Commit 52e18c8
authored
Distributed optimizer support for multiple dtypes (#1721)
* Distopt support for multiple dtypes
Rough draft.
Signed-off-by: Tim Moon <[email protected]>
* Add test for distopt with multiple dtypes
Signed-off-by: Tim Moon <[email protected]>
* Debug typos
Signed-off-by: Tim Moon <[email protected]>
* Support for contiguous buffers with multiple dtypes
Signed-off-by: Tim Moon <[email protected]>
* Fix bug when calling bf16 remainder kernel
Signed-off-by: Tim Moon <[email protected]>
* Return distopt checkpoint on all ranks
Signed-off-by: Tim Moon <[email protected]>
* Tweak docs
Signed-off-by: Tim Moon <[email protected]>
* Black formatting
Handle case where we load old checkpoints without multi-dtype support
Signed-off-by: Tim Moon <[email protected]>
* Tweak logic for BF16 param remainder kernel
Signed-off-by: Tim Moon <[email protected]>
* Skip distopt test if import fails
Signed-off-by: Tim Moon <[email protected]>
---------
Signed-off-by: Tim Moon <[email protected]>1 parent 730376f commit 52e18c8
2 files changed
Lines changed: 534 additions & 396 deletions
0 commit comments