Skip to content

Support hybrid multi-pass quantization (mxfp8_16 Pass 1 + mxfp4 Pass 2) in QWIX - #372

Open
copybara-service[bot] wants to merge 1 commit into
mainfrom
test_981380857
Open

copybara-service[bot] wants to merge 1 commit into
mainfrom
test_981380857

Conversation

@copybara-service

Copy link
Copy Markdown

Support hybrid multi-pass quantization (mxfp8_16 Pass 1 + mxfp4 Pass 2) in QWIX

  • Support residual_decompose_hybrid for multi-pass decomposition with heterogeneous precision rules across passes.
  • Support hybrid_fp8_fp4_multipass_dot using mxfp8_16 (block 16) for coarse Pass 1 and mxfp4 (block 32) for fine Pass 2.
  • Support multi-pass modes: 'triangular' (3 passes), 'full_cross' (4 passes), 'lhs_high_precision' (2 passes), 'rhs_high_precision' (2 passes), and 'reconstructed'.
  • Add unit tests verifying hybrid SQNR bounds (>42 dB for triangular, >44 dB for full cross) and residual decomposition accuracy.

@copybara-service
copybara-service Bot force-pushed the test_981380857 branch 2 times, most recently from 95aa34e to f647f9a Compare September 15, 2026 00:20
…2) in QWIX

- Support residual_decompose_hybrid for multi-pass decomposition with heterogeneous precision rules across passes.
- Support hybrid_fp8_fp4_multipass_dot using mxfp8_16 (block 16) for coarse Pass 1 and mxfp4 (block 32) for fine Pass 2.
- Support multi-pass modes: 'triangular' (3 passes), 'full_cross' (4 passes), 'lhs_high_precision' (2 passes), 'rhs_high_precision' (2 passes), and 'reconstructed'.
- Add unit tests verifying hybrid SQNR bounds (>42 dB for triangular, >44 dB for full cross) and residual decomposition accuracy.

PiperOrigin-RevId: 981380857
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