Fix post-merge DCP resharding review findings + harden the flaky handoff test - #90

Merged

thad0ctor merged 27 commits into

mainfrom

fix/dcp-resharding-review-followup

Jul 17, 2026

Commits on Jul 16, 2026

Fix DCP resharding review findings: bounded save transient, honest resume error
thad0ctor
committed
93fda44
Make the DCP handoff workers report failures instead of a bare assert (#89)
thad0ctor
committed
be25ef1
Synchronize rank-local DCP validation; correct the re-blocking regimes
thad0ctor
committed
a6bcede

Commits on Jul 17, 2026

Bound the DCP save's peak memory to one slot's dense form
thad0ctor
committed
50375f2
Scope the DCP async_save claim to where torch's staging actually copies
thad0ctor
committed
8d992b9
Reject positional collision names and overflowing hyper destinations
thad0ctor
committed
f4913ed
Synchronize the DCP save's rank-local validation; pin name provenance
thad0ctor
committed
de473cc
Support async_save with a staging writer that keeps the save bounded
thad0ctor
committed
0440664
Hold the DCP save's one-slot bound under a multi-threaded writer
thad0ctor
committed
b7a160a
Page-lock one slot at a time on the bounded DCP save path
thad0ctor
committed
921c310
Scope the planner-omission and host-memory claims precisely
thad0ctor
committed
e5d736a
Hold the DCP save's fail-before-write and one-slot page-locked bounds
thad0ctor
committed
3bc6495
Await the async save through the version-tolerant response
thad0ctor
committed
c78aba5
Await the async response everywhere, and share the drain's reap budget
thad0ctor
committed
85c9d9f
Reject block state that straddles devices before the save writes
thad0ctor
committed
b635f6e
Reject one name shared by two parameters before the save
thad0ctor
committed
9632e3d
Condense the DCP adapter's comments and docstrings
thad0ctor
committed
35e75d7
Pin the async and unplanned save memory bounds
thad0ctor
committed
5056c75
Import fully_shard tolerantly across the torch 2.5 floor (#92)
thad0ctor
committed
9de5c18
Fix two DCP resharding fail-atomic/sync gaps from PR #90 review
thad0ctor
committed
1564bad
Cover the reshard levers left implicit: 2->1, CPU-offload, mixed precision
thad0ctor
committed
4f1d721
Run the whole of _validate_layout inside the DCP sync region
thad0ctor
committed
bb0edcd
Harden DCP spawn drains; defer world>1 construct-time validation; bound pinned staging across saves
thad0ctor
committed
cef24d3
Revert the ineffective across-saves pinned bound
thad0ctor
committed
c5e3121
Reject a float hyperparameter that underflows the destination to zero
thad0ctor
committed
4509add
Reject beta round-to-one, non-finite save state, and cross-rank layout drift
thad0ctor
committed
05aacc5
State the torch 2.5 fully_shard import in the resharding example
thad0ctor
committed
39c0054

Navigation Menu

Fix post-merge DCP resharding review findings + harden the flaky handoff test - #90

Fix post-merge DCP resharding review findings + harden the flaky handoff test#90
thad0ctor merged 27 commits into
mainfrom
fix/dcp-resharding-review-followup

Commits on Jul 16, 2026

Fix DCP resharding review findings: bounded save transient, honest resume error

Make the DCP handoff workers report failures instead of a bare assert (#89)

Synchronize rank-local DCP validation; correct the re-blocking regimes

Commits on Jul 17, 2026

Bound the DCP save's peak memory to one slot's dense form

Scope the DCP async_save claim to where torch's staging actually copies

Reject positional collision names and overflowing hyper destinations

Synchronize the DCP save's rank-local validation; pin name provenance

Support async_save with a staging writer that keeps the save bounded

Hold the DCP save's one-slot bound under a multi-threaded writer

Page-lock one slot at a time on the bounded DCP save path

Scope the planner-omission and host-memory claims precisely

Hold the DCP save's fail-before-write and one-slot page-locked bounds

Await the async save through the version-tolerant response

Await the async response everywhere, and share the drain's reap budget

Reject block state that straddles devices before the save writes

Reject one name shared by two parameters before the save

Condense the DCP adapter's comments and docstrings

Pin the async and unplanned save memory bounds

Import fully_shard tolerantly across the torch 2.5 floor (#92)

Fix two DCP resharding fail-atomic/sync gaps from PR #90 review

Cover the reshard levers left implicit: 2->1, CPU-offload, mixed precision

Run the whole of _validate_layout inside the DCP sync region

Harden DCP spawn drains; defer world>1 construct-time validation; bound pinned staging across saves

Revert the ineffective across-saves pinned bound

Reject a float hyperparameter that underflows the destination to zero

Reject beta round-to-one, non-finite save state, and cross-rank layout drift

State the torch 2.5 fully_shard import in the resharding example