Build AITER from source at a pinned tag and prebuild every attention variant - #375

Merged

demandal25 merged 42 commits into

amd-integrationfrom

aiter-source-build

Sep 15, 2026

Commits on Sep 14, 2026

docs(skill): pick the backend by measurement, not by reputation

demandal25
and
claude
committed
b352dce
fix(rocm): key the AITER lib cache by ROCm as well as version

demandal25
and
claude
committed
01c7b73
fix(rocm): resolve AITER's csrc include dir without requiring aiter_meta

demandal25
and
claude
committed
5c0e653
docs(rocm): AITER's build machinery needs no GPU; only this driver does

demandal25
and
claude
committed
fbaf467
feat(rocm): name and compose every AITER attention variant's CK build

demandal25
and
claude
committed
b49e9ff
build(rocm): build AITER from source at v0.1.20 and prebuild the variants

demandal25
and
claude
committed
6dd43ef
build(rocm): move the AITER pin to v0.1.21.post2, ABI and all

demandal25
and
claude
committed
bfeab6e
build(rocm): prebuild all 40 attention variants, not 36

demandal25
and
claude
committed
c619c28
fix(rocm): re-measure the AITER soft-cap defect at 0.1.21.post2

demandal25
and
claude
committed
fa4383e
docs: the supported AITER install is a pinned source build

demandal25
and
claude
committed
2fb6b08
fix(rocm): guard the stream on the migrated norm and rope shims

demandal25
and
claude
committed
ad88c14
test(rocm): cover AITER shim stream affinity

demandal25
and
claude
committed
2b8527c
fix(rocm): drop the variant build trees, and make the floor check real

demandal25
and
claude
committed
bb3e6e5
fix(rocm): prebuild the asm module the loader now dlopens

demandal25
and
claude
committed
f07b293
fix(rocm): make --check cover the loader modules it builds

demandal25
and
claude
committed
13f8053
test(rocm): pin the AITER state these two tests depend on

demandal25
and
claude
committed
9f99d53
fix(rocm): prebuild the fp8 batch-prefill arm the loader can ask for

demandal25
and
claude
committed
de09d06
fix(rocm): scope --only's check, honour AITER's clang, keep artifacts readable

demandal25
and
claude
committed
f640a1b
refactor(rocm): one composer for the AITER cache tag, and trim to the comment cap

demandal25
and
claude
committed
0e28292
docs(rocm): record which stream-affinity cases actually catch a missing guard

demandal25
and
claude
committed
6d80af6
test(rocm): cover fp8 paged prefill, the only route that serves fp8

demandal25
and
claude
committed
7733443
fix(rocm): reap killed builds, chmod what the driver wrote, run the long job first

demandal25
and
claude
committed
18c224a
docs(skill): correct the native page-size guidance, and guard two probes

demandal25
and
claude
committed
ea14896
fix(rocm): make only the jit directories world-writable, not the kernels

demandal25
and
claude
committed
1d85803
fix(rocm): prebuild the two modules a source install otherwise leaves cold

demandal25
and
claude
committed
02bc2a3
fix(rocm): close the sticky-bit hole, export GPU_ARCHS, correct three stale claims

demandal25
and
claude
committed
6f481af

Commits on Sep 15, 2026

Merge branch 'amd-integration' into aiter-source-build
demandal25
authored
b5b0705
fix(rocm): keep the installed AITER tree read-only, give the fallback a HOME

demandal25
and
claude
committed
c50dac0
test(rocm): make the token-order check actually check order

demandal25
and
claude
committed
a977c15
docs(rocm): install AITER's build requirements before the no-isolation build

demandal25
and
claude
committed
e84279f
fix(rocm): settle HOME before the prebuild resolves AITER's jit directory

demandal25
and
claude
committed
a4daaa5
docs(rocm): the 0.1.21 floor is the norm/rope ABI move, not a struct change

demandal25
and
claude
committed
fb6bc3f
test(rocm): pin the direction literal, the Dockerfile floor, and the fp8 range

demandal25
and
claude
committed
4514b8a
fix(rocm): stop widening a shared directory, and correct two stated numbers

demandal25
and
claude
committed
d56e274
fix(rocm): pin where the prebuild writes, and scan the loader's call sites too

demandal25
and
claude
committed
a27cf11
fix(rocm): drop the fp8 LSE arm, and detect an artifact with no kernels in it

demandal25
and
claude
committed
e441b65
fix(rocm): put the gfx950 paged fault where routing can see it, and four others

demandal25
and
claude
committed
f3b52e1
fix(rocm): tell source installs how to get the asm module they do not have

demandal25
and
claude
committed
714d451
fix(rocm): delete a duplicate fp8 test, and make the install recipe actually work

demandal25
and
claude
committed
724b9e7
fix(rocm): gate HIP_CLANG_PATH on existence, and finish the wheel-to-source sweep

demandal25
and
claude
committed
3a423d1
fix(rocm): cgroup-aware job budget, arch-safe header probe, and five smaller ones

demandal25
and
claude
committed
e2fdd10
fix(rocm): include AITER's real norm and rope headers instead of declaring them

demandal25
and
claude
committed
644d725