{"version":"https://jsonfeed.org/version/1.1","title":"DeepSpeed — Lab Ledger Desk","home_page_url":"https://labledgerdesk.pages.dev/lab/deepspeed/","feed_url":"https://labledgerdesk.pages.dev/lab/deepspeed/feed.json","description":"Open-source DeepSpeed releases filed by Lab Ledger Desk.","items":[{"id":"https://labledgerdesk.pages.dev/b/2026/11/12/deepspeed-v0-20-0/","url":"https://labledgerdesk.pages.dev/b/2026/11/12/deepspeed-v0-20-0/","title":"DeepSpeed v0.20.0","content_text":"What's Changed Sum GPTBigCode's shared key/value gradient across tensor-parallel ranks by @Achyuthan-S in #8928 [AffineIR] Storage maps of AutoEP experts under EDP and ZeRO by @jinyouzhi in #8941 Add klxpu (KUNLUNXIN XPU) self-hosted CI for the v1 unit suite by @GuCarpenter in #8923 ZeRO-3: copy offloaded master…","date_published":"2026-11-12T23:02:58.000Z","authors":[{"name":"DeepSpeed"}],"tags":["note"],"external_url":"https://github.com/deepspeedai/DeepSpeed/releases/tag/v0.20.0"},{"id":"https://labledgerdesk.pages.dev/b/2026/11/12/nightly-last-green-checkpoint-batch-replica-selection-for-affine/","url":"https://labledgerdesk.pages.dev/b/2026/11/12/nightly-last-green-checkpoint-batch-replica-selection-for-affine/","title":"nightly-last-green: Fix unit tests failure on CPU with fp16 support (#9086)","content_text":"On runners whose CPU supports oneDNN fp16 (Sapphire Rapids or newer), fp16 tests that normally skip start running, and Muon's Newton-Schulz switches to fp16. cpu-torch-latest-v1-multirank then fails 10 tests, e.g. runs 43556673085 and 43396382279 (identical values). …","date_published":"2026-11-12T07:10:31.000Z","authors":[{"name":"DeepSpeed"}],"tags":["note","hardware"],"external_url":"https://github.com/deepspeedai/DeepSpeed/releases/tag/nightly-last-green"},{"id":"https://labledgerdesk.pages.dev/b/2026/10/29/v0-19-9-patch-release/","url":"https://labledgerdesk.pages.dev/b/2026/10/29/v0-19-9-patch-release/","title":"v0.19.9 Patch Release","content_text":"What's Changed Update version post 0.19.8 release by @loadams in #8871 Update to filter out version.txt from PR gates. …","date_published":"2026-10-29T23:42:23.000Z","authors":[{"name":"DeepSpeed"}],"tags":["launch","hardware"],"external_url":"https://github.com/deepspeedai/DeepSpeed/releases/tag/v0.19.9"},{"id":"https://labledgerdesk.pages.dev/b/2026/10/29/nightly-cpu-multi-last-green-add-a-v1-multi-rank-cpu-suite-manual/","url":"https://labledgerdesk.pages.dev/b/2026/10/29/nightly-cpu-multi-last-green-add-a-v1-multi-rank-cpu-suite-manual/","title":"nightly-cpu-multi-last-green: Fix multiple issues in Op Builder and refactor implementation (#9085)","content_text":"Problems fixed Pre-built ops lose GPU architectures. Ops that filter their targets (e.g. fp_quantizer ) wrote the reduced list back to TORCH_CUDA_ARCH_LIST , and every op setup.py built after them inherited it. With 7.0;7.5;8.0;9.0 , 10 ops shipped without sm_70/sm_75 code. …","date_published":"2026-10-29T00:12:13.000Z","authors":[{"name":"DeepSpeed"}],"tags":["launch","hardware"],"external_url":"https://github.com/deepspeedai/DeepSpeed/releases/tag/nightly-cpu-multi-last-green"},{"id":"https://labledgerdesk.pages.dev/b/2026/10/20/v0-19-8-patch-release/","url":"https://labledgerdesk.pages.dev/b/2026/10/20/v0-19-8-patch-release/","title":"v0.19.8 Patch Release","content_text":"What's Changed Deprecate leftovers of compression library by @hwchen2017 in #8535 Update version.txt after 0.19.7 release by @loadams in #8555 Deprecate loco zero++ by @hwchen2017 in #8534 Deprecate sparse_gradients by @pengdurice in #8504 docs: stop Codex false-alarm on DCO sign-off by @delock in #8561 Raise Modal…","date_published":"2026-10-20T15:21:47.000Z","authors":[{"name":"DeepSpeed"}],"tags":["launch"],"external_url":"https://github.com/deepspeedai/DeepSpeed/releases/tag/v0.19.8"},{"id":"https://labledgerdesk.pages.dev/b/2026/10/11/nightly-bisect-c/","url":"https://labledgerdesk.pages.dev/b/2026/10/11/nightly-bisect-c/","title":"nightly-bisect/c","content_text":"DeepSpeed published release nightly-bisect/c on its official releases feed. The feed did not include notes, so this brief is the version and the primary source, nothing more.","date_published":"2026-10-11T23:58:32.000Z","authors":[{"name":"DeepSpeed"}],"tags":["launch"],"external_url":"https://github.com/deepspeedai/DeepSpeed/releases/tag/nightly-bisect%2Fc"},{"id":"https://labledgerdesk.pages.dev/b/2026/10/11/nightly-bisect-b-autosp-honor-seq-dim-when-autosp-shards-inputs-8720/","url":"https://labledgerdesk.pages.dev/b/2026/10/11/nightly-bisect-b-autosp-honor-seq-dim-when-autosp-shards-inputs-8720/","title":"nightly-bisect/b: [AutoSP] Honor seq_dim when AutoSP shards inputs (#8720)","content_text":"Split from #8457 . Problem prepare_autosp_inputs() takes a seq_dim argument, but only passes it to mark_dynamic . The AutoSP graph passes ignore it: pass_shard_seq_dim and shard_tensor_node always read the sequence length from shape[1] and slice dimension 1. …","date_published":"2026-10-11T22:46:09.000Z","authors":[{"name":"DeepSpeed"}],"tags":["note","llm","hardware"],"external_url":"https://github.com/deepspeedai/DeepSpeed/releases/tag/nightly-bisect%2Fb"},{"id":"https://labledgerdesk.pages.dev/b/2026/10/11/nightly-bisect-a-bug-fix-fix-cpuadagrad-updates-with-nonzero-weight/","url":"https://labledgerdesk.pages.dev/b/2026/10/11/nightly-bisect-a-bug-fix-fix-cpuadagrad-updates-with-nonzero-weight/","title":"nightly-bisect/a: [Bug Fix] Fix CPUAdagrad updates with nonzero weight decay (#8781)","content_text":"Fix CPUAdagrad to use the weight-decay-adjusted gradient in both the accumulator and parameter update, matching torch.optim.Adagrad . Previously, the update numerator incorrectly used the original gradient. Covers scalar and SIMD implementations, including the XPU copies. …","date_published":"2026-10-11T15:21:59.000Z","authors":[{"name":"DeepSpeed"}],"tags":["note","hardware"],"external_url":"https://github.com/deepspeedai/DeepSpeed/releases/tag/nightly-bisect%2Fa"}]}