Skip to content

turboquant: guard L=1 value kernels behind not use_rht (fix masked decode under RHT) #1244

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
Blaizzy merged 2 commits into from
May 30, 2026
Merged

turboquant: guard L=1 value kernels behind not use_rht (fix masked decode under RHT) #1244

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
0113d46
turboquant: guard L=1 value kernels behind `not use_rht`
popfido May 30, 2026
7937ecb
Merge branch 'main' into fix/turboquant-rht-masked-decode
Blaizzy May 30, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
9 changes: 6 additions & 3 deletions mlx_vlm/turboquant.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4287,7 +4287,7 @@ def score(self, queries: mx.array, state: TurboQuantMSEState) -> mx.array:
return self.score_prepared(self.prepare_queries(queries), state)

def weighted_sum(self, weights: mx.array, state: TurboQuantMSEState) -> mx.array:
if weights.shape[-2] == 1:
if not self.use_rht and weights.shape[-2] == 1:
fast_output = _metal_mse_weighted_sum(
weights,
state,
Expand Down Expand Up @@ -4327,8 +4327,11 @@ def weighted_sum_stats_from_scores(
self, scores: mx.array, state: TurboQuantMSEState
) -> tuple[mx.array, mx.array, mx.array]:
max_scores = mx.max(scores, axis=-1)
# Metal kernel fast path: only for single-query decode (L=1)
if scores.ndim == 5 and scores.shape[-2] == 1:
# Metal kernel fast path: only for single-query decode (L=1).
# Skip under RHT: these L=1 value kernels undo the codec rotation with
# matmul(., rotation) and do not implement the RHT inverse, so they
# corrupt the output when use_rht is set (the codec default).
if not self.use_rht and scores.ndim == 5 and scores.shape[-2] == 1:
max_scores_2d = max_scores.reshape(
max_scores.shape[0],
max_scores.shape[1],
Expand Down
Loading