Deprecate unused DeepSpeed features #8489

Summary

Tracking issue for retiring unused / unmaintained DeepSpeed features so the runtime, docs, tests, and packaging stay consistent.

Goal: remove code paths that are no longer supported, then sweep the leftover docs, extras, CODEOWNERS, and tests so they cannot be re-enabled by accident.

PRs

There is no PR for the full follow-up sweep described below (MoQ + 1-bit comm + PLD + Apex AMP as one bundle). #8535 covers the 1-bit comm and MoQ/eigenvalue leftovers. Progressive Layer Dropping and Apex AMP still have no PR.

Coverage statistics

Line counts are from git diff --numstat. Code = runtime/library sources; tests = tests/; docs = docs/ plus CODEOWNERS / issue templates / CI where noted.

Removed in #8490

Follow-up sweep (no dedicated PR)

Local working-tree estimate vs master when this was written: +104 / −2,814. Combined vs pre-#8490: +253 / −11,092.

Leftover configs should fail at parse time (DeepSpeedConfigError or ZeRO-3 ValidationError) instead of silently training without the feature.

Features to deprecate

Training / checkpointing

• Nebula checkpoint engine (deepspeed/nebula, NebulaCheckpointEngine) — #8490
• MiCS ZeRO-3 sharding (MiCS_Init, mics_shard_size, mics_hierarchical_params_gather) — #8490
• ZeRO-3 elastic checkpoint (zero_optimization.elastic_checkpoint with stage 3; use Universal Checkpointing). ZeRO-1/2 elastic checkpoints remain. — #8099

Optimizers

• 1-bit Adam (OneBitAdam) — #8490
• 1-bit Lamb (OneBitLamb) — #8490
• 0/1 Adam (ZeroOneAdam) — #8490

Compression / quantization

• Model compression library (deepspeed.compression, compression scheduler, related tests) — #8490
• Mixture-of-Quantization (MoQ) / quantize_training / eigenvalue scheduling — #8535

Mixed precision

• Apex AMP (amp config). Use fp16, bf16, or torch_autocast. No PR.

Training schedule

• Progressive Layer Dropping (progressive_layer_drop). No PR.

1-bit communication leftovers (optimizers are gone; backends remain)

• Compressed allreduce backends (deepspeed/runtime/comm/{nccl,mpi,compressed}.py, deepspeed/runtime/compression/cupy.py) — #8535
• tests/onebit/ and remaining unit tests for those backends — #8535
• setup.py extras 1bit / 1bit_mpi and requirements/requirements-1bit-mpi.txt — #8535

Docs / repo hygiene

• Docs nav Model Compression (removed in Deprecate unused DeepSpeed features #8490)
• Docs nav Mixture-of-Quantization and Progressive Layer Dropping — MoQ nav in #8535; PLD still open
• .github/ISSUE_TEMPLATE/compression_bug_report.md — #8535
• CODEOWNERS /deepspeed/runtime/fp16/onebit/ (removed in Deprecate unused DeepSpeed features #8490)
• CODEOWNERS /deepspeed/runtime/compression/ — #8535
• Config / tutorial leftovers for MoQ, Apex AMP, PLD (bulk Nebula/one-bit optimizer/MiCS docs already removed in Deprecate unused DeepSpeed features #8490)

GitHub usage consensus (2026-09-14)

Three independent classifications (Opus, Codex, Grok) over the same GitHub pull agreed on the next deprecation set below.

Method

• Issues/PRs: unauthenticated /search/issues on 2026-09-14. Primary column is -org:deepspeedai. Relative score vs "zero_optimization" -org:deepspeedai = 1,362.
• Code: authenticated /search/code on 2026-09-14 via gh. Relative score vs "zero_optimization" -org:deepspeedai = 49,280. Code search over-counts vendored DeepSpeed copies (e.g. DeepSpeed-0.9.5/ inside other repos) and generic identifiers. Treat as ordinal rank, not user counts.
• Not consensus / do not add from this scan: inference v1/v2, Windows, ZeRO offload, pipeline, DeepCompile, Hybrid Engine, autotuning, DataStates, ZenFlow.

Consensus candidates

2-of-3 only (not unanimous; do not treat as consensus)

Data Efficiency / Random-LTD, dump_state (debug flag; "dump_state" code search is 164k unrelated hits), disable_allgather, FusedLion.

Other unused / already-deprecated candidates

• Sparse Attention (deepspeed/ops/sparse_attention, extras_require['sparse_attn']) — #8493
• sparse_gradients (config-json already calls this essentially deprecated) — #8504
• Elastic training (elasticity / max_train_batch_size + micro_batch_sizes)
• graph_harvesting
• Legacy curriculum learning (superseded by Data Efficiency)
• LoCo-Zero++ (zero_optimization.zeropp_loco_param) — #8534

Notes

• Checked items landed in #8490, #8099, or #8493.
• The original follow-up sweep is not a single open PR. #8535 takes 1-bit comm + MoQ leftovers; PLD and Apex AMP still need a PR.
• MiCS removal also dropped the AutoEP ZeRO-3 “no MiCS” compatibility gate and its unit test.
• 1-bit optimizers are gone in Deprecate unused DeepSpeed features #8490; 1-bit communication backends are in Deprecate leftovers of compression library #8535 (coalesced_collectives.py stays for ZeRO-3).
• deepspeed/compression/helper.py remains as a FutureWarning shim for DeepSpeed-Chat after Deprecate unused DeepSpeed features #8490.
• Apex AMP removal is still local / un-PRed; PyTorch torch_autocast is unchanged. Local import apex remains only to detect apex.optimizers.FusedAdam for FP16 optimizer wrapping.
• Stats methodology: Deprecate unused DeepSpeed features #8490 from git diff --numstat 71d316d60^1 71d316d60; follow-up from git diff --numstat origin/master on the working tree at the time; combined vs b726f4edb.