## deepspeedai/DeepSpeed — v0.19.0…v0.19.1

_22 commits._

### Features
- **Support flash-attn 2.7.0 in FPDT attention (#8022)** (4542922)
- **[Blog] Muon Optimizer Support in DeepSpeed (#7962)** (de47309)
- **Support bf16 optimizer states with CPU offload (#8010)** (3c337b5)
- **Add configurable torch-latest dependency versions (#8016)** (b387166)
- **Add office hours times/link on the README (#8004)** (883fefb)

### Fixes
- **Fix DeepCompile AOT kwargs patching for PyTorch >= v2.11 (#8024)** (510ebe5)
- **fix(zero): enable vmap on LinearFunctionForZeroStage3 (#8023)** (ae576f8)
- **fix: add setup_context for torch.func compatibility (#7916)** (67b9e21)
- **fix gemma4 num attention head bugs (from #7975) (#7990)** (bd71eed)
- **[fix] fix test_zf.py hang bug (#8012)** (26ee385)
- **Fix ZeRO-3 forward crash on modules with plain dict _parameters (#8009)** (d7a3972)
- **fix: use subprocess instead of os.system in data_analyzer.py (#7994)** (8cdf865)
- **fix(io): close aio_fd in FastFileWriter._fini to prevent fd leak (#8005)** (b01a091)

### Backend
- **Sort and dedupe -gencode flags emitted by op_builder.builder (#8021)** (4316646)
- **Make GitHub Actions job names unique (#8014)** (9f95452)
- **Run FastFileWriter fd-close test outside pytest-forked (#8015)** (d6d4b3d)
- **Remove stale step() docstring from DeepSpeedCPUAdam (#8011)** (ae07514)
- **Auto-detect CUTLASS for EvoformerAttention (#8000)** (d5356e0)
- **zero3: SDMA allgather via mori (sdma_allgather)  (#7999)** (66af8f0)
- **Optimize singleton MoE collectives (#7997)** (5549141)
- **Update topkgating probability-mask test expectation (#8007)** (5be4c70)
- **Update version post-v0.19.0 release (#7996)** (a781199)

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