[Triton/Gluon] [ASM] [HIP] Block-sparse MHAv4 with load-balancing - #5005

Merged

valarLip merged 42 commits into

nsakkine:blocksparse_mha_v4

Aug 28, 2026

Commits on Aug 18, 2026

perf(mha_v4): avoid copying odd-tail FP6 V inputs
jcaraban
committed
c5b1546
feat(mha_v4): support grouped query attention
jcaraban
committed
918947e
feat(mha_v4): add MXFP8 raw entrypoint
jcaraban
committed
2205978
docs(mha_v4): clarify grouped-query attention contract
jcaraban
committed
c49d98d
feat(mha_v4): add gfx942 native FP8 kernel
jcaraban
committed
4f0488a
fix(mha_v4): canonicalize rotated FP8 preprocessing
jcaraban
committed
7dc62a5
refactor(bench): simplify MHA v4 quantized runners
jcaraban
committed
4815d47
perf(mha_v4): deploy gfx942 XCD-swizzled kernels
jcaraban
committed
92deb9a

Commits on Aug 19, 2026

perf(mha_v4): deploy gfx942 block kernels
jcaraban
committed
e62e076
fix(mha_v4): handle singleton-head rotation strides
jcaraban
committed
e79b1c8
perf(mha_v4): deploy retimed gfx942 I8/FP8 kernels
jcaraban
committed
361ad58

Commits on Aug 20, 2026

fix(mha_v4): deploy corrected gfx942 PV LDS waits
jcaraban
committed
ef3a635

Commits on Aug 22, 2026

fix(fmha): deploy gfx942 V staging
jcaraban
committed
8b2acc8
fix(fmha): update gfx942 I8FP8 kernel
jcaraban
committed
75ff143

Commits on Aug 23, 2026

feat(mha): add bf16 to mha v4
jcaraban
committed
a2a7db1
perf(fmha): deploy optimized gfx942 block kernels
jcaraban
committed
b4d7c11

Commits on Aug 24, 2026

style(mha_v4): apply repository formatting
jcaraban
committed
550c90e
test(mha_v4): isolate compile parity cases
jcaraban
committed
c27ea90
Merge remote-tracking branch 'origin/main' into mha_v4_gqa_etc
jcaraban
committed
3b7e931
fix ruff warnings
jcaraban
committed
a1037ee
fix(mha_v4): enforce contiguous rotation layout
jcaraban
committed
932fb80

Commits on Aug 25, 2026

fix(mha_v4): update deterministic BF16 kernel
jcaraban
committed
18cc699
Revert "fix(mha_v4): handle singleton-head rotation strides"
jcaraban
committed
3f3e117

Commits on Aug 26, 2026

Sparse MHAv4 initial commit
nsakkine
committed
0f58938
Enable sparse GQA. Fix rebase issues. Fix rotate_activations bug
nsakkine
committed
1a47410
Give MHA v4 its own hd128 rotation instead of calling into dsv4
nsakkine
committed
1174bf6
Cut the fixed cost of the sorted-sparse work table
nsakkine
committed
189c648
Build small sorted-sparse work tables in one kernel
nsakkine
committed
76416c9
Rank work table entries across a wave instead of a thread
nsakkine
committed
4d1a516
fix(mha_v4): restore BF16 dense dispatch and launcher byte strides
nsakkine
committed
0c4248d
style: run black, ruff, and clang-format on block-sparse MHA v4 changes
nsakkine
committed
389c7bb
docs(mha_v4): trim sparse section implementation detail

nsakkine
and
cursoragent
committed
d9d4985
Add gfx942 sorted-sparse MHA v4 kernels
nsakkine
committed
3f71438
Merge remote-tracking branch 'origin/main' into blocksparse_mha_v4
nsakkine
committed
4793bb1