Consolidate PRs for docs and typos (#8689)
Consolidate the reviewed documentation fixes, typo corrections, and
small runtime cleanups from the original PRs below.
**Credit to @simpleqt for the original contributions.** The
cherry-picked commits preserve the original authorship and sign-offs.
This PR combines the valid changes, avoids duplicate edits, and adjusts
changes that no longer match the current code or documentation.
- Fix broken tutorial links, documentation anchors, and Python API
references.
- Correct parameter names, missing argument documentation, duplicated
words, and garbled text in docstrings, comments, and documentation.
- Simplify identical branches in the checkpoint writer and ZenFlow
gradient-copy path, restore the gradient assertion, and remove the
checkpoint writer's unconditional debug print.
- Additionally, repair the remaining ZeRO-3 Offload tutorial link so it
points to the current ZeRO-3/ZeRO-Infinity section.
## Original PRs
| Original PR | Change incorporated |
| --- | --- |
| [#8445](https://github.com/deepspeedai/DeepSpeed/pull/8445) | Restore
the non-None gradient assertion and remove the duplicated ZenFlow
gradient-copy branch. |
| [#8446](https://github.com/deepspeedai/DeepSpeed/pull/8446) | Remove
identical checkpoint-writer branches and the unconditional constructor
debug print. |
| [#8449](https://github.com/deepspeedai/DeepSpeed/pull/8449) | Point
CONTRIBUTING's installation link to `README.md#installation`. |
| [#8450](https://github.com/deepspeedai/DeepSpeed/pull/8450) | Include
only the `max_norm` documentation; omit the obsolete `use_graph` entry.
|
| [#8545](https://github.com/deepspeedai/DeepSpeed/pull/8545) | Both
scheduling-docstring corrections are covered by #8599, so they are
included only once. |
| [#8546](https://github.com/deepspeedai/DeepSpeed/pull/8546) | Correct
the "Add a new workflow" anchor in `TEST_SELECTION.md`. |
| [#8593](https://github.com/deepspeedai/DeepSpeed/pull/8593) | Correct
the two 1Cycle tutorial links to `/tutorials/one-cycle/`. |
| [#8594](https://github.com/deepspeedai/DeepSpeed/pull/8594) | Retain
the two valid URL case corrections in the ZeRO-3 Offload post; handle
its final tutorial link separately as described below. |
| [#8595](https://github.com/deepspeedai/DeepSpeed/pull/8595) | Correct
the Megatron example to use
`deepspeed.checkpointing.get_cuda_rng_tracker`. |
| [#8596](https://github.com/deepspeedai/DeepSpeed/pull/8596) | Correct
`deepspeed.com.log_summary` to `deepspeed.comm.log_summary`. |
| [#8597](https://github.com/deepspeedai/DeepSpeed/pull/8597) | Point
the AutoTP supported-models link to the correct Read the Docs page and
anchor. |
| [#8598](https://github.com/deepspeedai/DeepSpeed/pull/8598) | Align
the `fp_quantize` docstring's parameter names with its signature. |
| [#8599](https://github.com/deepspeedai/DeepSpeed/pull/8599) | Remove
duplicated words and garbled fragments in scheduling docstrings, runtime
comments, and documentation. |
## Integration adjustments
- **#8450:** `use_graph` was removed by
[#8579](https://github.com/deepspeedai/DeepSpeed/pull/8579), so only the
still-applicable `max_norm` documentation is retained.
- **#8594:** Its proposed final destination, `/tutorials/zero-offload/`,
teaches ZeRO stage 2 optimizer offloading. Instead, this PR separately
updates the ZeRO-3 tutorial link to [Training trillion-scale models with
ZeRO-Infinity](https://www.deepspeed.ai/tutorials/zero/#training-trillion-scale-models-with-zero-infinity),
which documents stage 3 and CPU parameter/optimizer offloading.
- **#8545 and #8599:** The overlapping scheduling-docstring fixes are
incorporated through #8599 rather than duplicated.
---------
Signed-off-by: simpleqt <89645338+simpleqt@users.noreply.github.com>
Signed-off-by: Hongwei Chen <hongweichen@microsoft.com>
Co-authored-by: simpleqt <89645338+simpleqt@users.noreply.github.com>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>