## deepspeedai/DeepSpeed — v0.18.9…v0.19.0

_28 commits._

### Features
- **[AutoSP] (Sequence Parallelism) support for Multimodal Models (ViT + LLM) (#7984)** (4e668fc)
- **Add Gram Newton-Schulz orthogonalization for Muon optimizer    (#7953)** (8a77f38)
- **Add DeepSpeed NVTX domain support (#7988)** (5b44102)
- **[SP] add SP deny list instead of allow (#7887)** (bf0126b)

### Fixes
- **fix: topkgating major bug (#7986)** (853c938)
- **Fix BF16_Optimizer last-microbatch grad leak under ZeRO-1 (#7985)** (5999fb0)
- **Fix modal ci timeout (#7989)** (6e291be)
- **Fix Adam subgroup inconsistency (#7982)** (44c51e3)
- **deepcompile: Fix backward graph recompilation due to unbalanced forward/backward visits (#7980)** (077bff5)
- **Fix ZeRO-1/2 CPU-offloaded gradient loss with multiple backward() per step (#7981)** (aeb10bb)
- **fix(op_builder): avoid duplicate/wrong -gencode flags (#7974)** (893c6d2)
- **fix(fp_quantizer): fix UB and negative shift warnings in fp_quantize_impl.cu (#7973)** (3ba3dc9)
- **Fix WarmupCosineLR multi-group initialization (#7969)** (3fd762c)
- **Fix DeepCompile+Z3 on PyTorch v2.9/2.10 (#7951)** (ecb26a5)
- **Fix overlap-comm buffer lifetimes (#7965)** (dac1525)
- **Fix zero 1 and 2 CPU-offloaded gradient norm (#7967)** (9d632f1)
- **Fix FPQuantizer build (#7963)** (3a5df51)
- **fix(zero): detach flat buffer to prevent autograd inplace error on CP… (#7948)** (37e232f)
- **Fix process hang in process-group shutdown (#7941)** (2f0924a)
- **Fix/fix autotp universal checkpoint ci (#7937)** (3bdebc0)

### Backend
- **Update version.txt before 0.19.0 release (#7995)** (98a2fbd)
- **Dynamic offload compatible with static optimizer offload (#7979)** (f6d7ae4)
- **Avoid CUDA reinit error in CI tests (#7977)** (83d43c3)
- **Rename dequantization template parameters (#7976)** (dc0fd29)
- **Enable PyTorch version selection for full test (#7968)** (0ba2352)
- **Zero3 defragment utility (#7940)** (5dce124)
- **Refactor consolidate transpose (#7934)** (89bf0d2)
- **Update version after latest release (v0.18.9) (#7936)** (607b55f)

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