[Lilo][Miles runtime] Multi-node LoRA support and 256k Qwen3.8-27B definitions - #39

Merged

micahtyong merged 23 commits into

mainfrom

devin/1789702581-lilo-27b-256k-multinode

Sep 22, 2026

Commits on Sep 18, 2026

Add 256k Qwen3.8-27B definition and multi-node Miles trainer support

micahtyong
and
devin-ai-integration[bot]
committed
8fcfa3e
Require TP groups to tile a node in multi-node configs

micahtyong
and
devin-ai-integration[bot]
committed
5696e7a
Fix multi-node checkpoint capture publication

micahtyong
and
devin-ai-integration[bot]
committed
7fa4d1d
Refresh multi-node trainer assets before Ray joins

micahtyong
and
devin-ai-integration[bot]
committed
71e3897
Widen the 256k trainer's collective timeout

micahtyong
and
devin-ai-integration[bot]
committed
c6c4267
Cap coalesced forward_backward batches on the 256k rung

micahtyong
and
devin-ai-integration[bot]
committed
49d8175
Align 256k sequences to the context/tensor-parallel layout

micahtyong
and
devin-ai-integration[bot]
committed
4d3c22b
Load actor module under its package name in test so relative profiling import resolves

micahtyong
and
devin-ai-integration[bot]
committed
cbed492

Commits on Sep 21, 2026

Elect checkpoint-commit representatives by Modal task id

micahtyong
and
devin-ai-integration[bot]
committed
55e941e
Publish multi-node checkpoints with every node's shards

micahtyong
and
devin-ai-integration[bot]
committed
aef7148
Fail checkpoint volume sync on every rank instead of hanging

micahtyong
and
devin-ai-integration[bot]
committed
c467cb4

Commits on Sep 22, 2026

Give multi-node trainer actors the checkpoint volume name

micahtyong
and
devin-ai-integration[bot]
committed
192b95e
Publish checkpoints from the volume's committed state

micahtyong
and
devin-ai-integration[bot]
committed
3ece760
Stop refreshing a node's volume view after committing checkpoints

micahtyong
and
devin-ai-integration[bot]
committed
faaaecb
Keep rank zero's own files in its view when publishing a checkpoint

micahtyong
and
devin-ai-integration[bot]
committed
186c33d
Fail a checkpoint publish that is missing a peer's shards

micahtyong
and
devin-ai-integration[bot]
committed
0f5aa78
Install a checkpoint capture from committed volume state

micahtyong
and
devin-ai-integration[bot]
committed
b7d53c0
Add the two-node 256k Qwen3.8-27B definition

micahtyong
and
devin-ai-integration[bot]
committed
035d3e4
Trim checkpoint publication docstrings

micahtyong
and
devin-ai-integration[bot]
committed
f2c12ab
Check node count when waiting for the Ray cluster

micahtyong
and
devin-ai-integration[bot]
committed
c795591
Finalize the two-node 256k definition registry

micahtyong
and
devin-ai-integration[bot]
committed
52753d5
Rename the two-node 256k definition symbols

micahtyong
and
devin-ai-integration[bot]
committed
1607559
Shorten _align_row docstring per review

micahtyong
and
devin-ai-integration[bot]
committed
04a8c7a