Model Checkpoint docs are incorrectly rendered on deepspeed.readthedocs.io

[akeshet]

The top google search hit for most deepspeed documentation searches is https://deepspeed.readthedocs.io/ or its various sub-pages. I assume this export is somehow maintained or at least enabled by the deepspeed team, hence this bug report.

The page on model checkpointing at https://deepspeed.readthedocs.io/en/latest/model-checkpointing.html#model-checkpointing shows empty sections for "Loading Training Checkpoints" and "Saving Training Checkpoints", which I have found confusing on a few occasions. Whereas the rst file in the repo has an autofunction declaration inside these sections, e.g. https://github.com/microsoft/DeepSpeed/blob/b692cdea479fba8201584054d654f639e925a265/docs/code-docs/source/model-checkpointing.rst

Evidently the autofunction there is not being correctly rendered by doc export.

[SubhamCPP]

Hi @akeshet, thank you for reporting this issue.

I've also noticed several other pages with missing documentation sections. Below, I'm listing the URLs along with the specific missing blocks:

Training API - Link to Documentation

• Forward Propagation
• Backward Propagation
• Optimizer Step
• Gradient Accumulation

ZeRO - Link to Documentation

• For Modifying Partitioned States
  • GPU Memory Management

The missing sections all seem to have the same root cause: the .. autofunction:: deepspeed.DeepSpeedEngine directive is not being correctly rendered by the documentation export process. This may be related to an issue with the Sphinx configuration or the specific way the documentation build process is set up for deepspeed.readthedocs.io.

It might be worthwhile to verify the Sphinx extensions being used and ensure the autodoc functionality is correctly configured for rendering these DeepSpeedEngine methods.

It might be helpful to add a documentation or bug label to this issue for visibility.

[loadams]

@akeshet - yes, we maintain the readthedocs page, so thanks for reporting this here. I'll take a closer look at what is happening here for why the export is failing.

@SubhamCPP - thanks for the suggestions on the tags, I added them and will take a look.