Skip to content

Abhinavsing/elastic#4434

Open
abhinavclemson wants to merge 31 commits into
lukebaumann/elastic-snapshotfrom
abhinavsing/elastic
Open

Abhinavsing/elastic#4434
abhinavclemson wants to merge 31 commits into
lukebaumann/elastic-snapshotfrom
abhinavsing/elastic

Conversation

@abhinavclemson

Copy link
Copy Markdown
Collaborator

No description provided.

lukebaumann and others added 30 commits June 30, 2026 17:42
…ing, dynamic mesh recovery, and GKE scale-up)

Integrates snapshot-based failure recovery directly into the training iteration.
- Implements asynchronous host snapshot manager (snapshot.py) to copy TrainState parameters to CPU pinned host memory in a background daemon thread, minimizing host memory overhead by freeing old snapshot buffers before allocating new transfers.
- Integrates recovery and scale-up pathways inside the training loop in train.py, including dynamic configuration mutation and JAX set_mesh compiler safety isolation.
- Utilizes the standard pathwaysutils.elastic.manager.Manager directly to track slice health, eliminating custom manager duplication and optimizing startup times by avoiding redundant checks.
- Optimizes RNG key sharding during JIT by defining out_shardings directly inside jax.jit(jax.random.fold_in, out_shardings=...).
- Incorporates configurable elastic_snapshot_interval parameter in Pydantic settings.
- Clarifies checkpoint restoration logging.
@codecov

codecov Bot commented Jul 11, 2026

Copy link
Copy Markdown

Codecov Report

❌ Patch coverage is 33.33333% with 2 lines in your changes missing coverage. Please review.

Files with missing lines Patch % Lines
src/maxtext/common/data_loader.py 33.33% 2 Missing ⚠️

📢 Thoughts on this report? Let us know!

@lukebaumann lukebaumann force-pushed the lukebaumann/elastic-snapshot branch 2 times, most recently from bee5614 to b6866ef Compare July 14, 2026 08:00
@lukebaumann lukebaumann force-pushed the lukebaumann/elastic-snapshot branch 9 times, most recently from 6afbbea to 1cf3c0a Compare July 14, 2026 08:51
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants