ReleaseDatabricks (DBRX)Databricks (DBRX)published Mar 14, 2025seen 5d

databricks/megablocks v0.8.0

databricks/megablocks

Open original ↗

Captured source

source ↗
published Mar 14, 2025seen 5dcaptured 8hhttp 200method plain

v0.8.0

Repository: databricks/megablocks

Tag: v0.8.0

Published: 2025-03-14T17:57:28Z

Prerelease: no

Release notes:

Breaking Changes

As a consequence of the torch 2.6.0 upgrade, sparse support is disabled for megablocks (meaning that only grouped support is available).

For additional context, torch 2.6.0 depends on triton 3.2.0, which introduced some change to how it handles dtype promotion when two binary operands have different dtypes, and as a result we're encountering an int16 overflow in the stk dependency of megablocks which results in an illegal memory access (IMA). Once this issue is resolved, we will release a new version of megablocks. View https://github.com/databricks/megablocks/pull/168 for additional details.

What's Changed

  • Updated pytorch and disabled sparse tests by @rithwik-db in https://github.com/databricks/megablocks/pull/168

New Contributors

  • @rithwik-db made their first contribution in https://github.com/databricks/megablocks/pull/168

Full Changelog: https://github.com/databricks/megablocks/compare/v0.7.0...v0.8.0

Notability

notability 5.0/10

Incremental release of MoE library