Pull requests / #2758

#2758 sycl: STRATA_QFUSE - commit one-token rounds (the commit skipped them, so QFUSE gave wrong output)

open · @aslater3 · 0 comentarios · En GitHub

AMD / HIPNVIDIA / CUDAModels & quants

Descripción

## sycl: STRATA_QFUSE - commit one-token rounds (the commit skipped them, so QFUSE gave wrong output)

record_window does not self-commit a one-token window under STRATA_QFUSE (`!g_qfuse()` in its self_commit), but
Verifier::commit skipped the commit graph for every one-token window whenever one_token_self_commit() is on, so with
QFUSE those rounds never advanced the GDN state: wrong tokens. commit() now uses the same condition. Default path
(QFUSE off) unchanged.

A770, Coder IQ1_M, 8K code prompt, STRATA_QFUSE=1, 3 pairs: output is now the reference tokens (before the fix it
diverged); no speed change, QFUSE stays opt-in.

The CUDA/HIP engine looks to have the same mismatch: `src/core/verify.cpp` line 879 (record_window) has `!g_qfuse()` in its self-commit condition, but `Verifier::commit` (line 2655) does not. QFUSE was built on HIP, where the one-token self-commit is off, so it did not show there. I could not build or test CUDA here, so this PR only changes the SYCL port.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

En el sitio

Enlaces a install, modelos, releases.