Skip to content

Support hierarchical scaling with per-axis BF16 and subchannel microscaling in QWIX. - #374

Open
copybara-service[bot] wants to merge 1 commit into
mainfrom
test_980172960
Open

copybara-service[bot] wants to merge 1 commit into
mainfrom
test_980172960

Conversation

@copybara-service

Copy link
Copy Markdown

Support hierarchical scaling with per-axis BF16 and subchannel microscaling in QWIX.

  • Support hierarchical scaling: outer per-axis scaling with BF16 scaling factors followed by inner power-of-2 subchannel microscaling (mxfp4, mxfp8, mxint4, mxint8).
  • Automatically no-op when the format does not use microscaling.
  • Add granular configuration options to DotGeneralQtConfig for forward, dlhs, and drhs (separated for grad and residual).
  • Add comprehensive unit tests in qarray_test and dot_general_qt_test validating accuracy, gradient correctness, and no-op behavior.

…caling in QWIX.

- Support hierarchical scaling: outer per-axis scaling with BF16 scaling factors followed by inner power-of-2 subchannel microscaling (mxfp4, mxfp8, mxint4, mxint8).
- Automatically no-op when the format does not use microscaling.
- Add granular configuration options to DotGeneralQtConfig for forward, dlhs, and drhs (separated for grad and residual).
- Add comprehensive unit tests in qarray_test and dot_general_qt_test validating accuracy, gradient correctness, and no-op behavior.

PiperOrigin-RevId: 980172960
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.

0 participants