Training without Pipeline Parallelism

[kshitijkg]

When training without pipeline parallelism, the sequential wrapper is used: https://github.com/floatingsnake/gpt-neox/blob/magma/megatron/training.py#L461. Code for to_sequential: https://github.com/floatingsnake/gpt-neox/blob/magma/megatron/model/gpt2_model.py#L343

However, all the adapters added are lost when this is done.

This is probably because the model is rebuilt with self.specs which wasnt updated when the adapters are added.

[floatingbigcat]

Hi, I have tested on a small model with pp=1 mp=1, but the output of model looks fine. Did you change this? maybe our code didn't make model sequential now https://github.com/floatingsnake/gpt-neox/blob/73cdd8692be8a2c579444434e60a01450a8c9a3c/megatron/neox_arguments/arguments.py#L992

https://github.com/floatingsnake/gpt-neox/blob/magma/mytests/test_model_build.py https://github.com/floatingsnake/gpt-neox/blob/magma/configs/summit-70m-openclipH.yml#L16-L17

part of output:

    )
    (6): ParallelTransformerLayerPipe(
      (input_layernorm): LayerNorm((512,), eps=1e-05, elementwise_affine=True)
      (attention): AdapterWrapper(
        (adapter): Sequential(
          (0): Linear(in_features=512, out_features=64, bias=True)
          (1): ReLU()
          (2): Linear(in_features=64, out_features=512, bias=True)
        )
        (attn_block): ParallelSelfAttention(
          (query_key_value): ColumnParallelLinear()
          (rotary_emb): RotaryEmbedding()
          (scale_mask_softmax): FusedScaleMaskSoftmax()
          (attention_dropout): Dropout(p=0, inplace=False)
          (dense): RowParallelLinear()
        )
      )
      (post_attention_layernorm): LayerNorm((512,), eps=1e-05, elementwise_affine=True)
      (mlp): AdapterWrapper(
        (adapter): Sequential(
          (0): Linear(in_features=512, out_features=64, bias=True)
          (1): ReLU()
          (2): Linear(in_features=64, out_features=512, bias=True)
        )
        (attn_block): ParallelMLP(
          (dense_h_to_4h): ColumnParallelLinear()
          (dense_4h_to_h): RowParallelLinear()
        )
      )
    )
    (7): ParallelTransformerLayerPipe(
      (input_layernorm): LayerNorm((512,), eps=1e-05, elementwise_affine=True)
      (attention): AdapterWrapper(
        (adapter): Sequential(
          (0): Linear(in_features=512, out_features=64, bias=True)
          (1): ReLU()
          (2): Linear(in_features=64, out_features=512, bias=True)
        )
        (attn_block): ParallelSelfAttention(
          (query_key_value): ColumnParallelLinear()
          (rotary_emb): RotaryEmbedding()
          (scale_mask_softmax): FusedScaleMaskSoftmax()
          (attention_dropout): Dropout(p=0, inplace=False)
          (dense): RowParallelLinear()
        )
      )
      (post_attention_layernorm): LayerNorm((512,), eps=1e-05, elementwise_affine=True)
      (mlp): AdapterWrapper(
        (adapter): Sequential(
          (0): Linear(in_features=512, out_features=64, bias=True)
          (1): ReLU()
          (2): Linear(in_features=64, out_features=512, bias=True)
        )
        (attn_block): ParallelMLP(
          (dense_h_to_4h): ColumnParallelLinear()
          (dense_4h_to_h): RowParallelLinear()
        )
      )
    )
    (9): NormPipe(
      (norm): LayerNorm((512,), eps=1e-05, elementwise_affine=True)
    )
    (10): ParallelLinearPipe(
      (final_linear): ColumnParallelLinear()
    )
  )
)
Current GPU memory usage: 9.01 GB

[floatingbigcat]

As we abandond the sequential wrapper. and mp=1, pp=1 works will without it. We can reopen the issue when it is needed

[kshitijkg]

Yes, I had changed that line to test sequential wrapper. But yeah, solving this is not high priority for now since we are moving away from the sequential wrapper :)