Skip to content

[lora] feat: multi lora async (full PR, see stacked diff PRs instead) #1638

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
yushengsu-thu merged 71 commits into from
Jul 21, 2026
+0 −0
Merged

[lora] feat: multi lora async (full PR, see stacked diff PRs instead) #1638

Changes from all commits
Commits
Show all changes
71 commits
Select commit Hold shift + click to select a range
def3841
[feat] multi lora fully async
mathewjhan Jul 12, 2026
2f4b851
[feat] multi lora fully async
mathewjhan Jul 12, 2026
c1134f8
[misc] revert minor changes
mathewjhan Jul 12, 2026
930a1d2
[fix] deterministic ordering
mathewjhan Jul 12, 2026
cf37c3f
[chore] precommit
mathewjhan Jul 12, 2026
1bee66b
Merge branch 'main' into feat/multi-lora-async
mathewjhan Jul 12, 2026
629dc85
[fix] import paths for tests
mathewjhan Jul 12, 2026
9ca4536
[test] move tests
mathewjhan Jul 12, 2026
e371b2e
[fix] tests + keep original lora behavior for non multi-lora
mathewjhan Jul 12, 2026
8bdd7d4
[fix] recompute-logprobs uses per-sample adapter lora_path in multi-LoRA
yushengsu-thu Jul 13, 2026
1e93ea4
[fix] harden retired-adapter teardown against orphaned rollout requests
yushengsu-thu Jul 13, 2026
50f6ff1
[test] mark recycle-aborted as xfail until re-queuing lands
yushengsu-thu Jul 13, 2026
a75a5b4
[fix] reject multi-LoRA with --sglang-tokenizer-worker-num > 1 at launch
yushengsu-thu Jul 14, 2026
c5cc40a
[feat] optimizer changes initial commit
mathewjhan Jul 15, 2026
5c4a4e8
[fix] typo lol
mathewjhan Jul 16, 2026
f115610
[fix] handle empty
mathewjhan Jul 16, 2026
52b0410
[feat] use num_step instead of num_row
mathewjhan Jul 16, 2026
d81bf52
[fix] missing import
mathewjhan Jul 16, 2026
5f1e83f
[fix] smoke
mathewjhan Jul 16, 2026
7245240
[test] move tests to test dir
mathewjhan Jul 16, 2026
9aea756
[chore] improve metrics
mathewjhan Jul 16, 2026
c4bed1d
[test] clean up some tests
mathewjhan Jul 16, 2026
2100479
[feat] improve metrics
mathewjhan Jul 16, 2026
d972436
[fix] metrics for queue don't need step
mathewjhan Jul 16, 2026
ce6c2dd
[fix] metrics glob expansion
mathewjhan Jul 16, 2026
54e13f8
[fix] steps
mathewjhan Jul 16, 2026
07d9e9b
[chore] minor naming
mathewjhan Jul 16, 2026
031c086
Revert "[fix] steps"
mathewjhan Jul 16, 2026
e22ca8d
Merge remote-tracking branch 'radixark/main' into feat/multi-lora-async
yushengsu-thu Jul 17, 2026
8fa298a
[fix] metric naming
mathewjhan Jul 17, 2026
75d7538
[feat] improve adapter metrics + adapter metric use real adapter steps
mathewjhan Jul 17, 2026
793715a
[test]
mathewjhan Jul 17, 2026
3c1e9d8
[misc] deprecate num_row in favor of num_step
mathewjhan Jul 17, 2026
3ba0d99
[fix] get_batch CP regressions: nonexistent Tensor.batch() and double…
yushengsu-thu Jul 18, 2026
2b22920
[test] pin get_batch CP=2 behavior: multi-LoRA token counts and singl…
yushengsu-thu Jul 18, 2026
ef2ba07
[fix] multi-LoRA validation crashes when --global-batch-size is unset
yushengsu-thu Jul 19, 2026
95ab736
[fix] re-registered adapter name must not inherit the previous tenant…
yushengsu-thu Jul 19, 2026
4455037
[fix] reject bad adapter registrations instead of killing the shared …
yushengsu-thu Jul 19, 2026
f0940c1
[fix] reset the retired slot's group-level FusedAdam clock, not just …
yushengsu-thu Jul 19, 2026
0d4cfb2
[fix] slice HF PEFT checkpoint exports to the adapter's real rank
yushengsu-thu Jul 19, 2026
45ef970
[fix] gate adapter checkpoint writers on the dp x cp group, not dp alone
yushengsu-thu Jul 19, 2026
9e51024
[fix] honor --accumulate-allreduce-grads-in-fp32 on the bridge DDP path
yushengsu-thu Jul 19, 2026
918b6db
[fix] engine-recovery full push must not bump unchanged adapters' ver…
yushengsu-thu Jul 19, 2026
c02f79a
[fix] honor --num-rollout in the multi-LoRA driver loop
yushengsu-thu Jul 19, 2026
1b6f69d
[fix] register --multi-lora-max-empty-wait-s as a CLI argument
yushengsu-thu Jul 19, 2026
0461e7c
[fix] reject non-Adam optimizers for multi-LoRA at launch
yushengsu-thu Jul 19, 2026
1a4ef27
[fix] reject multi-LoRA with MILES_EXPERIMENTAL_FT_TRAINER at launch
yushengsu-thu Jul 19, 2026
c42d9e5
[misc] warn when multi-LoRA runs with a decaying LR schedule
yushengsu-thu Jul 19, 2026
3804f78
add comment
yushengsu-thu Jul 19, 2026
f9381b7
change controller name to multilora controller
yushengsu-thu Jul 19, 2026
f8f8779
Revert "[fix] honor --num-rollout in the multi-LoRA driver loop"
yushengsu-thu Jul 19, 2026
7cc00a6
[fix] multi-LoRA does not require --num-rollout
yushengsu-thu Jul 19, 2026
d5305bc
Revert "[fix] multi-LoRA does not require --num-rollout"
yushengsu-thu Jul 19, 2026
ebdba18
[fix] per-adapter LR schedule for multi-LoRA
yushengsu-thu Jul 19, 2026
d5b81c1
[refactor] move multi-LoRA rollout infra from examples into miles/rol…
yushengsu-thu Jul 19, 2026
34bf4bc
[refactor] split miles/utils/multi_lora.py into the miles/ray/multi_l…
yushengsu-thu Jul 19, 2026
85a23c7
[refactor] move the multi-LoRA driver to the repo root, default its p…
yushengsu-thu Jul 19, 2026
0bd4291
[refactor] extract validate_multi_lora_args out of miles_validate_args
yushengsu-thu Jul 19, 2026
e16b310
[refactor] extract the multi-LoRA optimizer-step arm out of train_one…
yushengsu-thu Jul 19, 2026
06ea30d
[refactor] extract the actor's multi-LoRA blocks into multi_lora_util…
yushengsu-thu Jul 19, 2026
a83b90a
Merge remote-tracking branch 'radixark/main' into feat/multi-lora-async
yushengsu-thu Jul 20, 2026
a3c8daa
[refactor] split _update_multi_lora_weight_implementation out of the …
yushengsu-thu Jul 20, 2026
7839bd6
[refactor] drop the dead upsert plumbing, tidy _send_one_multi_lora_a…
yushengsu-thu Jul 20, 2026
e068a49
[lint] apply isort and black to the multi-LoRA files
yushengsu-thu Jul 20, 2026
39fcb29
[style] make both LoRA-instance branches lazy-import symmetrically
yushengsu-thu Jul 20, 2026
a1ce9e4
[refactor] get_batch fetches adapter_slots itself
yushengsu-thu Jul 20, 2026
c368a30
Merge remote-tracking branch 'radixark/main' into feat/multi-lora-async
yushengsu-thu Jul 20, 2026
323a991
[example] replace the multi-LoRA shell scripts with a typer launcher
yushengsu-thu Jul 20, 2026
38a576c
[docs] compress the multi-LoRA comments and docstrings to one or two …
yushengsu-thu Jul 21, 2026
47c3ec0
Merge remote-tracking branch 'radixark/main' into feat/multi-lora-async
yushengsu-thu Jul 21, 2026
ef7d030
Merge remote-tracking branch 'radixark/main' into feat/multi-lora-async
yushengsu-thu Jul 21, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view

These merge commits were added into this branch cleanly.

There are no new changes to show.