Repository navigation
What's Changed
- fix dstorage test in Linux by @takeshi-yoshimura in #112
- Add FGDS (an alternative solution of GDS) copier by @kk47 in #111
- Reuse checkpoint metadata across chunk batches by @takeshi-yoshimura in #113
- Account for fixed copier pools in memory budget planning by @takeshi-yoshimura in #114
- Build PyTorch tensor views from shared typed storage by @takeshi-yoshimura in #115
- Document which close() releases device memory by @gitbisector in #116
- Pass group ranks to collectives on process subgroups by @gitbisector in #117
- Reuse nogds workers and add NUMA placement for GPUs by @takeshi-yoshimura in #119
- Split cuFile reads across persistent workers by @takeshi-yoshimura in #120
- windows cuda test fixes by @takeshi-yoshimura in #121
- Add ParallelSaver and generic sharded checkpoint writing by @takeshi-yoshimura in #122
- improve broadcast perf by @takeshi-yoshimura in #123
- Overlap checkpoint IO with tensor consumption by @takeshi-yoshimura in #124
- tensor lifetime and borrow by @takeshi-yoshimura in #125
- Add bounded-memory GDS loading with aligned chunk reads by @takeshi-yoshimura in #126
- Default nogds to filesystem-aware O_DIRECT by @takeshi-yoshimura in #128
- Fall back to buffered reads when a filesystem rejects O_DIRECT in nogds by @gitbisector in #131
- fix tests by @takeshi-yoshimura in #133
- standalone benchmark by @takeshi-yoshimura in #134
- Add Python 3.15 support by @takeshi-yoshimura in #135
- Bump version to 0.5.0 by @takeshi-yoshimura in #136
New Contributors
Full Changelog: 0.4.0...0.5.0