ln
openai_gelu: false
normalize_attention_scores: true
position_embedding_type: rope
rotary_percentage: 1.0
attention_type: multihead
share_embeddings_and_output_weights: false
overlap_p2p_comm: false
batch_p2p_comm: true
num_query_groups: 4
tokenizer:
  library: sentencepiece
  type: null
  model: nemo:4f7bc9bb269d4abd9680ac15dcec4b16_tokenizer.model
  vocab_file: null
  merge_file: null
  delimiter: null
  sentencepiece_legacy: false
native_amp_init_scale: 4294967296
native_amp_growth_interval: 1000
hysteresis: 2
fp32_residual_connection: false
fp16_lm_cross_entropy: false
megatron_amp_O2: false
grad_allreduce_chunk_size_mb: 125
grad_div_ar_fusion: true
gradient_accumulation_fusion: false
bias_activation_fusion: false
bias_dropout_add_fusion: false
masked_softmax_fusion: true
get_attention_mask_from_fusion: true
apply_rope_fusion: false
seed: 1234
resume_from_checkpoint: null
use_cpu_initialization: false
onnx_safe: false
apex_transformer_log_level: 30
gradient_as_bucket_view: true
sync_batch_comm: false
activations_checkpoint_granularity: null
activations_checkpoint_method: null
activations_checkpoint_num_layers: null
num_micro_batches_with_partial_activation_checkpoints: null
activations_checkpoint_layers_per_pipeline: null
sequence_parallel: false
transformer_engine: true
fp8: false
fp8_e4m3: false
fp8_hybrid: true
fp8_margin: 0
fp8_interval: 1
fp8_amax_history_len: 1024
fp8_amax_compute_algo: max
reduce_amax: true
use_emha: false
data:
  index_mapping_dir: null
  data_impl: mmap
  splits_string: 900,50,50
  seq_length: 4096
  skip_warmup: true
  num_workers: 2
  dataloader_type: single
  reset_position_ids: false
  reset_attention_mask: false
  eod_mask_loss: false
  validation_drop_last: true
  no_seqlen_plus_one_input_tokens: false
  pad_samples_to_global_batch_size: false
  shuffle_documents: true
nsys_profile:
  enabled: false
  start_step: 10
  end_step: 10
  ranks:
  - 0
  gen_shape: false
optim:
  name: fused_adam
  lr: 0.0002
  weight_decay: 0.01
  betas:
  - 0.9
  - 0.98
  sched:
    name: CosineAnnealing
    warmup_steps: 500
    constant_steps: 50000
    min_lr: 2.0e-05
rotary_base: 5000000.0
precision: 16
target: nemo.collections.nlp.models.language_modeling.megatron_gpt_model.MegatronGPTModel
nemo_version: 1.23.0

[zirui]

Some additional information:

• I converted Llama-2-7b-hf to Llama-2-7b-hf.nemo, and set TP=4 PP=2, the reward model training tasks still get stuck, so the issue should not be related to my model's architecture
• I also tried SFT training(follow the user-guide), the PP(TP=4,PP=2) training works well

[ashors1]

Hi, could you try setting model.encoder_seq_length and model.data.seq_length to the same value? I was able to reproduce the hang, and making those seq_length arguments consistent fixed the problem for me.