12月1月2月3月4月5月6月7月8月9月10月11月
周一周三周五
一年内 260365 次贡献
更少更多
- frostbyte_neo 推送了仓库 github-featured/deepspeedai--deepspeed 的 tjruwase/pin-memory-route-zero 分支比较 3 提交 »aebb281901 Keep pinned reload sources alive until ZeRO-1/2 synchronizes.e6dc5f9d87 Fix CPU CI skip for offload pin-route unit test37a5bf184e Route ZeRO/SuperOffload pin sites through accelerator pin_memory
- frostbyte_neo 推送了仓库 github-featured/deepspeedai--deepspeed 的 tjruwase/native-pin-device-register 分支比较 2 提交 »bbfd83a623 Do not free native pin pages if device unregister fails.e6bea83b5d Register native mlock pages with CUDA so GPU DMA can use them.
- frostbyte_neo 推送了仓库 github-featured/deepspeedai--deepspeed 的 loadams/update-workflow-steps 分支4cd6ee0019 Update versions of steps in workflows
- frostbyte_neo 推送了仓库 github-featured/deepspeedai--deepspeed 的 nonreentrant-activation-offload 分支比较 10 提交 »9eb6e1d447 docs: note async cpu_checkpointing perf and expandable_segmentsb8f76a98eb Tersify activation offload comments50d0609ac4 Wire async CPU offload into native cpu_checkpointing687d5f6fd9 Drop Microsoft copyright from activation offload filesc9c579cd0b Tersify activation offload comments
- frostbyte_neo 推送了仓库 github-featured/deepspeedai--deepspeed 的 gma/autotp-per-model-meta 分支比较 10 提交 »0fae83e711 Keep the multi-model AutoTP regression tests inside tests/unit/v1e9612b3c77 Revert move of model_parallelism tests out of tests/unit/v1bb63a3e427 Import print_dist for the replicated-grad-hook log line606ec52338 Address review: fix step numbering; make lm-head meta test two real layers123912f945 Move tests/unit/model_parallelism to tests/unit/v1/autotp
- frostbyte_neo 推送了仓库 github-featured/deepspeedai--deepspeed 的 loadams/update-pytest-version-dependabot 分支73a4ed0ed3 Merge branch 'master' into loadams/update-pytest-version-dependabot
- frostbyte_neo 推送了仓库 github-featured/deepspeedai--deepspeed 的 docs/cpu-checkpointing-perf-note 分支3108014efd docs: note async cpu_checkpointing perf and expandable_segments
- frostbyte_neo 推送了仓库 github-featured/deepspeedai--deepspeed 的 copilot/rewrite-git-history-dco 分支3f86a99905 Update version.txt pre-0.19.4 release
- frostbyte_neo 推送了仓库 github-featured/deepspeedai--deepspeed 的 rtd-staging 分支比较 404 提交 »99b30652e0 RTD: remove native pin backend docs not in this PR7db4b06d2c RTD: drop libaio/async_io from pin_memory op docsbc07cc9e37 RTD: document pin_memory op in Host Memory Pinning98670e13f6 Trim shared behaviors from torch vs native pin comparison table1f64b281a1 Note that torch-pinned buffers are visible to AIO/GDS
- frostbyte_neo 推送了仓库 github-featured/deepspeedai--deepspeed 的 ci/model-parallelism-gpu-workflow 分支3ac9b67f85 Run tests/unit/model_parallelism in the modal GPU workflow
- frostbyte_neo 推送了仓库 github-featured/deepspeedai--deepspeed 的 master 分支比较 71 提交 »edaa722126 Fix ZeRO-1/2 with zero-sized parameters (#8280)11b518a00f [typo] Add raise for RuntimeError (#8281)9a297cfa74 Enable activation offloading (#8255)858e91eea0 Non-reentrant activation checkpoint CPU offload (#8282)35c8b03cd6 Make pass contracts cover the passes DeepCompile actually schedules; Renaming files (#8251)
