[BUG] max_seq_length paramter missing in DeepSpeedTransformerConfig

[delock]

Describe the bug
In comments of DeepSpeedTransformerConfig, there is a parameter max_seq_length, however it is not in init function

To Reproduce
Steps to reproduce the behavior:

1. https://github.com/microsoft/DeepSpeed/blob/master/deepspeed/ops/transformer/transformer.py
2. Search for DeepSpeedTransformerConfig
3. In comments see "max_seq_length: The sequence-length of the model being trained with DeepSpeed"
4. Goto init, there is no parameter max_seq_length
5. Goto TransformerConfig, there is no parameter max_seq_length either

Expected behavior
max_seq_length should appear in paramter list

[RezaYazdaniAminabadi]

Hi @delock,

Yes, you are right. Initially, we added this parameter to allocate the buffers properly, however, we later remove this as the sequence-length can be a dynamic parameter. So, we allocate the buffers based on the max_sequence_length that we see during the training: https://github.com/microsoft/DeepSpeed/blob/master/csrc/transformer/ds_transformer_cuda.cpp#L713-L720

Do you have any thought on where/why this parameter can be useful? Otherwise, we can remove it.

Thanks,
Reza

[delock]

Thanks! If it is no longer needed better remove from documentation.