Skip to content

[gluon][pa_mqa_logits] memory-safety: mask all OutLogits buffer_store lanes (does NOT fix long-context accuracy) #2936

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

Draft
maeehart wants to merge 3 commits into
base: main
Choose a base branch
from
+170 −27
Draft

[gluon][pa_mqa_logits] memory-safety: mask all OutLogits buffer_store lanes (does NOT fix long-context accuracy) #2936

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
69e6687
gluon(pa_mqa_logits): mask all OutLogits buffer_store lanes (preshuff…
maeehart Apr 28, 2026
b81f8e8
style(pa_mqa_logits): Black-format mask expressions for CI
maeehart Apr 28, 2026
900df6d
gluon(pa_mqa_logits): upper-bound masks on non-preshuffle OutLogits s…
maeehart Apr 28, 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
197 changes: 170 additions & 27 deletions aiter/ops/triton/gluon/pa_mqa_logits.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -282,9 +282,18 @@ def _gluon_deepgemm_fp8_paged_mqa_logits(
context_idx
+ gl.arange(0, ChunkK, layout=gl.SliceLayout(0, mfma_layout))
),
mask=context_idx
+ gl.arange(0, ChunkK, layout=gl.SliceLayout(0, mfma_layout))
>= 0,
mask=(
(
context_idx
+ gl.arange(0, ChunkK, layout=gl.SliceLayout(0, mfma_layout))
>= 0
)
& (
context_idx
+ gl.arange(0, ChunkK, layout=gl.SliceLayout(0, mfma_layout))
< max_model_len
)
),
)

context_idx = split_context_start + split_context_length - ChunkK
Expand All @@ -310,8 +319,18 @@ def _gluon_deepgemm_fp8_paged_mqa_logits(
ptr=OutLogits_buffer,
offsets=(pid_batch * next_n + pid_next_n) * stride_out_batch
+ (context_idx + gl.arange(0, ChunkK, layout=gl.SliceLayout(0, mfma_layout))),
mask=context_idx + gl.arange(0, ChunkK, layout=gl.SliceLayout(0, mfma_layout))
>= 0,
mask=(
(
context_idx
+ gl.arange(0, ChunkK, layout=gl.SliceLayout(0, mfma_layout))
>= 0
)
& (
context_idx
+ gl.arange(0, ChunkK, layout=gl.SliceLayout(0, mfma_layout))
< max_model_len
)
),
)


Expand Down Expand Up @@ -595,9 +614,16 @@ def _gluon_deepgemm_fp8_paged_mqa_logits_preshuffle(
context_idx
+ gl.arange(0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout))
),
mask=context_idx
+ gl.arange(0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout))
>= split_context_start,
mask=(
context_idx
+ gl.arange(0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout))
>= split_context_start
)
& (
context_idx
+ gl.arange(0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout))
< max_model_len
),
)

for context_idx in range(
Expand Down Expand Up @@ -663,10 +689,22 @@ def _gluon_deepgemm_fp8_paged_mqa_logits_preshuffle(
0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout)
)
),
mask=context_idx
+ ChunkKPerStage
+ gl.arange(0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout))
>= split_context_start,
mask=(
context_idx
+ ChunkKPerStage
+ gl.arange(
0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout)
)
>= split_context_start
)
& (
context_idx
+ ChunkKPerStage
+ gl.arange(
0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout)
)
< max_model_len
),
)

# =======================================================================================
Expand Down Expand Up @@ -735,6 +773,14 @@ def _gluon_deepgemm_fp8_paged_mqa_logits_preshuffle(
0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout)
)
),
mask=(
context_idx
+ ChunkK
+ gl.arange(
0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout)
)
< max_model_len
),
)

context_idx = split_context_start + split_context_length - ChunkK
Expand Down Expand Up @@ -769,10 +815,18 @@ def _gluon_deepgemm_fp8_paged_mqa_logits_preshuffle(
+ ChunkKPerStage
+ gl.arange(0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout))
),
mask=context_idx
+ ChunkKPerStage
+ gl.arange(0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout))
>= split_context_start,
mask=(
context_idx
+ ChunkKPerStage
+ gl.arange(0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout))
>= split_context_start
)
& (
context_idx
+ ChunkKPerStage
+ gl.arange(0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout))
< max_model_len
),
)
else:
context_idx = split_context_start
Expand Down Expand Up @@ -925,6 +979,11 @@ def _gluon_deepgemm_fp8_paged_mqa_logits_preshuffle(
context_idx
+ gl.arange(0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout))
),
mask=(
context_idx
+ gl.arange(0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout))
< max_model_len
),
)

for context_idx_ in range(
Expand Down Expand Up @@ -1000,6 +1059,14 @@ def _gluon_deepgemm_fp8_paged_mqa_logits_preshuffle(
0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout)
)
),
mask=(
context_idx_
+ ChunkKPerStage
+ gl.arange(
0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout)
)
< max_model_len
),
)

# =======================================================================================
Expand Down Expand Up @@ -1074,6 +1141,14 @@ def _gluon_deepgemm_fp8_paged_mqa_logits_preshuffle(
0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout)
)
),
mask=(
context_idx_
+ ChunkK
+ gl.arange(
0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout)
)
< max_model_len
),
)
context_idx = context_idx_ + ChunkK

Expand Down Expand Up @@ -1107,6 +1182,12 @@ def _gluon_deepgemm_fp8_paged_mqa_logits_preshuffle(
+ ChunkKPerStage
+ gl.arange(0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout))
),
mask=(
context_idx
+ ChunkKPerStage
+ gl.arange(0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout))
< max_model_len
),
)


Expand Down Expand Up @@ -1407,9 +1488,16 @@ def _gluon_deepgemm_fp8_paged_mqa_logits_preshuffle_varctx(
context_idx
+ gl.arange(0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout))
),
mask=context_idx
+ gl.arange(0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout))
>= split_context_start,
mask=(
context_idx
+ gl.arange(0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout))
>= split_context_start
)
& (
context_idx
+ gl.arange(0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout))
< max_model_len
),
)

for context_idx in range(
Expand Down Expand Up @@ -1475,10 +1563,22 @@ def _gluon_deepgemm_fp8_paged_mqa_logits_preshuffle_varctx(
0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout)
)
),
mask=context_idx
+ ChunkKPerStage
+ gl.arange(0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout))
>= split_context_start,
mask=(
context_idx
+ ChunkKPerStage
+ gl.arange(
0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout)
)
>= split_context_start
)
& (
context_idx
+ ChunkKPerStage
+ gl.arange(
0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout)
)
< max_model_len
),
)

# =======================================================================================
Expand Down Expand Up @@ -1547,6 +1647,14 @@ def _gluon_deepgemm_fp8_paged_mqa_logits_preshuffle_varctx(
0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout)
)
),
mask=(
context_idx
+ ChunkK
+ gl.arange(
0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout)
)
< max_model_len
),
)

context_idx = split_context_start + split_context_length - ChunkK
Expand Down Expand Up @@ -1581,10 +1689,18 @@ def _gluon_deepgemm_fp8_paged_mqa_logits_preshuffle_varctx(
+ ChunkKPerStage
+ gl.arange(0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout))
),
mask=context_idx
+ ChunkKPerStage
+ gl.arange(0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout))
>= split_context_start,
mask=(
context_idx
+ ChunkKPerStage
+ gl.arange(0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout))
>= split_context_start
)
& (
context_idx
+ ChunkKPerStage
+ gl.arange(0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout))
< max_model_len
),
)
else:
context_idx = split_context_start
Expand Down Expand Up @@ -1737,6 +1853,11 @@ def _gluon_deepgemm_fp8_paged_mqa_logits_preshuffle_varctx(
context_idx
+ gl.arange(0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout))
),
mask=(
context_idx
+ gl.arange(0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout))
< max_model_len
),
)

for context_idx_ in range(
Expand Down Expand Up @@ -1812,6 +1933,14 @@ def _gluon_deepgemm_fp8_paged_mqa_logits_preshuffle_varctx(
0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout)
)
),
mask=(
context_idx_
+ ChunkKPerStage
+ gl.arange(
0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout)
)
< max_model_len
),
)

# =======================================================================================
Expand Down Expand Up @@ -1886,6 +2015,14 @@ def _gluon_deepgemm_fp8_paged_mqa_logits_preshuffle_varctx(
0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout)
)
),
mask=(
context_idx_
+ ChunkK
+ gl.arange(
0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout)
)
< max_model_len
),
)
context_idx = context_idx_ + ChunkK

Expand Down Expand Up @@ -1919,4 +2056,10 @@ def _gluon_deepgemm_fp8_paged_mqa_logits_preshuffle_varctx(
+ ChunkKPerStage
+ gl.arange(0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout))
),
mask=(
context_idx
+ ChunkKPerStage
+ gl.arange(0, ChunkKPerStage, layout=gl.SliceLayout(0, mfma_layout))
< max_model_len
),
)
Loading