Describe universal checkpoint shards as affine maps #8385
delock
merged
12 commits into
deepspeedai:master
from
Achyuthan-S:affine-ir-shard-map
delock
commented
on 2026-09-04
delock
commented
on 2026-09-04
Add affine shard map for universal checkpoint
b6b0d505
Add scale and homogeneity to the affine shard map
13e2c649
Lower the current checkpoint metadata to affine maps
1d1d6039
Convert tp slices through an affine map when one is present
3cf12341
Build shard metadata per model in the affine map tests
166a3d2f
Refuse an affine map format newer than the reader understands
91499088
Keep the stride helper private
d63251bb
Address review: scale optimizer states by their own power
2e846d91
Address review: header, scale in the piece tuple, scaling example
fcec0621
Refuse to convert a scaled optimizer state
10bde3e7
delock
approved these changes
on 2026-09-10
Carry the scale on the replicated map, not the split
d02c56bf
Fix universal checkpoint resume across AutoTP sizes
dd5c57f1
Assignees
No one assigned
Labels
None yet
Milestone
No milestone
Login to write a write a comment.
Login via GitHub