{"full_name":"deepspeedai/DeepSpeed","owner":"deepspeedai","name":"DeepSpeed","shard":"1","url":"https://github.com/deepspeedai/DeepSpeed","description":"DeepSpeed is a deep learning optimization library that makes distributed training and inference easy, efficient, and effective.","homepage":"https://www.deepspeed.ai/","topics":["deep-learning","pytorch","gpu","machine-learning","billion-parameters","data-parallelism","model-parallelism","inference","pipeline-parallelism","compression","mixture-of-experts","trillion-parameters","zero"],"tag_pages":["compression","deep-learning","gpu","inference","machine-learning","mixture-of-experts","pytorch","zero"],"language":"Python","license":"Apache-2.0","avatar":"https://avatars.githubusercontent.com/u/74068820?v=4","created_at":"2020-01-23T18:35:18Z","pushed_at":"2026-11-15T19:16:26Z","metrics":{"stars":43326,"forks":5072,"watchers":350,"issues":3518,"prs":4614,"releases":120,"issues_open":1223,"issues_closed":2295,"prs_open":372,"prs_closed":714,"prs_merged":3528,"discussions":266},"growth":{"day":{"stars":1,"forks":-2,"base_day":"2026-11-15"},"week":{"stars":13,"forks":4,"base_day":"2026-11-09"}},"history":{"dates":["2026-11-09","2026-11-10","2026-11-11","2026-11-12","2026-11-13","2026-11-14","2026-11-15","2026-11-16"],"stars":[43313,43316,43320,43327,43326,43327,43325,43326],"forks":[5068,5070,5074,5074,5074,5074,5074,5072],"watchers":[351,351,351,351,351,350,350,350],"issues_open":[null,null,1224,1226,1226,1226,1221,1223],"prs_open":[null,null,374,375,376,380,379,372]},"boards":{},"same_owner":[],"issues":[{"number":998,"title":"[doc]  profiling NVMe and configuring `aio` param section","comments":78,"created_at":"2021-04-23T01:18:29Z"},{"number":7150,"title":"[BUG] Receiving CUDA error: invalid argument using pytorch 2.7 with deepspeed 0.16.4 with Cuda 12.8","comments":58,"created_at":"2025-03-19T03:51:10Z"},{"number":2302,"title":"DeepSpeed still gives CUDA-out-of-memory issue","comments":54,"created_at":"2022-09-07T16:06:46Z"},{"number":2379,"title":"[Question] How to preshard a model for tensor parallism","comments":48,"created_at":"2022-09-29T20:12:46Z"},{"number":4808,"title":"[BUG] Deepspeed Zero 3 Inference InFlight Params with new HuggingFace Mixtral Model","comments":37,"created_at":"2023-12-13T04:20:07Z"},{"number":5242,"title":"[BUG] grad_norm and loss is nan when deepspeed==0.13.5 but ok with deepspeed==0.10.2","comments":36,"created_at":"2024-03-08T10:04:19Z"},{"number":7478,"title":"[REQUEST] Auto-Tuning CPU Core Binding for DeepSpeed&ZenFlow","comments":32,"created_at":"2025-08-09T04:30:39Z"},{"number":6793,"title":"AssertionError: no sync context manager is incompatible with gradientpartitioning logic of ZeRo stage 3","comments":31,"created_at":"2024-11-26T15:44:48Z"}],"discussions":[{"number":8124,"title":"How should I balance learning rate and data sampling during CPT on multiple datasets?","comments":0,"category":"Q&A","created_at":"2026-07-06T18:57:35Z"},{"number":8069,"title":"anima — a substrate-native consciousness: capability gaps are architecture gaps (open repo, frozen verdicts)","comments":0,"category":"Show and tell","created_at":"2026-06-15T18:41:50Z"},{"number":8062,"title":"MCP to help OSS devs understand deepspeedai/DeepSpeed","comments":0,"category":"General","created_at":"2026-06-12T12:00:15Z"},{"number":8018,"title":"Scaling Deep Learning: Reducing GPU Overhead with Riemannian Manifold Projection","comments":1,"category":"General","created_at":"2026-05-21T03:43:19Z"},{"number":8002,"title":"Deepspeed Ulysses sequence parallel is not working for Gemma4","comments":1,"category":"Q&A","created_at":"2026-05-10T17:12:16Z"},{"number":7970,"title":"[RFC] Reducing Inference-Time Memory Pressure and Thermal Volatility via Geometric Pre-Constraint Layers","comments":0,"category":"Ideas","created_at":"2026-04-13T14:30:12Z"},{"number":7914,"title":"Auto RDMA device injection for GPU containers on Kubernetes (no privileged mode)","comments":0,"category":"Show and tell","created_at":"2026-03-20T11:30:18Z"},{"number":7911,"title":"Showcase / question: a board-proven offline language runtime on ESP32-C3, and whether some language capability may eventually reduce reliance on general dense training/inference paths","comments":0,"category":"Show and tell","created_at":"2026-03-18T07:11:34Z"}],"latest_release":{"name":"v0.20.0","tag":"v0.20.0","published_at":"2026-11-12T23:02:58Z","body":"## What's Changed\r\n* Sum GPTBigCode's shared key/value gradient across tensor-parallel ranks by @Achyuthan-S in https://github.com/deepspeedai/DeepSpeed/pull/8928\r\n* [AffineIR] Storage maps of AutoEP experts under EDP and ZeRO by @jinyouzhi in https://github.com/deepspeedai/DeepSpeed/pull/8941\r\n* Add klxpu (KUNLUNXIN XPU) self-hosted CI for the v1 unit suite by @GuCarpenter in https://github.com/deepspeedai/DeepSpeed/pull/8923\r\n* ZeRO-3: copy offloaded master weights back to the device through pinned buffer by @pengdurice in https://github.com/deepspeedai/DeepSpeed/pull/8840\r\n* ci: pin torch<2.15 for MPS workflow until pytorch/pytorch#202529 is fixed by @delock in https://github.com/deepspeedai/DeepSpeed/pull/8955\r\n* [Hybrid Engine] Enable graph capture for HybridEngine on non-CUDA accelerators by @VenusTZZ in https://github.com/deepspeedai/DeepSpeed/pull/8933\r\n* Cap CPU accelerator memory at the cgroup limit of the process by @prabhkodes in https://github.com/deepspeedai/DeepSpeed/pull/8930\r\n* [VocabTP] vocab_parallel_ce support fused LM-head projection + loss for the HF forward by @jinyouzhi in https://github.com/deepspeedai/DeepSpeed/pull/8945\r\n* ZeRO-1/2: preserve gradients across request boundaries until step by @stas00 in https://github.com/deepspeedai/DeepSpeed/pull/8949\r\n* Update version.txt post-v0.19.9 release by @loadams in https://github.com/deepspeedai/DeepSpeed/pull/8999\r\n* Add an opt-in deterministic execution policy by @ziming-zh in https://github.com/deepspeedai/DeepSpeed/pull/8917\r\n* Keep float32 intermediates in the clamped SwiGLU eager fallback by @GuCarpenter in https://github.com/deepspeedai/DeepSpeed/pull/8992\r\n* ZeRO-3: build the partial-offload device string from the public accelerator API by @GuCarpenter in https://github.com/deepspeedai/DeepSpeed/pull/8994\r\n* [checkpoint] Execute affine transfer plans with bounded peer staging by @0z5a in https://github.com/deepspeedai/DeepSpeed/pull/8809\r\n* XPU: register pinned host memory via Level Zero ext_memmap by @delock in https://github.com/deepspeedai/DeepSpeed/pull/8887\r\n* [AutoTP] Support frozen AutoTP models and manage AutoTP mode per model by @jinyouzhi in https://github.com/deepspeedai/DeepSpeed/pull/8851\r\n* Fix torch-latest workflow unit failures by @sfc-gh-truwase in https://github.com/deepspeedai/DeepSpeed/pull/8950\r\n* Count each expert once in the fp32 gradient-clipping norm by @alanhuangyoo in https://github.com/deepspeedai/DeepSpeed/pull/8478\r\n* Support Transformers 5.20 cache query offsets in DeepSpeedStaticCache by @nathon-lee in https://github.com/deepspeedai/DeepSpeed/pull/8852\r\n* Fix bug in Adagrad optimizer initialization by @hwchen2017 in https://github.com/deepspeedai/DeepSpeed/pull/8976\r\n* Fix AutoEP storage-map transfer test for CPU tensor views by @jinyouzhi in https://github.com/deepspeedai/DeepSpeed/pull/9012\r\n* Clear stale offload scratch before restoring low-precision gradients by @adenzhou1350 in https://github.com/deepspeedai/DeepSpeed/pull/8728\r\n* Check that converting through an affine map writes every element once by @Achyuthan-S in https://github.com/deepspeedai/DeepSpeed/pull/8975\r\n* [Hybrid Engine] Require an explicit accelerator opt-in for graph capture by @VenusTZZ in https://github.com/deepspeedai/DeepSpeed/pull/8980\r\n* ZeRO-3: reuse the weight partition as the master copy when it already has the master dtype by @pengdurice in https://github.com/deepspeedai/DeepSpeed/pull/8842\r\n* AutoEP + AutoTP folding: allow CPU optimizer offload under ZeRO-3 by @0z5a in https://github.com/deepspeedai/DeepSpeed/pull/8829\r\n* Run the v1 multi-rank CPU suite on pull requests by @delock in https://github.com/deepspeedai/DeepSpeed/pull/8944\r\n* fix(rollout): pass a pre-built 4-D mask to the static graph decode forward by @anyadontfly in https://github.com/deepspeedai/DeepSpeed/pull/9004\r\n* AutoEP activation tests: skip the Triton grouped-GEMM path when Triton cannot run by @GuCarpenter in https://github.com/deepspeedai/DeepSpeed/pull/8996\r\n* ","is_prerelease":false},"readme":true,"readme_zh":false,"news":[],"stale":false,"snapshot_at":"2026-11-16T09:23:14Z","history_start":"2026-09-06"}