Add versioned trainer graphs and logical-rank APIs - #925

Draft

bradhilton wants to merge 127 commits into

mainfrom

stark/trainer-v1-20260917

Commits on Sep 18, 2026

Commits on Sep 19, 2026

refactor: remove unused memory placement selector
bradhilton
committed
6d62ce2
Stamp trainer head registration and export observations
bradhilton
committed
5475043
fix: remap trainer payloads and compact operation retirement
bradhilton
committed
0746fca
Share trainer v1 test and development setup
bradhilton
committed
d5b4fb2
Fix logical head completion and lookup boundaries
bradhilton
committed
5a8e491
Merge commit '5b7acf06d209dffa70327a810d3a9e303411ab77' into stark/review-fixes-merge-20260919
bradhilton
committed
daf6210
Expose callback release fence for checkpoint entry
bradhilton
committed
4c89eb4
Separate abandoned batch pull cleanup from iterator close
bradhilton
committed
04334c3

Commits on Sep 20, 2026

Deduplicate live-head test construction
bradhilton
committed
bbb81f3
Reuse tensor argument container traversal
bradhilton
committed
e92db78
Consolidate trainer-rank test process scaffolding
bradhilton
committed
bcaa2f0
Inherit duplicate TrainerRank public methods
bradhilton
committed
7365d9a
Remove obsolete facade type variables and imports
bradhilton
committed
8f0f35e
Reuse trainer test process group fixtures
bradhilton
committed
3fc523e
Reuse live parameter test fixtures
bradhilton
committed
1bcf119
Reuse custom tensor wrapping for module members
bradhilton
committed
0429b71
Share policy choice checks and preserve public hint bindings
bradhilton
committed
1d9d6ae
Share argument container traversal with replay restore
bradhilton
committed
101fd1f
Use real trainer typing and share CPU gradient stubs
bradhilton
committed
6e61553
Share native live-head test registration setup
bradhilton
committed
1a0bc40
Share strict gradient reducer installation in validation tests
bradhilton
committed
d99dde4
Share native and live head tensor inventories
bradhilton
committed
ee5415e
test: share trainer snapshot setup in version tests
bradhilton
committed
fb3be23
test: share CP2 attention layout setup
bradhilton
committed
f197898
test: finish sharing initial version snapshots
bradhilton
committed
36b127f
test: share inline trainer operation capture
bradhilton
committed
de6a8d9
test: run trainer operation cases directly as async tests
bradhilton
committed
30ef515
test: remove redundant trainer async wrappers
bradhilton
committed
c2fb6f8

Commits on Sep 21, 2026

test: share trainer operation imports
bradhilton
committed
797f463
Consolidate initialized trainer test imports
bradhilton
committed
ba8e720
Share standard-library imports in trainer tests
bradhilton
committed
38b53de
Consolidate remaining trainer test imports
bradhilton
committed
5b7e1c6
Share trainer head mutation-name constants
bradhilton
committed
5ef8489
Coordinate authority buffer snapshot failures across ranks
bradhilton
committed
1a13ee6
Release undelivered trainer output graphs on attachment failure
bradhilton
committed
b73cd1b
Preserve delivery errors while closing trainer iterators
bradhilton
committed
a75373b
Merge main vocabulary-head lifetime fix into trainer v1
bradhilton
committed
696e24b
Fix trainer output delivery test typing
bradhilton
committed
022eeaa
Keep Tinker client imports isolated from backend initialization
bradhilton
committed
595f12c
Release logical forward batches before advancing
bradhilton
committed
4d9c631

Commits on Sep 26, 2026

Commits on Sep 27, 2026

Merge ART main c0b1296d for repeated Tau cancellation cleanup
bradhilton
committed
3a159e7
Share exact checkpoint runtime constructor across trainer tests
bradhilton
committed
70c0b33
Retain explicit Linear type in command transport worker
bradhilton
committed
2099ee2
Share lightweight GPT model fixture across trainer tests
bradhilton
committed
0debb26
Simplify packed trainer test setup
bradhilton
committed
924000d
Remove redundant correction kind state
bradhilton
committed
9f6724f
Remove unused native module owner state
bradhilton
committed
70588af
Simplify inherited checkpoint finalization arbitration
bradhilton
committed
a4187dc
Merge exact ART main 762c89df into trainer-v1
bradhilton
committed
6918176
Allow cleanup-only abort after checkpoint finish
bradhilton
committed
5d61a6b

Commits on Sep 28, 2026

Share fixed fake rank constructors in memory tests
bradhilton
committed
c57cb1c
Merge current ART main with fake rank test cleanup
bradhilton
committed
9e19ef6
Merge current ART main into trainer-v1 candidate
bradhilton
committed
a704674
Share full recompute config in memory fixtures
bradhilton
committed
e4545fc
Merge ART main checkpoint-floor reuse and GPU validation reuse
bradhilton
committed
788299c
Share real transformer model shells in memory fixtures
bradhilton
committed
4a3da11
Simplify checkpoint memory test cases and config setup
bradhilton
committed
067b203
Consolidate checkpoint profile and TP pricing test scenarios
bradhilton
committed
f89d83b
Normalize packed GPT-OSS exports in the native builder
bradhilton
committed
63b91ca
Merge qualified checkpoint fixture reductions
bradhilton
committed
ef7367f
Merge fixed ART main trajectory serialization updates
bradhilton
committed
949f03e
Merge fixed ART main string subclass interning update
bradhilton
committed
f347e1d
Compose native export correction and checkpoint test reuse
bradhilton
committed
46c9110
Remove unused GPT-OSS interleaved padding helper
bradhilton
committed
9bfdc4e
Compose checkpoint export synchronization and accepted private simplifications
bradhilton
committed
b745db5
Share LoRA publication runtime validation
bradhilton
committed
3777c0b
Require validated runtime for LoRA publication preparation
bradhilton
committed
0fe6f91
Test second export metadata failure on rank one
bradhilton
committed
e2852ca
Merge ART main e65b0dd preserving literal assistant history
bradhilton
committed
6222b3c
Qualify joined-reasoning test fixture import
bradhilton
committed
8496d79
Merge ART memory extraction while preserving trainer-v1 admission
bradhilton
committed
b4da074
Remove obsolete thinking workaround and strengthen branch coverage
bradhilton
committed
bcb3cc8
Merge ART planner and checkpoint extractions preserving trainer-v1
bradhilton
committed
e80192e
Merge exact ART tokenization update preserving trainer-v1 controls
bradhilton
committed
74b5d9f
Fix runtime annotations on extracted public trainer methods
bradhilton
committed
1a02ed7
Remove stale imports and repeated extraction notes
bradhilton
committed
7ce1cc8
Preserve literal renderer oracles across explicit override modes
bradhilton
committed
7c0d0b9
Respect prompt provenance in the explicit template override control
bradhilton
committed
b13b5f5
Share recorded boundary invocation setup in tests
bradhilton
committed
8711cc5
Merge main event-loop isolation into trainer-v1
bradhilton
committed
ebf3ccf
Merge ART main 71af7792 planner input snapshots
bradhilton
committed
d9e9061
Merge main a68fa500 planner retention diagnostics
bradhilton
committed
620a1f0
Merge current main planner diagnostic bounds into trainer-v1
bradhilton
committed
b5e65a8
Release registered forward graphs when correction setup fails
bradhilton
committed
24bdb52
Merge current main planner diagnostic nesting bounds into trainer-v1
bradhilton
committed
19db288
Compose trainer handoff and callback cleanup fixes
bradhilton
committed
35498de
Merge bounded grouped planner replay into trainer v1
bradhilton
committed
22cd501
Preserve placed report completeness and failed graph ownership
bradhilton
committed
f002b10

Commits on Sep 29, 2026

Release failed native forward capture aliases
bradhilton
committed
d758c0a
Release failed correction capture copies
bradhilton
committed
4fc438c
Release failed forward captures before graph registration
bradhilton
committed
43b65bb