[modular] improve auto offload - #14304

Open

yiyixuxu wants to merge 18 commits into

mainfrom

auto-offload-v2

Commits on Jul 27, 2026

[modular] auto offload: tunable memory reserve, OOM retry, non-disruptive add/remove

yiyixuxu
and
claude
committed
780e112

Commits on Jul 28, 2026

[modular] drop stale "dynamic mode" wording from the offload test comments

yiyixuxu
and
claude
committed
d4b4edd
Record what the auto-offloader does

yiyixuxu
and
claude
committed
34036c1

Commits on Jul 29, 2026

Group the offload record table by decision

yiyi@huggingface.co
and
claude
committed
24460ae
Split display helpers into components_manager_utils; drop activation measurement; cover encode/decode in the OOM retry

yiyixuxu
and
claude
committed
b755e08
Trim the offload record to consumed fields; hook reads available memory itself

yiyixuxu
and
claude
committed
d2eb134
Make offload record events pure moves

yiyixuxu
and
claude
committed
09d5ec7
Drop OffloadRecord.summary(): the table is the record

yiyixuxu
and
claude
committed
d3c1e9d
Address review: discover entry points via an apply_forward_hook marker

yiyixuxu
and
claude
committed
b360420
Inline the OOM-retry helper into wrap_forward, its only caller

yiyixuxu
and
claude
committed
b313f79
Shrink ComponentsManager to its used surface

yiyixuxu
and
claude
committed
0e512dd
link issue to TODO
yiyixuxu
committed
77615c0

Commits on Jul 30, 2026

Address review: record at call sites, one device rule, fewer moving parts

yiyixuxu
and
claude
committed
ba14c2a
Overhaul the ComponentsManager test suite around real behavior

yiyixuxu
and
claude
committed
3f56149
Add slow modular-pipeline integration tests with declared offload behavior

yiyixuxu
and
claude
committed
29bd6e0

Commits on Jul 31, 2026

Restore ModularPipelineOffloadTesterMixin to its original form

yiyixuxu
and
claude
committed
e20dc87
add set_offload_strategy, Peak column in offload record, and offloading docs

yiyixuxu
and
claude
committed
0750134
make style

yiyixuxu
and
claude
committed
9a1c508