- Extract lambda to a named function to ensure cache hits. by @copybara-service[bot] in AI-Hypercomputer#127
- Always write array metadata if self._array_metadata_store is not None. by @copybara-service[bot] in AI-Hypercomputer#125
- Add "Shared Pathways Service" for Pathways-on-Cloud by @copybara-service[bot] in AI-Hypercomputer#128
- Exposes
pathwaysutils.profiling._start_trace_from_profile_requestaspathwaysutils.experimental.profiling.start_trace. by @copybara-service[bot] in AI-Hypercomputer#141 - Patch internal JAX profiler functions (enabling
jax.profiler.trace) and add a test forjax.profiler.trace. by @copybara-service[bot] in AI-Hypercomputer#140 - Expose
_split_by_mesh_axisdirectly inpw_jax. by @copybara-service[bot] in AI-Hypercomputer#147 - Use Pathways
_transfer_to_shardingfor resharding in experimental reshard API. by @copybara-service[bot] in AI-Hypercomputer#148 - Allow specifying a custom proxy job name in _ISCPathways by @copybara-service[bot] in AI-Hypercomputer#157
- Update the github action for PyPI by @lukebaumann in AI-Hypercomputer#105
- Expose
is_pathways_backend_usedinpathwaysutils. by @copybara-service[bot] in AI-Hypercomputer#107 - Treat additional error types as potential slice down issues. by @copybara-service[bot] in AI-Hypercomputer#109
- Adding split_by_mesh_axis to experimental for use by a new experimental reshard. by @copybara-service[bot] in AI-Hypercomputer#112
- Update pathways.experimental.reshard so that PyTrees with arrays that have different device sets can be resharded. by @copybara-service[bot] in AI-Hypercomputer#113
- Update GitHub action unittest matrix by @lukebaumann in AI-Hypercomputer#111
- Handle new style PRNG keys in
reshardandCloudPathwaysArrayHandler
- Updates to Pathways orbax handler. In AI-Hypercomputer#81 AI-Hypercomputer#89 AI-Hypercomputer#90
- Improvements to pathwaysutils/elastic by @copybara-service[bot] inhttps://github.com/AI-Hypercomputer/pull/72 AI-Hypercomputer#84 AI-Hypercomputer#85 AI-Hypercomputer#86 AI-Hypercomputer#99 AI-Hypercomputer#100 AI-Hypercomputer#102
- Handle JAX API interface compatibility by @copybara-service[bot] in AI-Hypercomputer#88 AI-Hypercomputer#93 AI-Hypercomputer#95
- Adding jax version to the github actions matrix by @lukebaumann in AI-Hypercomputer#94
- Added LRU cache and tests to pathwaysutils. by @copybara-service[bot] in AI-Hypercomputer#83
- Moving initialization logic into its own module by @copybara-service[bot] in AI-Hypercomputer#96
- Port the
collect_profilescript from JAX to PathwaysUtils - Remove support for legacy initialize
- Add collect_profile as a script of pathwaysutils
- Make CloudPathwaysArrayHandler compatible with async directory creation feature in orbax
- Bump the JAX requirement to 0.5.1
- Introduce
pathwaysutils.initialize()to remove relying on side-effects fromimport pathwaysutils. by @copybara-service in AI-Hypercomputer#47 - Add a test for proxy backend registration by @copybara-service in AI-Hypercomputer#55
- Adding debugging utilities by @copybara-service in AI-Hypercomputer#57
- Adding an elastic manager and reshard modules by @copybara-service in AI-Hypercomputer#58
- Update README for Colocated Python Sidecar. by @copybara-service in AI-Hypercomputer#60
- Disabled JAX's compilation cache
- Updated Orbax handler to use bulk APIs
- Updates to support JAX 0.5.0
- Updated
setup.pytopyproject.toml - Added this changelog
- Added unittests
- Prepared for PyPI release
- Decreased logging severity for most logs
- Persistence enabled
- General argument type fixes