## deepspeedai/DeepSpeed — v0.19.2…v0.19.3

_57 commits._

### Features
- **Z3: Support for activation ckpt with frozen params (#8148)** (a31eb31)
- **[DeepCompile] Add lightweight pass contracts for optimization passes (#8139)** (b048a7d)
- **Add EXAONE 4.5 model support for Inference V2 (#8121)** (d36079f)
- **Add AutoEP + AutoTP parallel folding (#8064)** (104193b)
- **Add Hybrid Engine rollout in DeepSpeed to support On-Policy Distillation (OPSD) Trainer (#8027)** (429e2ad)
- **feat(types): add type hints to public comm API functions (#8118)** (9b106c4)
- **Add input validation for save_checkpoint and warmup LR schedulers (#8126)** (54ebdec)
- **feat(zenflow): run the overlapped CPU optimizer in a native process (#8058)** (1ddbe48)
- **Add optional torchembed RoPE backend to apply_rotary_pos_emb (#8052)** (5ca819b)
- **Feat: zero3 deprecate elastic checkpoint (#8099)** (ac5c49f)
- **Add numerical-correctness test for Muon under ZeRO-1/2 (#8091)** (0ecb987)
- **Support AutoEP with ZeRO-3 zero.Init source modules (#8060)** (02663d6)
- **Add configurable engine log level (#8067)** (a163245)
- **Add type hints to top-level public API functions (#8079)** (aaaed26)
- **Add merge queue DCO workflow (#8086)** (8bda651)

### Fixes
- **Fix repeated gradient scaling across pipeline stages (#8154)** (d99a642)
- **Fix ZeRO-3 hooks for attribute-delegating modules (#8144)** (75f395f)
- **Fix DeepCompile fallback handling for scalar outputs (#8149)** (397efe8)
- **fix(muon): support ZeRO-1/2 reduce scatter (#8141)** (de1846c)
- **fix(exceptions): narrow broad Exception clauses to specific types (#8136)** (9102e8e)
- **Fix DeepSpeedInferenceConfig crash on bool moe backward-compat value (#8145)** (17e19b1)
- **fix(autotp): when using autotp, ignore the consistency of certain data within the tp_group. (#8125)** (53a2ac4)
- **Fix CommsLogger.stop_profiling_comms not disabling global profiling (#8137)** (10db691)
- **Fix ZeRO-3 autocast gather with mixed parameter dtypes (#8113)** (195255e)
- **Fix ZeRO-3: use per-param dtype for quantize-scale buffer in _allgather_params_coalesced (#8128)** (b079c30)
- **Fix Python 3.9 import-time TypeError in AutoEP ep_router (#8119)** (ae1dba5)
- **Fix FlopsProfiler crash when dp_world_size is None under sequence parallelism (#8122)** (6e38d56)
- **fix: add buffer-length check in shm.cpp (#8082)** (7720656)
- **Fix minor comment/docstring typos in runtime and inference modules (#8046)** (0b987c0)
- **Fix DeepCompile profiling memory cleanup (#8106)** (131779c)
- **fix: log eigenvalue monitor values (#8049)** (32d009b)
- **Fix eigenvalue parsing for compression-only quantize configs (#8057)** (ea3d65e)
- **[DeepCompile] fix gather params in dynamo skipped frames for ZeRO3 (#8059)** (1e85ce2)
- **Fix DeepCompile profile metadata backfill (#8094)** (a33655c)
- **fix: AutoTP partition_config uses full hierarchical module path (#8088)** (7aafb32)
- **fix: close file descriptor in deepspeed_io_handle_t::wait() to prevent fd leak (#8075)** (3c3c68f)
- **Fix ZeRO-3: Use per-param dtype for output buffers in _allgather_params_coalesced (#8073)** (b5b3fde)

### Backend
- **Adopt PTF Code of Conduct (#8165)** (0c36f6d)
- **Update Zhipeng Wang affiliation to Google in COMMITTERS.md (#8153)** (4c275f9)
- **Update CI badge (#8150)** (ec8411e)
- **Use synthetic CIFAR data in AlexNet tests (#8143)** (b81c6eb)
- **Reduce blockDim in fake_quantize_kernel for improved SM occupancy (#8115)** (000ca19)
- **Guard WarmupCosineLR against total_num_steps == warmup_num_steps (ZeroDivisionError) (#8142)** (2163a01)
- **ZeRO 1/2: wait on all IPG-bucket producer streams in average_tensor (#8061) (#8080)** (711bd11)
- **Make DCO workflow Probot compatible (#8110)** (46c9e20)
- **doc: Remove suggestion to build extensions in parallel (#7899)** (05a481b)
- **Avoid CUDA context initialization during op compatibility checks at import (#8078)** (5aef6d8)
- **Don't swallow KeyboardInterrupt/SystemExit in NPUOpBuilder (#8096)** (10f76c2)
- **Reject Muon optimizer with reduce_scatter in ZeRO-1/2 (#8090)** (d714251)
- **Update publication list in README.md (#8095)** (f765822)
- **Enable bf16 check_grad_overflow by default (matching fp16) (#8035)** (f0253c8)
- **[CI] diff driven test selection (#8077)** (f346082)
- **activation_checkpointing: default num_layers to None so configure() assert fires (#8041)** (3a47d1d)
- **Default gradient_clipping to 1.0 (#8068)** (4421665)
- **Stop obsolete CI jobs on workflow cancellation (#8034)** (329f937)
- **Validate fp16 dynamic loss scaling parameters are positive (#8050)** (ded2349)
- **Update version post release (#8070)** (ad026a1)

_Recap by [Repo Wrapped](https://repowrapped.com/gh/deepspeedai/DeepSpeed?utm_source=github-action)._