Releases: google/ml-flashpoint
Releases · google/ml-flashpoint
Release list
v0.0.9
Release Notes: v0.0.8 -> 97f5c54
🏆 Features
✅ Bug Fixes
- (9e10862) adapter/nemo: also close alt async calls queue for regular nemo checkpoints (#106)
- (6e1c211) replication: safely shutdown transfer service (#105)
- (8962c0b) adapter/nemo: safely call maybe_finalize_save_checkpoint in on_train_end (#104)
♻️ Refactoring
- (67fbe4a) checkpoint-object-manager: deprecate C++ object_manager layer for lightweight Python subprocess async deletion (#102)
📔 Documentation
Generated with: ./scripts/create_release.py
Full Changelog: v0.0.8...v0.0.9
v0.0.8
Release Notes: v0.0.7 -> cf7264e
🏆 Features
- (cf7264e) adapter/nemo: adapt to NeMo's new CheckpointIO interface and fix metadata persistence (#99)
- (569d32c) adapter/nemo: cleanup checkpoint container on_train_end (#94)
- (ec2462a) replication: support single-node execution by disabling replication (#92)
- (995f6d8) scripts: support ModelCheckpoint.on_train_batch_end in log parser (#86)
- (e9eb9d6) adapter/nemo: add fully parallel save/load wrapper option (#49)
✅ Bug Fixes
🕐 Performance
- (ceffdab) adapter/nemo: set use_fully_parallel_wrapper default to True (#93)
- (17457aa) adapter/nemo: reduce NUM_OF_BUFFERS_PER_OBJECT to 2 to reduce memory pressure (#81)
- (71217a4) Implement BufferPool for efficient memory reuse (#61)
🔍 Tests
📔 Documentation
- (d6b5ed6) add codewiki link! (#98)
- (1ad7beb) site: update latest write throughput numbers (#97)
- (998bd61) add logging guidelines (#96)
🔃 CI
- (fe2ca6a) bifurcate nemo2.0 and nemo-rl build profiles, use ci-success aggregate check in build (#91)
- (bf23474) run build on python 3.12 as well (#87)
- (8f760ae) fix VERSION extraction from TAG_NAME in cloudbuild.yaml (#85)
- (cba3c3a) fix $$VERSION env var syntax (#82)
- (75bf47a) validate TAG_NAME and force the version used for pypi upload (#79)
🔩 Chores
- (c0b755a) config: exclude .gemini and .worktrees from git, ruff, and pytest (#90)
- (5ee67ae) scripts: skip first two writes for throughput averaging (#89)
Generated with: ./scripts/create_release.py
New Contributors
- @devyanic11 made their first contribution in #49
Full Changelog: v0.0.7...v0.0.8
Contributors
devyanic11
Assets 2
v0.0.7
Release Notes: v0.0.6 -> e5ce6e4
🏆 Features
- (e5ce6e4) adapter/nemo: Allow
initial_write_buffer_size_bytes=Noneto use default size (#76) - (78f20d7) Cache checkpoint saving plan in each local rank. (#60)
🕐 Performance
- (ae95fb6) adapter/pytorch: offload spawn manager creation to background thread (threading.Thread) (#73)
♻️ Refactoring
📔 Documentation
🔃 CI
🔩 Chores
Generated with: ./scripts/create_release.py
Full Changelog: v0.0.6...v0.0.7
v0.0.6
What's Changed
Release Notes: v0.0.5 -> ae84f35
🏆 Features
✅ Bug Fixes
- (219007d) core/saver: create checkpoint_dir in each rank when initialize_checkpoint (#63)
- (e6c2a59) core/saver: create checkpoint base dir in checkpoint saver init (#62)
- (7075ce4) adapter/nemo: use spawn context for mp_manager to prevent OOM on NVRX restart (#59)
🕐 Performance
♻️ Refactoring
📔 Documentation
🔀 Build
- (ae84f35) use cibuildwheels for PyPi compatible packages (#64)
- (7bb914f) Upgrade GitHub Actions for Node 24 compatibility (#56)
- (06ebf54) set a fallback version for installs in dirs without .git (#53)
🔃 CI
- (5f49b38) fix download-artifact pinned version (#54)
- (d923449) implement workflow_run pattern for coverage comments (#51)
- (ec4aed7) make published package version dynamic (#45)
🔩 Chores
Generated with: ./scripts/create_release.py
New Contributors
- @ronaldw07 made their first contribution in #43
- @salmanmkc made their first contribution in #56
Full Changelog: v0.0.5...v0.0.6
Contributors
salmanmkc and ronaldw07
Assets 2
v0.0.5
Release Notes: v0.0.4 -> 85d780d
🏆 Features
- (150d52c) scripts: use on_train_batch_start for precise step timing (#39)
- (06ebcbf) adapter/nemo: add EventLoggingCallback for lifecycle monitoring (#34)
- (efd8560) core: implement optimized serialization tensor saving (#12)
🕐 Performance
♻️ Refactoring
- (22d8691) Move C++ tests to tests/ directory (#33)
- (066f2a3) core: Move context recovery logic strictly to the NeMo layer (#19)
🎨 Styles
📔 Documentation
- (6f36c9c) fix CI badge (#41)
- (d9f8864) Update README.md front matter and links (#36)
- (14b9566) site: clarify perf points (#24)
- (6e20be2) tweak README verbiage and user doc perf section (#23)
- (e10fc58) update readme intro with more info and add build badge (#22)
- (a2450dc) site: add PyTorch DCP usage guide; segregate imports (#17)
- (59ca9da) site: fix perf section formatting; add code package links (#15)
- (89f4cf1) site: clarify performance section (#14)
- (359c825) site: Update user-guide for Megatron-LM usage (#13)
🔀 Build
🔃 CI
- (85d780d) add GCB configuration yaml for pypi releases (#35)
- (6ec066c) Add code coverage summaries (#25)
- (182922f) re-enable clang-format check in build workflow (#38)
- (1d08a8d) add license header validation check (#27)
🔩 Chores
- (bff1c08) scripts: Add aggregate and per-node throughput calculations (#32)
- (e113677) add feature request issue template (#16)
Generated with: ./scripts/create_release.py
Full Changelog: v0.0.4...v0.0.5
v0.0.4 - First GitHub Release
Note that the latest internal release prior to migrating to GitHub was v0.0.3, so continuing version numbering for consistency and clarity.
Release Notes: BEGINNING -> b976758
♻️ Refactoring
📔 Documentation
- (b976758) update readme and user doc site for GitHub and public usage (#11)
- (123596a) fix site deployment workflow (#4)
🔃 CI
🔩 Chores
- (1df527e) example_cpp: remove unused example C++ extension (#10)
- (ab85caa) import codebase from Gerrit (#1)
- (99886c4) init repo with placeholder README.md
Generated with: ./scripts/create_release.py
Full Changelog: https://github.com/google/ml-flashpoint/commits/v0.0.4