Remove elastic training (#8563)
## Summary
- Remove elastic training: the `elasticity` configuration block, the
batch-size solver, the
`DSElasticAgent` rendezvous agent, and the `--elastic_training` launcher
surface.
- Reject a leftover `elasticity` block with `DeepSpeedConfigError`
instead of silently
training without it.
- Tracking: #8489
### What it did
Elastic training let a job declare a GPU range and a set of acceptable
micro batch sizes:
```json
"elasticity": {
"enabled": true,
"max_train_batch_size": 2000,
"micro_batch_sizes": [2, 4, 6],
"min_gpus": 1, "max_gpus": 10000,
"version": 0.2
}
```
`compute_elastic_config` then searched for a global batch size that
factors correctly at
**every** GPU count in that range, so the same job kept an identical
effective batch size on
8 GPUs or 512, and wrote the result over `train_batch_size`,
`train_micro_batch_size_per_gpu` and `gradient_accumulation_steps`.
Separately,
`deepspeed --elastic_training` replaced the normal worker spawn with
`DSElasticAgent`, a
subclass of PyTorch's `LocalElasticAgent` doing c10d rendezvous.
### Why remove it
The GitHub usage scan in #8489 found **zero** issues outside the
DeepSpeed org for the
`"max_train_batch_size"` + `"micro_batch_sizes"` pair, the lowest score
of any candidate in
that table.
It was also incompatible with most DeepSpeed workloads.
`DeepSpeedEngine` asserted against
model parallelism and `PipelineEngine` asserted against pipeline
parallelism, in both cases
unless GPUs-per-node divided evenly by the elastic model-parallel size.
### Removed
| Path | Lines |
|---|---:|
| `deepspeed/elasticity/` (6 files) | 806 |
| `tests/unit/elasticity/test_elastic.py` | 289 |
| `bin/ds_elastic` | 40 |
| docs section in `config-json.md` | 31 |
Plus the wiring: config parsing and the batch-size override in
`runtime/config.py`, the two
compatibility gates in `runtime/engine.py` and `runtime/pipe/engine.py`,
the
`--elastic_training` / `--min_elastic_nodes` / `--max_elastic_nodes`
surface across
`launch.py`, `runner.py` and `multinode_runner.py`, the `setup.py`
scripts entry, the HPU CI
selector, the `CODEOWNERS` entry, and the pre-commit exclude.
### Explicitly NOT touched
- **`zero_optimization.elastic_checkpoint`** is a different feature.
#8099 already removed it
for ZeRO-3; ZeRO-1/2 keep it. Nothing in `deepspeed/runtime/zero/`
changes here.
- **The autotuning `max_train_batch_size` key** is a homonym of the
elasticity one. This is
why #8489 scores the pair `"max_train_batch_size" "micro_batch_sizes"`
rather than either
key alone. `deepspeed/autotuning/` is untouched.
### Notes for review
**The launcher change is structural, not a deletion.** In `launch.py`
the normal spawn path
lived inside `if not args.enable_elastic_training:` with the agent in
the `else`. Removing the
`else` required de-indenting the normal path back to function level: 49
lines, of which 40 are
non-blank and so appear in the diff. That hunk is the one worth reading
closely; everything
else is a straight removal.
**Three things became orphaned and are removed with it.** `typing.Dict`
in `launch.py` and
`__version__` in `runtime/config.py` were each used only by elastic
code, both flagged by
flake8 as F401. `is_elastic_model_parallel_supported` in the engine lost
its only caller.
**`parse_num_nodes` in `runner.py` was already dead before this PR.** It
existed for the
`MIN:MAX` node format that only elastic training accepted, and nothing
called it.
**One test rename.** `test_moq_config_is_rejected` becomes
`test_removed_config_keys_are_rejected`, since it now covers
`quantize_training`,
`eigenvalue` and `elasticity`. It already parametrizes over `None`,
`{}`, `False` and
`"auto"`, so the new key is covered at all four shapes. No CI selector
referenced the old
name.
## Test plan
- [ ] `tests/unit/runtime/test_ds_config_dict.py` — the rejection test
at all four values,
plus the rest of the file as a regression check on `config.py`
- [ ] `deepspeed --help` and a single-node `deepspeed` launch, covering
the de-indented
`launch.py` path
- [ ] A short ZeRO training step, confirming the batch-size path still
resolves with the
elasticity override gone
- [ ] CI unit / accelerator workflows on this PR
## Diffstat
```
21 files changed, 48 insertions(+), 1475 deletions(-)
```
---------
Signed-off-by: pengdurice <pengduhit@gmail.com>
